Writing Sample

An Analysis of Fatal Car Accidents in 2001 By Leigh Ann Mastrini November 18, 2008

Introduction For the average American consumer, buying an automobile is one of the most important financial decisions they will make for that year or even decade. Approximately $389.7 billion is spent on motor vehicles every year, with 52% of that, or $203.7 billion, spent on light trucks that include Sport Utility Vehicles.1 The term, “Sport Utility Vehicle” is actually a generic marketing term used to describe an automobile with the body of a station wagon that is built on top of a light-truck base frame. Although many consumers believe that SUVs are safer than their sedan or wagon counterparts, studies have shown that SUVs are no safer than any other vehicle on the road. Their increased size, however, has advertisers and manufacturers claiming a higher price. Ford is the third largest vehicle manufacturer in the country; selling 12% of automobiles sold in September 2008 alone.2 In October 2008, Ford sold 128,976 light trucks, second only to GM, which sold 168,785. Despite the current financial climate, Ford still commands a substantial portion of the light truck market. A PBS Frontline video excerpt titled, “Rollover: The Hidden History of the SUV” examines the development of the best selling Sport Utility Vehicle in the world, the Ford Explorer. 3 Released in March 1990, the rise of the Explorer was not without controversy. Frontline’s case study looks at the decision by Ford’s management to continue the launch of the Explorer even though Ford knew both the Explorer and its predecessor, the Bronco II, had failed numerous stability tests. The 1995 products liability case, Denny vs. Ford Motor Company4, examined the actual versus perceived design flaws of the Ford Bronco II. The plaintiff, Nancy Denny, was severely injured in the Bronco II when it rolled over after she hit the brakes to avoid a deer. The Denny’s sued Ford, claiming negligence, strict products liability, and breach of implied warranty. The plaintiffs' claim was that the Bronco II was not fit for its ordinary purpose because of its alleged propensity to rollover and lack of warnings to the consumer about this propensity. The Denny’s offered evidence that suggested the vehicle’s design contributed to the rollover. Ford argued that such a design was necessary for the vehicle’s off-road capabilities, and the vehicle was not defective. The vehicle was designed for off-road use and not designed as a passenger vehicle. In fact, a Ford engineer stated he would not recommend the vehicle for conventional passenger use. The Denny’s responded with a Ford marketing manual, which stated that the vehicles were “suitable to contemporary life styles” and “fashionable” in some suburban areas. In the verdict, the jury found that Ford was negligent in "designing, testing, and marketing the Bronco II," and this negligence was the immediate cause of Nancy Denny's injuries. The jury next concluded that the Ford Bronco II was not “defective” based on the definition of the termI, and therefore not liable on the strict products liability claim. However, on the third claim, the jury found that Ford had breached an implied warranty based on the marketing language written in the manual and Nancy Denny was awarded 40% of the $3 million claim for a total of $1.2 million. Despite the growing number of lawsuits against Ford, they still continue to be a top-ranking vehicle manufacturer in the United States, and the Ford Explorer is one of their most popular vehicles. The purpose of this study is twofold: to understand the factors that define different types of fatal vehicle accidents, and to verify the relationship between the Ford Explorer, rollover occurrences, and fatal accidents. Many studies previously done on Ford have found fault in the fundamental design of the I

According to strict liability instructions, "[a] manufacturer who places a product on the market in a defective condition is liable for injury which results from use of the product when the product is used for its intended or reasonably foreseeable purpose. To prevail on this claim the plaintiffs must prove beyond reasonable doubt that Ford Motor Company placed the Bronco II on the market in a defective condition." The Denny’s failed to prove this to the jury.

Explorer, but this study focuses on the accident, not the vehicle. The data was taken from 2001, ten full years after the rollout of the Ford Explorer. In 2001, 57,9165 vehicles were in fatal accidents, and Ford was involved in 19% of the accidents: the highest of any manufacturer. The BroncoseriesII alone accounted for about 2% of the total number of vehicles involved in fatal accidents, the most of any other Sport Utility Vehicle. Analysis My analysis is based on individual cases collected from the Fatality Analysis Reporting System (FARS) database. The data received from the FARS database included a detailed account of every vehicle involved in a fatal accident in 2001. From looking at Figure 1 (See Results in Appendix), passenger vehicles had the highest number of fatal accidents throughout 2001. This chart is a good representation of the total number of vehicle fatalities in 2001, yet it does not account for the total number of vehicles in use. I created a ratio, defined as vehicle accidents per year per million vehicles in transit. Note that this ratio is restricted in that it uses only the same vehicles, and it is in no means a full measure of risk, only a simplified one. Based on this ratio (Figure 2), motorcycles have the highest risk of fatality. The ratio for motorcycles, about 4.7 times that of passenger vehicles, can be explained by a number of reasons: vehicle design, driver behavior, and relative number of motorcycles in use. The same reasoning can be applied to buses, the lowest ratio. We can assume that the number of buses in use is very small when compared with passenger vehicles. Their size is much larger than motorcycles, and buses are also commonly used in safe driving situations such as taking children to school or public transit: situations where safety is a top priority. Figure 3 clearly shows that light trucks contribute the most to the total number of rollover fatalities. In Rollover Occurrences by Vehicle Manufacturer (Figure 4), Ford has the highest number of rolloverrelated fatalities. Both charts (Figures 3 and 4) support my hypothesis that asserts a relationship between Ford and rollover fatalities, but a factor analysis of the data in its entirety could show that there are other variables contributing to the rollover propensity of a vehicle. In a factor analysis performed in SPSS, the constraints for extraction were defined as eigenvalues over one, and ten components were extracted from the data. As a general rule of thumb, factors with eigenvalues over one should only be used. A scree plot (Figure 5) is another way to determine how many factors should be used in a factor analysis to explain the data. Scree plots show the fraction of total variance in the data as explained by each factor. The scree plot can often show a clear separation in fraction of total variance where the 'most important' components stop and the 'least important' components begin. The point of separation is often called the 'elbow'. My analysis of the scree plot is that five components would be sufficient, however this study will use all ten to explain as much about the data as possible. The rotated component matrix (Figure 6) displays factor loadings for the ten extracted factors that are interpreted in the Conclusion. For the purpose of this study, my analysis focuses on how the data responds to the Rollover variable from the original data. The other variables that are explained in the same factor as Rollover were First Harmful Event (a variable that describes the property damage of the vehicle) and Manner of Collision. This factor is called Rollover Accidents because the first harmful property damage to the vehicle will most likely be an effect of the rollover during the accident. Note that Manner of Collision had a negative factor loading compared with the other three. This negative number suggests that the manner of collision (striking, struck, or both) involves II

Bronco (thru 1977)/Bronco II/Explorer

more than one vehicle, whereas when a rollover accident occurs, it normally involves only one vehicle, and therefore would not include a collision. My analysis concluded with examining the output from two separate MANOVA procedures. In Ftests from the first MANCOVA (See Figure 7 and 8), with p-values > 0.05, we can conclude that First Harmful Event, Manner of Collision, and Rollover Occurrence taken together have a significant effect on Rollover Accidents. This is good. This should happen because these variables defined this factor in the factor analysis. In the second MANCOVA, I plotted the means of Rollover Accidents for each Vehicle Manufacturer. To fully comprehend the meaning of Figure 10II, we must look back at the recoded Rollover coefficients from the factor analysis. A score of 5 corresponds to coefficients -3 < x < -1, and a score of 6 refers to -1 < x < 1. A score above 6 indicates positive rollover coefficients for this manufacturer, and a score below 6 indicates negative ones. As these coefficients approach 6, they become less correlated with Rollover Accidents. Therefore, in Figure 10 we are looking for means that are far from 6. Take, for example, the manufacturer AM General (not included in the graph due to sample size). In 2001, AM General had two fatal crashes, both rollover fatalities. The Rollover Accident coefficient score for AM General was 5.5—the lowest mean of all manufacturers. The main outlier in Figure 10 was Volvo, widely accepted as a ‘safe’ vehicle. When compared with manufacturers with adequate sample sizes, Ford ranks third from the bottom (in front of Mitsubishi and Isuzu). Based on the confidence intervals in SPSS, Ford was not a statistically significant outlier. Conclusions 1. Body Type Comparisons In general, it is extremely difficult to determine the inherent safety of a vehicle type or model because of the difficulty in separating the contribution of driver characteristics and behavior from the contribution of vehicle design. An ideal study would also measure the risk of vehicle models to drivers and non-drivers in order to determine the relative risk of certain vehicle types. The addition of a Risk measurement considers driver accountability.III Much of this study was focused on understanding the factors that describe the types of fatal accidents amongst passenger vehicles and light trucks, including SUVs. By looking at Figure 3, we can see that a rollover in a light truck will significantly increase the passengers’ chance of death. The graph, however, does not answer the question about if driving a light truck will significantly increase your chance of a rollover occurrence. 2. Factor Loadings Driver Ability

Vehicle Size

Pile-Up Accident

Truck Accidents

Rollover Accident

Faulty Driver

Hit and Run Accident

Fire Accident

Speeding Accident

Emergency Vehicle

The chart above represents the ten factors extracted from the factor analysis. In my interpretation, these factors often defined the type of fatal accidents that occurred, not the cause. The first factor was defined by driver-related variables. The ability of the driver and their environment is significantly more important than the relative safety of any vehicle. This analysis does verify that II

This graph used only passenger vehicles and light trucks because consumers don’t purchase buses or semi-trucks for everyday use and motorcycles don’t rollover. III

For example, some car models may attract relatively aggressive drivers, who increase the fatalities in the model, independent of its design. Due to the limitations of statistical analysis, one can only study some general behavioral characteristics of crashes that may correlate with risk.

many of the factors that contribute to fatal traffic accidents are driver-centric. That is, they directly relate back to the driver. A more technical analysis would omit these driver-related variables and observe more aspects of the vehicle that contribute to fatalities. This method, however, would neglect driver behavior, and may ultimately prove inconclusive. As I previously mentioned, this is due to the difficulty in measuring driver behavior. In the second factor, Vehicle Configuration, Number of Axles, Cargo Body Type, and Special Use Vehicles are all variables that define buses, trucks, or large vehicles. Vehicles that are very large or very small often share a lot of the same accident characteristics, so this factor can be explained as the general size of the vehicle. In Pile-Up Accidents, the types of impact are usually the same for most of the vehicles involved, and they share the same impact points as well (front-end or back-end). The variable Number of Vehicles explains the magnitude of the Pile-up accident and the Underride/Override variable explains the intensity. Factor four could be explained as either Truck Accidents or Vehicle Make. According to FARS, Jack Knife occurrences are restricted to semi-trucks and reflect a loss of control by the driver. Vehicle Body type, a variable that also loaded highly on this factor, is also restricted to Trucks and Buses. The high factor loading of the Vehicle Make factor can be explained by the way in which truck data is coded in the FARS database. Many of the truck manufacturers in this data set were separated from their parent company, which could explain this representation in the data. The next factor was addressed earlier in my analysis and is explained by variables that represent Rollover Accidents. Highly loaded factors on First Driver Factor and Second Driver Factor variables explain how the driver contributed to the accident. These are accidents where the driver was mainly at fault for the fatality. Since alcohol consumption is generally at the willing risk of the driver, I included this variable in Factor 6, Faulty Driver Accidents. The next factor can be explained by Hit-and-Run incidents, and includes fatalities where the driver wasn’t present and the victim could be defined as a pedestrian or other non-motorist. In Factor 8, or Fire Accidents, a large fire in a vehicle crash typically renders the deformation of the vehicle as “totaled,” which explains the high factor loadings on both variables. The next factor is Speeding Accidents since the only two variables that highly load onto this factor are Travel Speed and Crash Avoidance. The higher the traveling speed of the vehicle, the less time the driver has to use an avoidance maneuver. The final factor has high loadings on the Emergency Vehicles variable only, so this factor is a measure of vehicles that got into fatal accidents while they were being used in an emergency. 3. Ford and Rollover Propensity This study was able to find clear empirical evidence to support the conclusions made about Ford in the video and in the 1995 Denny vs. Ford Motor Company court case. Although the classic F-tests failed to accept Vehicle Make as an acceptable explanation of Rollover accidents (Figure 9), Ford is still the leader of fatal rollover accidents involving SUVs. We can also conclude that a low mean score (less than 6) in Figure 10 for Rollover Accidents indicates vehicle manufacturers with a higher propensity to rollover. This study serves as a valid basis from which more complicated analysis can be made to determine where the vehicle can be improved in cases where the driver was not at fault. For example, the high number of cases where the vehicle exhibited “overturn” may suggest a widening of the wheelbase and further manipulation of the vehicle specifications. Although it has failed to identify causes of accidents as they relate to the model, this study has clearly identified what types of accidents are killing drivers. It is important to not just look at who (manufacturer) is involved, but also how the accidents are classified. APPENDIX The Data

The data used in my analyses was from the Fatality Analysis Report System (FARS). FARS is the same data source used by the Insurance Institute for Highway Studies (IIHS), and is published by the National Highway Traffic Safety Administration (NHTSA). It is a census of all fatal traffic accidents within the United States and includes a detailed record for every fatal highway accident. The total number of vehicles or case subjects was N = 57,918, and the original number of variables was 75 but reduced to 38 after considering only important, measurable variables. Method The majority of this study focused on a clear approach to functioning data. Creating preliminary graphs (Figures 1 – 4) gave me some evidence to suspect my hypothesis would be supported by the data, but correlations between the variables couldn’t be clearly identified. My analytic approach used a factor analysis since the fundamental design of this study was to decrease the number of variables. I wanted to use a large number of variables to identify the few variables that could help in effectively understanding (not define) the causes of fatal traffic accidents. A factor analysis is a statistical technique applied to a single set of variables to discover which sets of variables form subsets that are independent of one another. Variables that are correlated with one another and also independent of other subsets of variables are combined into factors. These techniques are designed to reduce the number of variables needed to explain the data by using a smaller number of factors.6 First, a correlation matrix was created to examine the multicollinearity of the variables before I could perform a factor analysis. Multicollinearity (also known as singularity) occurs when two or more independent variables in a multiple regression model are highly correlated. In a singular correlation matrix, the factors may change unpredictably in response to small changes in the model or the data. The first time this was done, many factors had correlation coefficients greater than .9, a sign of singularity. I specifically looked at the variables that counted the number of previous accidents, DWI convictions, moving violations, speeding violations, and license suspensions. By creating a single variable for the instances, PREV_TROUBLE, I was able to collapse those four variables into one. After each new variable was created, another correlation matrix was made and then analyzed for singularity. This happened many times as the data set began with 75 variables. When a sufficient set of 38 variables was determined, ten factors with eigenvalues greater than 1 were extracted using a factor analysis with a Varimax rotation in SPSS. Then, these ten factors were saved as variables so that each subject in the original data set had a corresponding coefficient for each of the ten extracted factors. This process, known as ordination, is a way to summarize the variation in all of the variables and make it possible to display the data graphically. In order to generate a graphical representation of Ford’s response to Rollover Accidents, I performed a Multiple Analysis of Variance (or MANOVA) in SPSS. MANOVA is used to see the effects of categorical variables (such as Vehicle Make) on multiple dependent interval variables (such as the recoded factor scores for Rollover Accidents). The specific method I used was called MANCOVA (Multiple Analysis of Covariance) because it uses categorical independent variables (Rollovers, Manner of Collision, and First Harmful Event) as predictors. This technique was done twice. The first time was to verify the results of the factor analysis, and thereby supporting the use of Rollover Accidents as a dependent variable. The second time was to observe Ford’s response to the dependent variable, Rollover Accidents. Results Figure 1.

Number of Fatalities that Occurred in Each Vehicle Type Passenger Car (automobile) Light Truck Motorcycle Truck* Other or Unknown

359 653 3,402

11,689

20,233

*Truck is defined as a combination truck (Tractor Trailer) or a single unit with 2 axles and 6 tires or more

Figure 2. Fatal Accidents per Million Motor Vehicles in Use 800

694 600

400

200

147

139 83

40

0

Passenger car Light Truck

Motorcycle

Truck

Bus

Figure 3. Number of Rollover Occurrences by Vehicle Model 12,000

11,156

First Event

Subsequent Event

10,000

8,000

6,000

4,293 4,000

2,900

2,000

216

54

0

Passenger Car

Figure 4.

Light Truck

Motorcycle

Truck

Bus

Number of Rollover Occurrences by Vehicle Manufacturer First Event

Subsequent Event

3,000 2,431

2,500

2,325

2,000

1,500

1,000 664

639 423

500

413

375

368 328

0

Ford

Chevrolet

Dodge

Toyota

Pontiac

GMC

Nissan

Honda

Jeep

Figure 5. Scree Plot

6

4

Eigenvalue 2

0 1

2

3

4

5

6

7

8

9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36

Component Number

Figure 6. Rotated Component Matrixa Previous Violations License Compliance Previous Accidents Driver License State Out of State Driver Registered Vehicle Driver License Status Vehicle Configuration Number of Axles Cargo Body Type Special Use Vehicle Number of Occupants Type of impact Number of Vehicles Under ride / Override Principal Impact Point Leaving the Scene Vehicle Make Jack-Knife Vehicle Body Type First Harmful Event Rollover Involved Manner Of Collision First Driver Factor Second Driver Factor Alcohol Involvement Hit and Run Incident Driver Presence Vehicle Factors Vehicle Deformation Fire Occurrence Number of Fatalities Crash Avoidance Travel Speed Emergency Vehicle Vehicle Maneuver

1 .812

2

3

5

6

7

8

9

10

.752 .666 .569 .540 .478

.365

.307

.451

-.342 .770

.352

.692

.331

.690 .656

.412

.536 .738 .625 .351

.590 .462 .451 .781 .764 .748 .710 .537 .417

-.670 .685 .682 .475

.345

.394 .649 .611 .427 .679 .361

.391

.582 .512 .769 .686 .960

.354

.387

Extraction Method: Principal Component Analysis. Rotation Method: Varimax with Kaiser Normalization. a. Rotation converged in 20 iterations.

Figure 7.

4

Multivariate Testsd,e Effect

Value

HARM_EVENT * ROLLOVER * MAN_COLL

Pillai's Trace Wilks' Lambda Hotelling's Trace Roy's Largest Root

F

.000 1.000 .000 .000

.966 .966a .966 1.696c

Hypothesis df 6.000 6.000 6.000 3.000

Error df 114748.000 114746.000 114744.000 57374.000

Sig. .446 .446 .446 .166

Figure 8. Tests of Between-Subjects Effectsd Source HARM_EVENT * ROLLOVER * MAN_COLL

Dependent Variable FAC5

Type III Sum of Squares 8.571

df 3

Mean Square 2.857

F

Sig.

.400

.753

Figure 9. Multivariate Testsc Effect MAKE

Value F Hypothesis df Pillai's Trace .062 24.408 150.000 Wilks' Lambda .939 24.564a 150.000 Hotelling's Trace .065 24.720 150.000 Roy's Largest Root .053 40.449b 75.000 c. Design: Intercept + ROLLOVER + MAN_COLL + HARM_EVENT + MAKE

Error df 114920.000 114918.000 114916.000 57460.000

Sig. 0.000E+00 0.000E+00 0.000E+00 0.000E+00

Figure 10. Means of Rollover Accidents for Vehicle Manufacturers Sorted Least Rollover-Prone to Most Rollover-Prone

6.3

Volvo 6.229 6.2

6.1

Observed Means Hyundai 6.063 6.0

5.9

Ford 5.963

Isuzu 5.931

1

REFERENCES U.S. Department of Commerce, Bureau of Economic Analysis, Underlying Detail for the National Income and Product Account Tables. As of Oct. 24, 2008. Internet site: 2 Automotive News Data Center, U.S. Monthly Sales (according to sales volume), As of October 24, 2008. Internet Site: http://www.autonews.com/section/DATACENTER. 3 Rollover: The Hidden History of the SUV. PBS Frontline. 4 Cornell Law School. Nancy Denny Et Al., Plaintiffs, v. Ford Motor Company, Defendant. 87 N.Y.2d 248, 662 N.E.2d 730, 639 N.Y.S.2d 250, December 5, 1995 5 National Highway Traffic Safety Administration. FARS Database, FARS FTP Files. 2001: DBS: FARS 2001.zip. 6 Manly, Bryan FJ. Multivariate Statistical Methods: A Primer. 12-13.