Abstract
Objective:
Although driving simulators are powerful tools capable of measuring a wide-ranging set of tactical and operational level driving behaviors, comparing these behaviors across studies is problematic because there is no core set of driving variables to report when assessing driving behavior in simulated driving scenarios. To facilitate comparisons across studies, researchers need consistency in how driving simulator variables combine to assess driving behavior. With inter-study consistency, driving simulator research could support stronger conclusions about safe driving behaviors and more reliably identify future driver training goals. The purpose of the current study was to derive empirically and theoretically meaningful composite scores from driving behaviors of young people in a driving simulator, utilizing driving data from across a variety of driving environments and from within the individual driving environments.
Method:
One hundred ninety adolescent participants aged 16 years or 18 years at enrollment provided demographic data and drove in a high-fidelity driving simulator. The simulated scenario included 4 distinct environments: Urban, Freeway, Residential, and a Car Following Task (CFT). A Principal Components Analysis (PCA) was conducted on the variable output from the driving simulator to select optimal factor solutions and loadings both across the multi-environmental drive and within the four individual driving environments.
Results:
The PCA suggested two components from the multi-environmental simulated drive: vehicle control and speed. The individual driving environments also indicated two components: vehicle control and tactical judgment.
Conclusion:
These findings are among the first steps for identifying composite driving simulator variables to quantify theoretical conceptualizations of driving behavior. Currently, driving behavior and performance measured by driving simulators lack “gold standards” via driving scores or benchmarks. The composites derived in this analysis may be studied for further use where driving behavior standards are increasingly sought by clinicians and practitioners for a variety of populations, as well as by parents concerned about the readiness of their novice driving teen.
Keywords: driving simulator, teen driving, principal components, novice drivers
INTRODUCTION
Driver Behavior Models – Michon
Unsafe driving behaviors are often measured in examinations of the elevated crash risk in newly licensed drivers. No driving behavior models currently exist for simulated driving, but existing real-world driving behavior models may inform hypotheses for the development of conceptualizing simulated driving outcomes. Specifically, although standards for interpreting driving outcomes from a driving simulator do not currently exist, models conceptualizing real-world driving behavior may aid in the interpretation and terminology of statistically derived components from driving simulator data. Driving behaviors are in service of a larger goal: convenient, efficient, and safe transportation. The behaviors in which drivers engage to achieve that goal can be organized into three levels as posed by Michon (1985): strategic (planning), tactical (maneuvering), and operational (control). The strategic level is closely related to the overall goal of driving: drivers engage in strategic behaviors when they decide on destinations, plan routes, and check weather or traffic along that route. Almost any decision about driving that happens before the vehicle starts moving is a strategic one. Decisions drivers make while behind the wheel are primarily tactical. A driver’s intentions for how their car should maneuver (e.g., speed, following distance, lane choice, overtaking other vehicles) are defined by tactical-level behaviors, or tactical judgment. However, the extent to which tactical intentions are realized is dependent upon the driver’s ability to control their vehicle. That ability is captured by operational-level behaviors. The physical actions required to make a car accelerate, decelerate, reverse, and turn are all examples of operational-level driving behaviors (Michon 1985). Operational-level behaviors are among the first skills acquired when people learn to drive and are ideally performed with minimal cognitive effort by the time a driver is licensed.
Notably, the strategic, tactical, and operational levels encompass all driving behaviors, both safe and unsafe. For example, choosing to drive during adverse weather conditions can be part of a driver’s strategy to avoid traffic, although such a strategy may increase crash risk. Tailgating does not reduce travel time, but drivers employ the tactic because they believe it provides desirable results. Unsafe strategic and tactical choices may indicate a lack of experience or a greater acceptance of risk, both of which may be exhibited by young and/or newly licensed drivers (McGwin and Brown 1999; Borowsky et al. 2007; Lee 2007). At the operational level, newly licensed drivers have typically mastered the basics of controlling a car, but they may not be as adept at reacting to the complex, dynamic scenarios that can occur on the road (Durbin et al. 2014). However, there are currently neither standard methods of empirically measuring these skills nor benchmarks for evaluating expertise.
Measuring Tactical and Operational Behaviors
Driving simulators provide an immersive, realistic, and safe environment to study driving behavior using real-world driving situations. In addition to the experimental control afforded to study designs, driving simulators provide a prodigious amount of driving data and, thus, a variety of ways to quantify driving behavior in diverse populations. Studies utilizing driving simulators may measure and report driving variables related to specific aspects of driving, such as within vehicle kinematics (e.g., speed, steering, and acceleration (Maxwell et al. 2021), spatial maneuvering and reaction (e.g., lane positioning or deviation (McManus et al. 2017), gap acceptance (Beck et al. 2007), reaction time (Karimi et al. 2020), and collisions (Stavrinos et al. 2013).
The computational power of driving simulators provides a wealth of scenario design options and data. The range of driving-related variables available may vary among driving simulators and may contribute to the variety of ways in which driving behavior has been measured. Although driving simulators are powerful tools capable of measuring a wide-ranging set of tactical and operational level driving behaviors, comparing these behaviors across studies is problematic because there is no core set of driving variables to report when quantifying driving behavior in simulated driving scenarios. An understanding of how various individual driving variables group together may provide greater insight into driving behaviors, particularly with respect to maneuvering and control, and help facilitate the development of standard composite scores quantifying driving behaviors.
Variability Across Studies Using Driving Simulators
Driving simulators can vastly differ in the driving variables they measure or record, as well as how they calculate those variables. Additionally, even among studies that use similar simulators, the reported driving outcomes often differ. This inconsistency limits conclusions drawn from comparing results across driving simulator studies. Formulating a consistent set of driving simulator variables, as well as a consistent, empirically, and theoretically supported method of quantifying driving behavior, could result in stronger theories of safe driving behavior, more reliable assertions to facilitate relationships between behaviors, and help driver training goals that could inform public policy. Some approaches have attempted to address these issues by assessing driving performance as an unobserved factor (i.e., latent variable) (Papantoniou 2018). Papantoniou’s (2018) work was an important step in computing a composite “score” for diving behavior, but the primary goal was to estimate the effect of driving distraction, among other factors, on overall driving performance while assessing the feasibility of using Structural Equation Modelling (SEM) approaches with simulator data. Additionally, the inclusion of distraction, driving environment, and driver characteristics as factors in the analysis to estimate driving performance as a latent variable reduces the potential standardization of driving scores for other simulator studies that may not include such factors. To date, no studies have developed composite scores that combine individual driving simulator variables into theoretically supported driving behavior components that can be generalized across studies and across driving simulator designs that may vary with study (e.g., driving environments, weather, traffic conditions).
Purpose of Current Study
This purpose of the current study was to derive empirically meaningful composite scores from driving behaviors of young people in a driving simulator, utilizing driving data from across a variety of driving environments as well as within each driving environment. Although the derivation of these composite scores is largely exploratory, this work will consider how the composites align with real-world driving behavior theories. This work aims to provide standards in how to assess driving performance in a driving simulator. Rather than examine if composite scores fit existing models of driver behavior, such as Michon’s, this study will use Michon’s terminology to aid in conceptualizing the composites that emerge. This work also builds upon Papantoniou’s (2018) goal of producing driving behavior indices beyond specific driving variables but distinguishes itself by aiming to derive components only through the driving simulator variables and assessing these components across various driving environments to maximize generalizability and initiate the process of standardization across simulator studies.
METHOD
Study Design and Sample
Data were collected from one hundred ninety participants (n = 190) enrolled in the longitudinal Roadway Environments and Attentional Change in Teens (REACT) study examining the effects of age and driving experience on driving attention development (Whittington et al. 2020). Participants were recruited at either 16 years of age (n = 113, 59%) or 18 years of age (n = 77, 41%). Participants were enrolled either within two weeks of licensure (n = 81, 43%) or with neither a driver’s license nor intent ever to obtain one (n = 109, 58%). Informed consent was obtained from all participants aged 18 years, and informed assent of participants aged 16 years was obtained with their parent/guardian’s consent. All study procedures were approved by the University of Alabama at Birmingham’s Institutional Review Board.
Measures and Materials
Driving simulator:
Participants drove in a Realtime Technologies, Inc., high-fidelity driving simulator to provide indices of driving. No participants had prior exposure to this driving simulator, or the scenarios utilized herein. This fully immersive driving simulator was a 2016 Honda Pilot with fully functioning steering wheel, throttle, brake, gear selector, turn signals, and dashboard. The simulator was on a 1-degree-of-freedom motion base allowing pitch cues for accelerating and braking. The visual system consisted of three 80” projection screens providing a 180° field of view, as shown in Figure 1. Drivers could view the simulated environment behind the vehicle through LCD displays in the side mirrors and through the center rearview mirror, which faced a large screen behind the cab. The driving simulator recorded data at 60 Hz.
Figure 1.
Realtime Technology Inc. Driving Simulator
Simulated driving scenario:
A 7-mile driving scenario was developed to consist of three distinct environmental sections (Urban, Freeway, and Residential) and a car following task (CFT), presented in a randomized order. An overhead view of the map is displayed in Figure 2 that depicts a route beginning in the Urban section, then traversing to the Freeway section, then CFT, and ending with the Residential section. The Urban section consisted of several intersections with traffic signals. The traffic signals were all green in the direction the participant was going to maintain the flow of their driving. The area surrounding the road had large city buildings and pedestrians on the sidewalks. The Freeway section consisted of two long, curved segments and one long, straight road segment. The roadway had two lanes, and the surrounding area contained grass and trees. Ambient traffic was present in the Freeway and Urban sections but not the Residential section. The Residential section consisted of a straight, two-way, two-lane road. Driveways connected to residential homes lead from the road on the left and right. Green grass, trees, and bushes separated the homes. The posted speed limits of the freeway, residential, urban, and CFT were 70 miles per hour (mph), 35 mph, 35 mph, and 45 mph, respectively. Participants were instructed not to take any turns within an environmental section (e.g., intersections within the Urban environment). Participants were instructed to turn only as necessary to go from on environmental section to another (e.g., turning onto an on-ramp to get from Urban to Freeway).
Figure 2.
Overhead view of Driving Simulator Scenario Map with an example depicting an Urban, Freeway, Car Following Task, and Residential environment route with arrows.
The 0.75-mile CFT, adapted from prior research (Strayer et al. 2011), was set in a no-passing zone on a flat, two-lane road (See Figure 2). The participant was instructed to follow and remain behind a lead car (moving at 45 mph). While on a straight portion of the road, the lead car braked intermittently on five different occasions despite having no external provocation, requiring the participant to brake to avoid a collision. After the lead vehicle engaged the brakes two times, it entered a single curve during which there were no braking episodes followed by another straight road segment. On this straight segment, the lead vehicle engaged the brakes three more times. No ambient traffic was present during the CFT. Prior to beginning the experimental scenario, participants completed a 4.5-mile practice drive which contained each experimental environment and the CFT to demonstrate their physical ability to operate the driving simulator. As the data utilized in this work represent the first exposure participants had to the driving simulator and experimental driving scenarios, this practice drive standardized the amount of experience with the driving simulator across participants.
Data Analyses
Descriptive statistics were examined for all simulator drive variables to review amounts of missing data and variable distributions. Measurements and operationalization of driving simulator variables are provided in Table 1. As the driving simulator records data at 60 Hz, instances of “standard deviation” are also conceptualized as the root mean square of successive differences (RMSSD) from data line to data line.
Table 1.
Driving Simulator Outcomes and Definitions
| Variable | Definition |
|---|---|
|
| |
| Mean speed (m/s) | Average speed |
| Speed variability (m/s) | Standard deviation of speed |
| Mean acceleration (m/s2) | Average acceleration of speed |
| Acceleration variability (m/s2) | Standard deviation in acceleration of speed |
| Steering variability (radians) | Standard deviation of the steering wheel position |
| Lane variability (meters) | Standard deviation of distance from the center of lane |
| Collisions (n) | Number of collisions with another vehicle or object |
| Mean reaction (RXN) to lead car braking (seconds)1 | Average time elapsed from the start of the lead car braking until the consecutive brake value difference equals or exceeds 1 pound of force |
| RXN variability to lead car braking (seconds)1 | Standard deviation of Mean RXN to lead car braking |
| Mean recovery to lead car speed1 | Average time elapsed for the participant to achieve lead car’s speed (45 mph) after a braking event |
| Recovery variability1 | Standard deviation in recovery time |
Note. m/s = meters per second, n = number, RXN = Reaction time, superscript indicates the variable was measured in the CFT only.
Only variables that were present for at least 90% (172 of 190) of participants and had some variability were selected for further analyses. The data from the remaining 10% (n = 18) were not analyzed due to simulator sickness preventing the completion of the entire simulated driving protocol (n = 2) and failure to adhere to instructions (e.g., taking uninstructed turn) that rendered simulator data less comparable (e.g., taking additional routes, turns, and maneuvers to be redirected to the appropriate route in the simulator). Each variable was first evaluated for outliers and normal distribution, and its distribution was corrected through a natural log transformation, inverse transformation, or truncating outliers, depending on which approach best improved the distribution. The driving variables were analyzed in two groups, multi-environmental drive and individual environments (Urban, Freeway, Residential, and CFT). These two groups of variables were analyzed separately because their observations were not independent of one another (i.e., the individual environments composed the multi-environmental drive). Additionally, given the different environmental elements within each environment segment (e.g., visual stimuli, speed limit), the individual environments may evoke different driving behaviors. For each group of variables, the assumption of linear relationships among the variables was checked through bivariate scatterplots. Then, factorability was evaluated by examining correlations, Bartlett’s test of sphericity, and Kaiser-Meyer-Olkin Measure of Sampling Adequacy (KMO). Next, principal component analysis (PCA) was conducted, with the number of suggested factors to retain evaluated with the Kaiser criterion, scree plot, parallel analysis, and Velicer’s MAP test (O’Connor, 2000). The suggested numbers of factors were then retained and rotated with an oblimin rotation. If all inter-factor correlations fell below 0.30, varimax rotation was used instead to simplify the solution. Based on the sample size, factor loadings above .40 were considered practically significant (Hair, 1998). In case of an insufficient number of marker variables on a principal component, an additional solution with fewer factors was examined. After selecting the most optimal solution, dimension scores were computed as averages of standardized variables loading above .40 on each component, excluding variables with complex loadings and variables whose loadings did not reach this cut-off. Variables with negative loadings were reverse-coded. The averages of standardized variables were used instead of factor scores because they are simpler to interpret and can be more easily reproduced in future research. Internal consistency of each dimension score was evaluated with Cronbach’s alpha, and if indicated, the dimension scores were further modified to improve internal consistency. Sensitivity analyses repeated the PCA analyses on drivers only.
RESULTS
Participant Demographics
Participants were mostly Black or African American (n = 138, 73%), and approximately half were female (n = 101, 53%). Licensure groups differed by age as recruited (16 vs. 18), and Race was associated with licensure status (χ2 = 18.63, p<.01). See Table 2 for a full description of the sample stratified by age and licensure groups.
Table 2.
Participant Demographics by Age and Licensure Categories
| Variable | Licensed | Unlicensed | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
|
|
|
|||||||||
| 16 (n = 58) | 18 (n = 23) | 16 (n = 55) | 18 (n = 54) | |||||||
|
|
|
|||||||||
| M (SD) | n (%) | M (SD) | n (%) | M (SD) | n (%) | M (SD) | n (%) | F or χ2 | p | |
|
| ||||||||||
| Age (years) | 16.17a (0.21) | 18.31b (0.35) | 16.39c (0.33) | 18.42b (0.31) | 807.23 | <0.01 | ||||
| Female | 33 (53%) | 15 (65%) | 28 (51%) | 27 (50%) | 1.67 | 0.64 | ||||
| Race | 18.63 | <0.01 | ||||||||
| Black | 27 (47%) | 19 (83%) | 47 (85%) | 45 (83%) | 17.82 | <0.01 | ||||
| White | 26 (45%) | 1 (4%) | 7 (13%) | 4 (7%) | 15.69 | <0.01 | ||||
| Other | 5 (9%) | 3 (13%) | 1 (2%) | 5 (9%) | 1.30 | 0.25 | ||||
Note. F and χ2 indicate differences among the 4 stratifications of age and licensure. Different superscripts indicate Bonferroni corrected pairwise differences
Multi-Environment Drive Variables
Seven overall drive variables had data for 187 participants: mean acceleration, acceleration variability, collisions, lane variability, mean speed, speed variability, and steering variability (see Table 3). All variables were not normally distributed (absolute skewness 0.54 to 5.75, absolute kurtosis 0.55 to 46.28), but distributions improved substantially after transformations or truncating of outliers (skewness 0.00 to 0.56, kurtosis 0.04 to 0.55, except for number of collisions which had skewness of 2.39 and kurtosis of 3.77). Specifically, outlier truncation best improved the distribution of mean acceleration (to −0.90 to 1.40), number of collisions (0 or 1), and mean speed (to 14.70 to 25.20); natural log transformation best normalized lane variability and speed variability; and inverse transformation best normalized acceleration variability and steering variability. These transformed variables were then used for all subsequent analyses.
Table 3.
Descriptives and Principal Component (PC) Analysis Factor Loadings for Multi-environmental Drive Variables
| M (SD) | PC1 Vehicle Control Variability | PC2 Speed | |
|---|---|---|---|
| Acceleration variability | 21.77 (48.37) | .89 | −.17 |
| Steering variability | 2.45 (1.54) | .77 | −.12 |
| Lane variability | 0.50 (0.15) | −.78 | −.08 |
| Collisions a | 0.15 (0.51) | −.42 | −.47 |
| Mean speed | 20.05 (2.95) | −.28 | .76 |
| Speed variability | 6.98 (1.42) | −.32 | .72 |
| Mean acceleration a | 0.85 (13.29) | .14 | .45 |
Note: M (SD) before transformations. Varimax rotation was used. Loadings above .40 are bolded. N=187.
Variable was not included in dimension scores.
Absolute correlations among these variables ranged from .00 to .71 and examination of scatterplots confirmed linear relationships among the variables. Factorability was supported by a significant Bartlett’s test (χ2(21) = 331.86, p <.001), and sampling adequacy was adequate but low (KMO = 0.62). The PCA yielded two eigenvalues greater than 1 (2.51, 1.42). The scree plot and parallel analysis suggested the retention of two components; Velicer’s MAP test suggested retaining one. Based on these results, two components were retained, explaining 56.22% of the total variance. Using oblimin rotation, the inter-factor correlation was 0.02, so the analysis was repeated with the orthogonal varimax rotation. Acceleration variability, steering variability, and lane variability (reverse direction) loaded on Component 1, which was labeled “vehicle control variability” to reflect the direction of the loadings (i.e., high variability in both acceleration and steering indicate greater variation in vehicle control; see Table 3 for loadings). Mean speed, speed variability, and mean acceleration loaded positively on Component 2, which was labeled “speed.” Number of collisions loaded on both components, and therefore was not included in the dimension composites. Using standardized variables, Cronbach’s alphas (α) for the two dimension scores were .80 and .48, respectively. However, with the exclusion of average acceleration, the Cronbach’s α for speed improved to .70. Thus, the final dimension scores for multi-environmental vehicle control variability were computed as the average of standardized acceleration variability, steering variability, and reversed lane variability. Multi-environmental drive speed was computed as the average of standardized mean speed and speed variability. As noted earlier, the dimension scores were computed as averages of standardized variables rather than factor scores to facilitate interpretability and replicability in future research of these dimension scores. Sensitivity analysis on drivers only yielded a similar PCA solution.
Individual Environment Variables
Thirty-one variables were analyzed for the four driving environments (Urban, Freeway, Residential, and CFT). All four yielded 6 common variables: mean acceleration, acceleration variability, collisions, lane variability, mean speed, speed variability, and steering variability. The CFT provided four additional variables: mean recovery to lead car speed, recovery variability, mean reaction (RXN) to lead car, and RXN variability. Finally, the Freeway, Residential, and Urban environments each included a measure of lane variability. Between 177 and 181 participants had valid data for these 31 variables. Examination of variable distributions showed no variability in Residential collisions (all values were 0); thus, this variable was removed from further analyses.
Most variables were not normally distributed (absolute skewness 0.05 to 11.06, absolute kurtosis 0.26 to 136.28), but distributions improved substantially after transformations or truncating of outliers (skewness 0.00 to 1.15, kurtosis 0.09 to 0.78, except for number of collisions which had skewness of 4.35 to 9.19 and kurtosis of 17.08 to 83.45). Specifically, no transformation was needed for car following mean speed and speed variability, freeway mean speed, mean acceleration, and lane offset variability; log transformation was used for freeway speed variability and steering variability, residential acceleration variability, speed variability, and lane offset variability, and urban speed variability; and inverse transformation was used for freeway acceleration variability. All remaining variables had outliers truncated; numbers of collisions were truncated to 0 vs. 1 or more. These transformed variables were then used for all subsequent analyses.
Correlations among the remaining 30 variables ranged from −.90 to .90 (from .01 to .90 in absolute values) and linear relationships among the variables were confirmed by examination of scatterplots. Factorability of the correlation matrix was supported by a significant Bartlett’s test (χ2(435) = 3767.00, p < .001) and sufficient sampling adequacy (KMO = .77). The PCA yielded 9 eigenvalues greater than 1, ranging from 8.14 to 1.07. The scree plot and parallel analysis suggested retaining three components and Velicer’s MAP test suggested two. Based on these results, both two- and three-component solutions were examined further. These solutions accounted for 40.13% and 47.05% of the total variance, respectively. Using oblimin rotation, all inter-factor correlations fell below .15, so both analyses were repeated with the orthogonal varimax rotation.
Examination of factor loadings (Table 4) indicated the presence of multiple complex loading variables in each solution. Specifically, four variables had complex loadings in the two-component solution and six variables had complex loading in the three-component solution (five between PC1 and PC2, one between PC2 and PC3). Based on the solutions’ simplicity and theoretical considerations (i.e., clearer grouping of CFT variables), the two-component solution was selected as the best solution. Excluding variables with complex loadings (i.e., split across both components), PC1 included 13 variables across acceleration variability, lane variability, steering variability, speed variability, and number of collisions on the CFT. This component was named “vehicle control variability” to be consistent with the direction of its constituent variables (i.e., greater variability in acceleration, lane, steering, and speed; number of collisions). PC2 included average speed on freeway and the CFT and negatively loading mean RXN to lead car braking, RXN variability to lead car breaking, and recovery on the CFT. This component was named “tactical judgment.” Seven variables did not have sufficiently high loadings on either component: mean recovery to lead speed on the CFT; mean acceleration on the CFT and in Residential, Freeway, and Urban areas; and collisions in Freeway and Urban areas.
Table 4.
Descriptives and Principal Component (PC) Factor Loadings for Individual Driving Environment Variables
| M (SD) | PC1 | PC2 | PC1 | PC2 | PC3 | |
|---|---|---|---|---|---|---|
| Acceleration Variability, Residential | 6.40 (6.85) | 0.86 | 0.21 | 0.84 | 0.31 | 0.06 |
| Lane Variability, Residential | 0.36 (0.15) | 0.81 | −0.10 | 0.81 | 0.02 | 0.16 |
| Steering Variability, Residential | 9.97 (3.9) | 0.75 | 0.35 | 0.72 | 0.42 | 0.00 |
| Acceleration Variability, Urban | 2.26 (3.68) | 0.74 | 0.04 | 0.75 | 0.00 | −0.24 |
| Lane Variability, Urban | 0.32 (0.18) | 0.71 | −0.23 | 0.73 | −0.20 | 0.00 |
| Lane Variability, Freeway | 0.51 (0.16) | 0.71 | −0.04 | 0.71 | 0.03 | 0.07 |
| Acceleration Variability, CFT | 5.79 (5.76) | 0.66 | 0.24 | 0.63 | 0.32 | 0.05 |
| Speed Variability, CFT | 3.77 (1.23) | 0.64 | 0.03 | 0.63 | 0.10 | 0.04 |
| Acceleration Variability, Freewaya | 10.74 (4.19) | −0.59 | −0.54 | −0.55 | −0.58 | 0.08 |
| Steering Variability, CFTa | 10.26 (9.17) | 0.57 | 0.50 | 0.52 | 0.58 | −0.01 |
| Steering Variability, Urban | 2.07 (4.67) | 0.55 | −0.10 | 0.59 | −0.21 | −0.35 |
| Speed Variability, Freeway | 2.71 (1.31) | 0.51 | −0.07 | 0.52 | −0.07 | −0.07 |
| Speed Variability, Residential | 3.06 (1.83) | 0.49 | 0.24 | 0.49 | 0.18 | −0.28 |
| Speed Variability, Urban | 3.10 (1.76) | 0.49 | 0.17 | 0.48 | 0.17 | −0.11 |
| Mean Speed, Residentiala | 16.98 (3.43) | 0.47 | 0.55 | 0.41 | 0.66 | 0.12 |
| Steering Variability, Freewaya | 7.12 (1.95) | 0.47 | 0.65 | 0.42 | 0.69 | −0.11 |
| Collisions, CFTa | 0.04 (0.25) | 0.41 | 0.04 | 0.41 | 0.05 | −0.07 |
| Mean Speed, Urbana | 16.80 (3.79) | 0.43 | 0.44 | 0.38 | 0.53 | 0.07 |
| RXN Variability to Lead Car Braking, CFT | 2.71 (3.33) | 0.19 | −0.86 | 0.23 | −0.74 | 0.42 |
| Mean RXN to Lead Car Braking, CFT | 3.54 (2.1) | 0.02 | −0.84 | 0.06 | −0.72 | 0.45 |
| Mean Speed, CFT | 16.00 (1.93) | 0.06 | 0.83 | −0.01 | 0.85 | −0.12 |
| Mean Speed, Freeway | 28.96 (3.72) | 0.21 | 0.74 | 0.15 | 0.80 | −0.04 |
| Recovery Variability, CFT | 2.21 (1.18) | −0.03 | −0.51 | −0.05 | −0.23 | 0.82 |
| Mean Recovery to Lead Speed, CFTa | 1.52 (0.63) | −0.02 | −0.21 | −0.06 | 0.07 | 0.75 |
| Mean Acceleration, Freewaya | 0.66 (1.77) | 0.05 | 0.07 | 0.08 | −0.08 | −0.42 |
| Mean Acceleration, CFTa | 0.34 (0.43) | −0.01 | 0.39 | −0.02 | 0.27 | −0.40 |
| Collisions, Urbana | 0.01 (0.11) | 0.05 | −0.21 | 0.07 | −0.20 | 0.05 |
| Mean Acceleration, Urbana | −0.02 (0.25) | 0.01 | −0.12 | 0.01 | −0.04 | 0.23 |
| Collisions, Freewaya | 0.04 (0.21) | −0.03 | −0.17 | −0.03 | −0.10 | 0.23 |
| Mean Acceleration, Residentiala | 0.34 (0.64) | −0.10 | 0.07 | −0.12 | 0.09 | 0.08 |
Note: M (SD) before transformations. Varimax rotation was used. Loadings above .40 are bolded. N=177–187.
Variable was not included in dimension scores.
Based on conceptual considerations, collisions on the CFT were dropped from PC1 prior to computing the dimension scores. Excluding collisions did not change the scale’s reliability of Cronbach’s alphas of .89. Thus, the PC1 individual driving environment “vehicle control variability” score was computed as the average of 12 standardized variables indexing variability in acceleration, steering, and lane positioning across the individual driving environments. PC2 “tactical judgment” score was computed as the average of five standardized variables, including average speed on Freeway and CFT, and reverse-coded mean RXN to lead car braking, RXN variability to lead car breaking, and recovery on the CFT, with Cronbach’s alpha of .85. Consistent with the multi-environmental drive composites, the individual drive dimension scores were computed as averages of standardized variables rather than factor scores to facilitate interpretability and replicability in future research of these composite scores. Sensitivity analysis on drivers only yielded a similar PCA solution.
DISCUSSION
This study analyzed the commonalities among driving simulator variables to understand the association between these simulated indices of driving and existing theories of driving behavior. The results identified two composite variables in the multi-environmental drive: Vehicle control variability and speed. Two compositive variables were identified in the individual environments: Vehicle control variability and tactical judgment. These findings suggest there may be theoretically and empirically meaningful composites of individual driving simulator variables that can relate to theoretical conceptualizations of driving behavior. Currently, driving behavior and performance measured by driving simulators lack “gold standards” via driving scores or benchmarks. However, these are increasingly sought by many clinicians and practitioners in a variety of populations (e.g., returning to drive after traumatic brain injury), as well as parents concerned about the readiness of their novice driving teen.
The results suggest driving simulator variables can be grouped into driving behaviors aligning with Michon’s (1979) conceptualization of operational (control) and tactical (maneuvering) driving. It should be reiterated, however, that this work aimed to provide empirically derived driving behavior groupings derived from driving simulator variables and Michon’s terminology and model were used to provide a broad guiding framework in how to conceptualize the derived components. Among simulator variables evaluated across the overall drive, those reflective of variability (steering, acceleration, and lane positioning) loaded together and indicate basic driving skill in operational control of the simulated vehicle. Meanwhile, both variables directly related to speed, including average speed and speed variability, loaded together. The distinguishing factor differentiating the control from maneuvering variables is risk taking (Michon 1979). In this study, participants were instructed to drive as if in a real car on a real road, but they were not explicitly instructed to maintain the posted speed limits. Thus, the driving speed was a tactical choice exhibited by the participants, and the speed component derived from the PCA is reflective of the tactical level of driving behavior. It should be noted that operational control of the simulated vehicle was weakly associated with driving speed as indicated in bivariate correlations, yet the PCA distinguished these as two separate factors.
The consideration of the simulated driving variables with respect to individual environments (Urban, Freeway, Residential, and CFT) elucidated a finer grain of commonalities among the numerous simulator variables. Interestingly, the most optimal loadings of the driving variables into two components grouped speed variables (operationalized as tactical without respect to environments) with the operational control variables. Here, the CFT variables clearly loaded as a factor apart from these variables now operationalized as vehicle control. These loadings likely indicate how distinguished the CFT was from the rest of the simulated driving scenario. Given the task demands of the CFT (Strayer et al. 2011), the following distance and speed largely influence the reaction times, both of which are at the discretion of the driver. These CFT variables similarly reflect risk taking and thus tactical driving judgment and behavior.
Tactical and operational driving skills develop along distinct trajectories in the driver learning phases, and inexperienced drivers may benefit from training that targets these levels separately. A strength of this study was to include both novice drivers (within 2 weeks of licensure) and unlicensed drivers to examine how driving variables may form composite factors. Additionally, the inclusion of unlicensed drivers allowed the examination of if driving behavior, and subsequent PCA grouping of simulator variables, would be dependent upon real-world driving exposure. As many current and future driving simulator researchers may aim to use driving simulator platforms to assess the use of such simulators as training tools to non-drivers, it is important to consider the potential effect of real-world driving experience. With improvements in using combinations of commonly measured driving simulator variables to quantify tactical and operational levels of driving skill and/or behavior, these driving scores may lead to and provide comparable benchmarks to be used in driving training.
Limitations
No study is without limitations, and a few are discussed herein. Although the sample was diverse and unique in enrolling participants soon after licensure and comparatively large for a driving simulator study, the sample size was relatively small for more complex analyses, such as invariance testing of the principal components by licensure status. The sample was recruited with the primary purpose of assessing attention development in licensed and unlicensed adolescents, so conclusions drawn regarding driving behaviors in the unlicensed adolescents may be interpreted with some caution. To fully understand how simulated driving variables can form comprehensive scores of driving behavior, a strategic (planning) level of driver behavior should be assessed. In the driving scenario, no strategic goal was set or measured. As instructions involved “driving as if you were in a real car on a real road,” it can only be assumed, or implied at-best, that participants shared a common strategy. It should be noted that we do not have local crash data with which to compare findings. Future studies may consider assessing early exposure to driving situations (e.g., driving video games) as contributors to operational control. Future studies may also consider providing goal-based scenarios or measuring a driver’s strategy if the scenario provides such ability (e.g., multiple routes available). Additionally, replicability is critical to any research endeavor. Future studies may wish to try to replicate this study, perhaps using a different platform, to compare results.
Conclusions and Future Directions
This work represents the first step to developing standard measures of simulator driving that can be used to compare results across driving simulator studies, enhance rigor of transportation science, and produce translatable driving scores reflective of driving behavior theory. The wide variation in how driving behavior and performance is defined across driving simulator studies makes it difficult to compare results and reduces rigor and reproducibility. In other words, a standard set of driving “scores” used in research should increase replicability by providing a holistic picture of driving performance, rather than focusing on the individual simulated driving variables that suggest statistical significance. As these individual metrics utilizes are commonly collected, we anticipate these components can be constructed from simulators with similar capabilities and in platforms beyond the one used in this work. Future studies should test this across multiple platforms. It should be noted that driving simulator outcomes may vary dependent upon the purpose of the simulated scenario, as the driving behaviors will be heavily influenced by the simulated driving scenario. Still, conceptually and empirically informed composites of driving simulator variables are also necessary to translate findings into meaningful conclusions, recommendations, and training goals. Future work should examine the replicability and stability of these driving scores and factors across studies and within drivers over time, particularly in young drivers where driving behavior is still developing. Additionally, future studies should evaluate applicability and invariance of these composite scores across populations with different demographic, cultural, and driving characteristics, as well as clinical populations of drivers (e.g., those with developmental disabilities or after injury). Ultimately, driver education and training programs that utilize simulators should evaluate the utility of these standard driving scores to inform how they assess skill and determine driving behaviors in need for additional practice.
Figure 33.
Car Following Task Environment
Highlights.
Driving simulator variables form factors aligning with driving behavior theories.
Driving simulator variables factor into operational and tactical driving behaviors.
Operational driving is informed by simulator variables quantifying variability.
Tactical driving is most informed by simulator variables representing speed.
ACKNOWLEDGMENTS
Special thanks to the research team of the UAB Translational Research for Injury Prevention Laboratory and families who participated in this research.
FUNDING
Research reported in this publication was supported by the Eunice Kennedy Shriver National Institute of Child Health & Human Development of the National Institutes of Health under Award Number R01HD089998. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.
Footnotes
DISCLOSURE STATEMENT
The authors report there are no competing interests to declare.
CRedIT AUTHORSHIP CONTRIBUTION STATEMENT
Benjamin McManus, PhD: conceptualization, methodology, software, validation, formal analysis, investigation, data curation, writing (original, review & editing) visualization; Sylvie Mrug, PhD: formal analysis, writing (original review & editing); William P. Wagner, PhD: investigation, writing (original, review & editing); Andrea Underhill, PhD: writing (original, review & editing), visualization; Piyush Pawar, MS: software, data curation, writing (review & editing); Thomas Anthony, MS: software, data curation, writing (review & editing); Despina Stavrinos, PhD: conceptualization, methodology, validation, investigation, resources, writing (original, review & editing), data curation, visualization, supervision, project administration, funding acquisition;
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
REFERENCES
- Beck KH, Yan F, Wang MQ. 2007. Cell phone users, reported crash risk, unsafe driving behaviors and dispositions: a survey of motorists in Maryland. J Safety Res. 38(6):683–688. eng. [DOI] [PubMed] [Google Scholar]
- Age, skill, and hazard perception in driving. The Fourth International Driving Symposium on Human Factors in Driving Assessment, Training, and Vehicle Design; 2007; Iowa City, IA. University of Iowa. [Google Scholar]
- Currie CE, Elton RA, Todd J, Platt S. 1997. Indicators of socioeconomic status for adolescents: the WHO Health Behaviour in School-aged Children Survey. Health education research. 12(3):385–397. eng. [DOI] [PubMed] [Google Scholar]
- Durbin DR, Mirman JH, Curry AE, Wang W, Thiel MCF, Schultheis M, Winston FK. 2014. Driving errors of learner teens: Frequency, nature and their association with practice. Accident Analysis & Prevention. 72:433–439. [DOI] [PubMed] [Google Scholar]
- Karimi A, Bassani M, Boroujerdian A, Catani L. 2020. Investigation into passing behavior at passing zones to validate and extend the use of driving simulators in two-lane roads safety analysis. Accident Analysis & Prevention. 139:105487. [DOI] [PubMed] [Google Scholar]
- Lee JD. 2007. Technology and teen drivers. J Safety Res. 38(2):203–213. eng. [DOI] [PubMed] [Google Scholar]
- Maxwell H, Weaver B, Gagnon S, Marshall S, Bedard M. 2021. The validity of three new driving simulator scenarios: detecting differences in driving performance by difficulty and driver gender and age. Hum Factors. 63(8):1449–1464. [DOI] [PubMed] [Google Scholar]
- McGwin G Jr., Brown DB. 1999. Characteristics of traffic crashes among young, middle-aged, and older drivers. Accident Analysis & Prevention. 31(3):181–198. [DOI] [PubMed] [Google Scholar]
- McManus B, Heaton K, Stavrinos D. 2017. Commercial motor vehicle driving performance: An examination of attentional resources and control using a driving simulator. J Exp Psychol Appl. 23(2):191–203. [DOI] [PubMed] [Google Scholar]
- Michon JA. 1979. Dealing with danger. Verkeerskundig Studiecentrum, Rijksuniversiteit Groningen Haren, The Netherlands. [Google Scholar]
- Michon JA. 1985. A critical view of driver behavior models: What do we know, what should we know? New York: Plenum Press. (Human behavior and traffic safety. [Google Scholar]
- Papantoniou P 2018. Structural equation model analysis for the evaluation of overall driving performance: A driving simulator study focusing on driver distraction. Traffic Inj Prev. 19(3):317–325. [DOI] [PubMed] [Google Scholar]
- Shults RA, Banerjee T, Perry T. 2016. Who’s not driving among US high school seniors: A closer look at race/ethnicity, socioeconomic factors, and driving status. Traffic Inj Prev. 17(8):803–809. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Stavrinos D, Jones JL, Garner AA, Griffin R, Franklin CA, Ball D, Welburn SC, Ball KK, Sisiopiku VP, Fine PR. 2013. Impact of distracted driving on safety and traffic flow. Accident Analysis & Prevention. 61:63–70. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Strayer DL, Cooper J, Drews FA. 2011. Chapter 27, Profiles in cell-phone induced driver distraction. In: Fisher DL, Rizzo M, Caird JK et al. , editors. Handbook of Driving Simulation for Engineering, Medicine, and Psychology. Boca Raton, FL: CRC Press; p. 1–9. [Google Scholar]
- Whittington C, McManus B, Albright G, Beck H, Buford K, Gresham B, Svancara A, Underhill A, Anthony T, Mrug S et al. 2020. Methodological considerations for longitudinal investigations of young drivers. Transportation Research Part F: Traffic Psychology and Behaviour. 75:214–221. [Google Scholar]