Older driver fitness-to-drive evaluation using naturalistic driving data

Journal of Safety Research 54 (2015) 29

Contents lists available at ScienceDirect

Journal of Safety Research journal homepage: www.elsevier.com/locate/jsr

Editorial

Letter from the Editors - Fourth international symposium on naturalistic driving research The Journal of Safety Research is pleased to present this collection of papers that were originally presented at the Fourth International Symposium on Naturalistic Driving Research. The symposium, hosted by the National Surface Transportation Safety Center for Excellence (NSTSCE) at Virginia Tech, was held in August 2014. From over 40 papers and posters exploring a wide range of naturalistic driving topics, these studies have been selected through our peer-reviewed process to be presented in this special issue. Although all of the studies included in this special issue use naturalistic driving research methods, the topics explored and analysis methods used vary widely. Studies in this collection can be roughly categorized into three broad groups: Novice driving: • • • •

Naturalistic teenage driving study — Findings and lessons learn Using naturalistic driving data to examine drivers' seatbelt use behavior, comparison between teens and adults Personality and crash risk Conducting in-depth naturalistic riding study: examples from beginner motorcyclists Distracted driving:

• • • •

Creation of the NEST distracted driving dataset Are cellular phone blocking applications effective for novice teen drivers? Drivers' visual behavior when using handheld and hands-free cell phones Examination of drivers' cell phone use behavior at intersections by using naturalistic driving data Methodological papers exploring innovative techniques in data extraction and analysis:

• • • • •

Population distributions of time to collision at brake application during car following from naturalistic driving data Evaluation of a video-based measurement of driver heart rate Drunk driving detection based on classification of multivariate time series Naturalistic drive cycle synthesis for pickup trucks Older driver fitness-to-drive evaluation using naturalistic driving data

We hope you find this collection of naturalistic driving research valuable. Through programs like SHRP 2 (see accompanying letter and articles in this issue) naturalistic driving research will become more prevalent in the years to come with the potential of revolutionizing our understanding of motor vehicle safety. However, all research methodologies have limitations, and no single methodology can fully explain the complex causal nature of crashes. The Journal invites all researchers conducting rigorous evidence-based investigations, regardless of the methods used or conclusions made, to consider submitting their studies. These studies add to the understanding of us all. Only through the publishing of findings in peerreviewed journals and through the subsequent debate on the merits of the research can the field of motor vehicle safety research advance. In this light, the Journal invites thoughtful commentary on this collection of studies. Thomas W. Planek Editor-in-Chief Sergey Sinelnikov Jonathan Thomas Kenneth Kolosh Associate Editors Kathleen Porretta Managing Editor 9 June 2015

http://dx.doi.org/10.1016/j.jsr.2015.06.003 0022-4375/© 2015 National Safety Council and Elsevier Ltd. All rights reserved.

Journal of Safety Research 54 (2015) 31

Contents lists available at ScienceDirect

Journal of Safety Research journal homepage: www.elsevier.com/locate/jsr

Editorial

The 4th International Symposium on Naturalistic Driving Research

The Virginia Tech Transportation Institute is proud to have hosted the 4th International Symposium on Naturalistic Driving Research in August of 2014. The papers presented in this special issue are expanded versions of the papers and posters presented at that symposium, and they represent the first dedicated collection of papers in this new area of research. In the past 20 years, we have seen the field of naturalistic driving research expand in incredible fashion. Advances have occurred in all aspects: from vehicles with car trunks and truck cabs filled with analog recording equipment to state-of-the-art miniaturized data collection systems, from a few participants to thousands of participants per study, from manual coding of data using video tape players and spreadsheets to sophisticated data coding and extraction software, and from simple parametric statistical analysis to advanced statistical modeling techniques. Most importantly, naturalistic driving has progressed to the point that the methods, equipment, and data are now available to a wide variety of researchers. This is what made the 4th Symposium so special: for the first time there were enough researchers doing work in the field that we were able to have a call for papers. By contrast, the three previous symposia were introductory in nature — introducing the methods, equipment, and analysis techniques to a new generation of researchers, with invited papers from those known to be working in the field. We hope that you find the papers presented in this issue to be useful in your own research, and that you will consider adding the naturalistic driving techniques and data to your research portfolio. Most importantly, we hope that the research highlighted in this issue will provide the impetus to help save lives and improve transportation efficiency worldwide. Jon Hankey Senior Associate Director Virginia Tech Transportation Institute, USA 21 June 2015

http://dx.doi.org/10.1016/j.jsr.2015.06.004 0022-4375/© 2015 National Safety Council and Elsevier Ltd. All rights reserved.

Journal of Safety Research 54 (2015) 49–54

Contents lists available at ScienceDirect

Journal of Safety Research journal homepage: www.elsevier.com/locate/jsr

Older driver fitness-to-drive evaluation using naturalistic driving data Feng Guo, ⁎ Youjia Fang, 1 Jonathan F. Antin 2 Virginia Tech Transportation Institute, 3500 Transportation Research Plaza, Blacksburg, VA 24061, USA

a r t i c l e

i n f o

Article history: Received 23 December 2014 Received in revised form 27 February 2015 Accepted 24 June 2015 Available online 29 July 2015 KEYWORDS: Fitness to drive Older driver Naturalistic driving study Driving risk Contrast sensitivity

a b s t r a c t Problem: As our driving population continues to age, it is becoming increasingly important to find a small set of easily administered fitness metrics that can meaningfully and reliably identify at-risk seniors requiring more indepth evaluation of their driving skills and weaknesses. Method: Sixty driver assessment metrics related to fitness-to-drive were examined for 20 seniors who were followed for a year using the naturalistic driving paradigm. Principal component analysis and negative binomial regression modeling approaches were used to develop parsimonious models relating the most highly predictive of the driver assessment metrics to the safetyrelated outcomes observed in the naturalistic driving data. Results: This study provides important confirmation using naturalistic driving methods of the relationship between contrast sensitivity and crash-related events. Practical applications: The results of this study provide crucial information on the continuing journey to identify metrics and protocols that could be applied to determine seniors' fitness to drive. © 2015 National Safety Council and Elsevier Ltd. All rights reserved.

1. Introduction In a letter to the editor of the British Medical Journal, Martin Stratford, presumably a general practitioner (GP), laments being put between the ostensibly competing interests of the insurance company and the elderly driver and his or her family when making fitness to drive determinations. He suggests that an independent and specifically trained physician would be better suited to make the such determinations, or at least that the GPs who currently make such decisions be furnished with a better, more objective set of tools for this purpose (Stratford, 1959, February 14). This is remarkable because the identical concerns articulated in his letter are still very much with us and largely unresolved more than a half century later. Some 40 years after the appearance of Stratford's letter, Marshall and Gilbert (1999) conducted a survey of physicians in Saskatchewan, Canada, who were likely to be involved in making fitness-to-drive determinations. They found that while 57.6% of the respondents indicated that they do not hesitate to report patients whom they believe to be medically unfit to drive, an even greater percentage (59.5%) felt as Stratford did, that while necessary, this type of reporting harms the physician–patient relationship. Although physicians around the turn of the 21st century did have better tools and information available than Stratford and his peers, Marshall and Gilbert still concluded that physicians' understanding ⁎ Corresponding author at: Department of Statistics, Virginia Tech, Virginia Tech Transportation Institute, 3500 Transportation Research Plaza, Blacksburg, VA 24061, USA. Tel.: +1 540 231 1038; fax: +1 540 231 1555. E-mail addresses: [email protected] (F. Guo), [email protected] (Y. Fang), [email protected] (J.F. Antin). 1 Tel.: +1 540 231 1518; fax: +1 540 231 1555. 2 Tel.: +1 540 231 1579; fax: +1 540 231 1555.

http://dx.doi.org/10.1016/j.jsr.2015.06.013 0022-4375/© 2015 National Safety Council and Elsevier Ltd. All rights reserved.

regarding the relationship between specific medical conditions and the resulting increments in driving risk tended to be poor. 1.1. Functional impairment Previous studies have demonstrated that seniors' driving risk is associated with certain dimensions of functional impairment typically associated with aging. Anstey, Wood, Lord, and Walker (2005) reviewed several studies reporting some manner of statistical association between predictors and driving performance (i.e., on-road metrics or crash rates) in adults 60 +. They found studies that reported associations for cognitive, perceptual, physical metrics as well as certain medical conditions. Based on this literature review, Anstey et al. (2005) developed a conceptual model whereby all of these associated factors affect one's capacity to drive safely, but this capacity must also be combined with sufficient metacognitive ability so that the senior driver can sensibly moderate or alter his/her driving habits in a way that successfully manages risk. In a similar vein, De Raedt and Ponjaert-Kristoffersen (2000) used a road test to classify senior drivers into: “good,” “average,” and “bad,” driver groups. They further found that those “bad” drivers who had reported having an at-fault crash within the past year had reported engaging in significantly fewer strategic compensation strategies (e.g., avoiding rush-hour traffic or other complex driving scenarios) than those “bad” drivers who had reported no at-fault crashes in the past year. These data imply that age-related loss of functionality may be successfully compensated via the application of safer strategic driving decisions and behaviors. Festa, Ott, Manning, Davis, and Heindel (2013) used naturalistic driving methods to demonstrate that both a group of Alzheimer patients as well as an age-matched control group tended to drive in conditions perceived to impose lower risk

50

F. Guo et al. / Journal of Safety Research 54 (2015) 49–54

(e.g., daylight, sunny weather, and light traffic). They also included nopassenger trips as lower risk; however, the presence of senior passengers has a complex effect on safety, but can more often reduce risk by helping with navigation, hazard detection, and general alertness (Meyers & Meyers, 2004). If senior drivers, even Alzheimer patients, typically self-regulate to the point where risk is reduced to an acceptable level, then this may mask or mute the ability of fitness-to-drive metrics to reliably predict safety-related driving outcomes. Antin, Lockhart, Stanley, and Guo (2012) compared the fitness profiles of seniors who were still driving with those of an agematched cohort of seniors who had recently ceased driving in an effort to devise fitness-to-drive models. Results of that study showed that the fitness profiles of the drivers and non-drivers were, not surprisingly, very different, with the drivers demonstrating better functional abilities than their non-driving counterparts for virtually every metric where there was a statistically significant difference. In addition, Antin et al. (2012) developed parsimonious models to investigate if driver or nondriver group membership could be predicted based solely upon the individual's fitness profile. A five-factor model was 100% successful at predicting group membership. An even more parsimonious threefactor model was nearly perfect at predicting group membership; this model included the following factors: dynamic visual acuity, average upper body maximum torque, and a metric for visual-cognitive ability (Trail Making B). The naturalistic driving study (NDS) research paradigm is an advanced method used to evaluate driving safety, especially for driver-related risk factors (Dingus et al., 2006; Fitch et al., 2013; Klauer, Guo, Sudweeks, & Dingus, 2010; Klauer et al., 2014; Olson, Hanowski, Hickman, & Bocanegra, 2009). The NDS is characterized by the unobtrusive installation of advanced instrumentation on vehicles and continuous recording on driving data in a natural, nonexperimental setting. The data typically include vehicle kinematics, multiple video camera images, GPS data, radar, and vehicle network variables. This rich source of information provides a means for evaluating the safety impact of driver distraction, speeding, and many other risky driving behaviors or other factors (e.g., environmental or scenario-based factors). In particular, the NDS data can be used to detect multiple safety-related events, such as crashes and near-crashes, thus providing a powerful tool to identify the relative risks associated with particular driver characteristics (Guo & Fang, 2013; Ouimet et al., 2014). Anstey et al. (2005) lamented the fact that studies investigating the relationship of senior driver functional impairment to driving ability and crash risk in older drivers have often lacked a sufficiently broad framework incorporating cognitive, perceptual, motor, and physical factors. The objective of the current study is to bring all of these factors and more to bear on evaluating whether the fitness profile data used in Antin et al. (2012) are also associated with safety-related outcomes observed in that same study's naturalistic driving record. Specifically, the current study attempts to proscriptively relate a broad array of seniors' functional assessment metrics to crash and near-crash (CNC) event rates observed in a 12-month sample of naturalistic driving data. It is our hypothesis that some combination of these factors will be able to reliably predict safety-related outcomes observed in the driving data.

2. Methods 2.1. Older driver naturalistic driving study The older driver naturalistic data collection included 50 participants living in the vicinity of the New River Valley area of Virginia, among which 27 were active drivers at the time of study and 23 individuals who had ceased driving within the past two years of that study (Antin et al., 2012). That study consists of two major data collection components: a thorough examination of participants' functional abilities (i.e., the fitness metrics) and a one-year naturalistic driving data collection for a subgroup of drivers. Note that of the 27 active drivers who participated the functional examination, only 20 of them went on to participate in the naturalistic driving data collection that consisted of them freely driving on public roads without restriction in a nonexperimental setting. The study was approved by the Virginia Tech Institutional Review Board. Sixty fitness-to-drive assessments were collected from all 50 participants. In this analysis, 7 of the 60 metrics were dropped from the analyses for various reasons. For example, the color vision metrics have essentially identical values for almost all participants, thus providing no modeling-relevant information. The data structure for the current analysis is illustrated in Fig. 1. The 53 remaining metrics included in the analysis are shown in Table 1. Details on how the metrics were measured are presented in Antin et al. (2012). There are a small number of drivers with part of the fitness data missing. Due to the relatively small sample size, a data imputation approach was used to make maximum use of all data available. Antin et al. (2012) suggested that older driver and non-driver groups may have significantly different fitness profiles, the missing values were therefore imputed by the group mean of either drivers or non-drivers, depending on each participant's driving status at the time of the study. The contrast sensitivity test consists of nine categories of contrast sensitivity scores under each spatial frequency for both left and right eyes (Table 2). The scores represent the magnitude of the contrast, with Category-1 representing the lowest contrast sensitivity and Category-9 representing the greatest sensitivity. To satisfy the assumptions of the PCA, a natural logarithmic transformation was conducted to transfer the exponentially-scaled raw score to linear scale and the transformed scores are denoted as CSL/CSRs (Table 2). The transformed scores are used in the analysis throughout this paper. One issue associated with the log-transformation is that if a participant fails to detect the lowest contrast sensitivity (Category-1), he/she will be recorded as Category-0, with no corresponding raw score. To address this issue, we imputed the score for Category-0 using the following procedure. For each spatial frequency, a linear model was fitted as follows, Y ¼ α þ βX where Y is the logarithm of raw scores; X is the category number (1–9); a and β were estimated using a regular least square estimation. In this setup, a is the logarithm of raw score when X equals to 0 and we impute the raw score for Category-0 with exponential of a.

Fig. 1. Data structure used in the current analysis.

F. Guo et al. / Journal of Safety Research 54 (2015) 49–54

transferred to a secure data server at the Virginia Tech Transportation Institute. IRB approval was secured both for the original data collection as well as the current follow-on data mining effort.

Table 1 Fitness assessment metrics. Functional evaluation metrics

Abbreviation

Physical ability (13 total) Ankle torque max/plantar and dorsiflexion (2) Hip torque max/flex and extend (2) Upper body torque max/left and right (2) Ankle initial reaction time (mean of plantar and dorsiflexion)(1) Ankle peak reaction time (mean of plantar & dorsiflexion) (1) Hip initial reaction time (mean of flex and extend) (1) Hip peak reaction time (mean of flex and extend) (1) Upper body initial reaction time (mean of left and right) (1) Upper body peak reaction time (mean of left and right) (1) Head–neck–torso flexibility (1)

Ankle tm/p-d Hip tm/f-e Upper tm/l-r Ankle i rt Ankle p rt Hip i rt Hip p rt Upper i rt Upper p rt Flex

Visual ability (24 total) Dynamic visual acuity at 12, 24, and 36 degree per second (3) Discomfort glare rating (1) Glare static acuity (1) Glare contrast sensitivity at 4, 8, and 16 cycles per degree (3) Static visual acuity (Snellen) (1) Contrast sensitivity by left and right eye at spatial frequencies of 1.5, 3, 6,12, and 18 cycles per degree (10) Total number of color vision plates correct (1) Stereopsis (1) Far acuity (optec) (1) Far vertical and lateral phoria (2)

Dvac12–36 Dgr Glare acuity Glare cs 4–16 Acuity Csl 1.5–18 Sr 1.5–18 Color sum Opt1 Opt2 Opt3–4

General and health-related info (10 total) Height (in.)—self report (1) Weight (lbs)—self report (1) Total number of reported health problems (1) Faces pain scale (1) Who (five) well-being index 1998 version (1) Total number of sleep problems (1) Total number of sleep disorders (1) Total hours of sleep estimated per day (1) Education (1) Total years driving (1)

Ht Wt Health Pain Well Sleep p Sleep d Hrs/day Ed Yrs drv

Cognitive ability (6 total) Abbreviated Mental Test Score (AMTS) (1) Self-estimate: how mentally sharp compared with 40s and 50s (1) Metacognition ratio: self estimate/AMTS (1) Visualizing missing information (1) Useful field of view™ (1) Trail making B (1)

Cog Meta Ratio MI UFoV TM

51

2.1.1. Safety-related events The collected data were analyzed through a rigorous multi-step protocol to identify two types of safety-related events, crashes and nearcrashes. The first step is to run automated algorithms across the time series data to determine threshold boundary transgressions indicating possible crash-related events. These potential events were then visually inspected by trained analysts and confirmed by the research staff as described in Dingus et al. (2006). A crash is defined as “any contact with an object, either moving or fixed, at any speed in which kinetic energy is measurably transferred or dissipated. Crashes include a participant's vehicle making contact with other vehicles, roadside barriers, objects on or off the roadway, pedestrians, cyclists, or animals;” a near-crash is a crash surrogate defined as “any circumstance requiring a rapid, evasive maneuver by the participant (or his/her vehicle) or any other vehicle, pedestrian, cyclist, or animal to avoid a crash” (Dingus et al., 2006). As the number of crashes is small, we combined crash and near-crash events to represent crash risk, which afforded increased statistical power. We believe this is a valid approach as previous studies have shown that near-crashes share kinematic and behavioral similarities with crashes and provide useful information on driving risk (Guo, Klauer, Hankey, & Dingus, 2010; Klauer et al., 2014). 2.2. Principal component regression A principal component regression (PCR) analysis was conducted to assess the relationship between fitness-to-drive metrics and crash and near-crash (CNC) risk. The method consists of two components: a principal component analysis (PCA) to reduce the dimensionality of the covariate matrix followed by negative binomial (NB) regression to link the CNC with the reduced fitness-to-drive metrics. The advantages of PCA include: (a) it maximizes the useable information in the data via dimension reduction; (b) the derived PCs are not correlated, which eliminates the multicollinearity issue in subsequent modeling; and (c) the factor loading pattern of the principal components (PCs) can help detect which metrics are closely related and have a significant impact on a particular PC. The first step of analysis is to apply PCA to reduce the dimensionality of the redundant and correlated fitness profile data (Jolliffe, 2002). All metrics were standardized before PCA (i.e., raw values minus mean and then divided by standard deviation). Fitness profiles comprised 53 assessment metrics (P = 53 columns) and 50 participants (N = 50 rows). Many of these metrics are highly correlated, as they measure similar or related constructs. When performing the regression analysis, it is not possible to include all metrics as covariates because (a) the number of observations N is smaller than the number of covariates P; and (b) there is a severe multicollinearity issue due to the high correlations among several of the fitness metrics.

As noted above, 20 drivers participated in a one-year naturalistic driving data collection in which the participants drove in their own vehicles instrumented with an advanced onboard data acquisition system (DAS). During the naturalistic driving data collection period, the DAS collected continuous driving data from ignition-on to ignition-off. The DAS included multiple video cameras that captured the forward and rear roadway, driver face, and the center stack. In addition, driving kinematics and other data were collected including accelerometers in three dimensions, yaw rate, vehicle speed, forward radar, and a GPS location trace. The driving data were downloaded regularly from the participants' vehicles and Table 2 Contrast sensitivity raw and transformed score table. Spatial frequency

1.5 3 6 12 18 1.5 3 6 12 18

Contrast sensitivity category

Raw score

Transformed score

0

1

2

3

4

5

6

7

8

9

4.75 7.25 8.31 5.57 2.92 1.56 1.98 2.12 1.72 1.07

7 10 12 8 4 1.95 2.30 2.48 2.08 1.39

9 15 16 11 6 2.20 2.71 2.77 2.40 1.79

13 20 23 15 8 2.56 3.00 3.14 2.71 2.08

18 29 33 22 12 2.89 3.37 3.50 3.09 2.48

25 40 45 30 17 3.22 3.69 3.81 3.40 2.83

36 57 64 43 23 3.58 4.04 4.16 3.76 3.14

50 80 90 60 33 3.91 4.38 4.50 4.09 3.50

71 114 128 85 46 4.26 4.74 4.85 4.44 3.83

100 160 180 120 65 4.61 5.08 5.19 4.79 4.17

52

F. Guo et al. / Journal of Safety Research 54 (2015) 49–54

PCA uses orthogonal transformation to convert correlated metrics into a set of uncorrelated principal components (PCs), which are linear combinations of optimally weighted observed metrics. The first PC has the highest variance and accounts for the highest proportion of the variability of the data. Each succeeding component in turn has the highest possible variance among the remaining components, while maintaining orthogonality with preceding components. The second step is a regression analysis to link the PCs reduced from the first step with the CNC risk. Preliminary analysis has shown that the CNC event data have a moderate variance over-dispersion issue (chisquare/degree of freedom (DF) value at 1.7–2.0). NB regression models were used to model over-dispersed count data. In NB models, the number of CNCs is assumed to follow a NB distribution, Yi  NBðEi λi ; γÞ with E(Yi) = Eiλi, Var(Yi) = Eiλi + (Eiλi)2 × γ, where Yi is the number of CNCs for driver i; λi is the expected CNC rate (the number of events per 100 h driven) for driver i; the exposure Ei is measured by hours driven in a unit of 100 h; γ is the dispersion parameter; when γ converges to 0, the variance of Y converges to the mean of Y, and the NB model converges to a Poisson model. The expected CNC rate λi (per 100 h driven) follows as: logðλi Þ ¼ β0 þ

J X β j Xi j ; j¼1

where Xij is the jth covariate, typically the PC for driver i; the β ' s are the regression coefficients. For regression models with a single PC, J equals 1. 2.3. NB regression scanning The PCR analyses discussed above are based on highly efficient principal components. However, one drawback of the PCR is that, because the four fitness categories were predefined by the researchers, the categorization schema might not be optimal. There is a possibility that some metrics that are individually significant but were still not detected, either because they were not included in significant PC or they were included in a non-significant PC with respect to the NB model. Either way, the potential significant metrics will be masked in the PCR. To prevent missing any potential significant metrics masked by the PCR analysis, we used NB models to individually screen all 53 metrics. This scanning procedure complements the PCR and ensures all fitness metrics significantly associated with driving risk are detected. 3. Results 3.1. Safety outcomes Eighty CNC events, including 6 crashes and 74 near-crashes, were identified from 4,158 driving hours for the 20 participating drivers. 3.2. Principal component regression The PCA identify significant principal components for a set of correlated metrics. The eigenvalue-one criterion was used to choose the important or information-rich PCs (i.e., ones with eigenvalue greater than 1; Kaiser, 1960). Although the PCA reduces the dimensionality of the covariate matrix, it is constrained by the ratio of N/P (i.e., the ratio between sample size and the number of variables). In general, N/P should be at least 2 and ideally be greater than 5 to 10 for the PCA to perform well. As the fitness profile data comprise N = 50 participants and P = 53 assessment metrics, it is not appropriate to conduct PCA for all metrics simultaneously. To overcome this obstacle, we divided the fitness metrics into

four categories by the nature of the metrics: (a) physical ability, (b) visual ability, (c) health, and (d) cognitive ability. PCA was conducted for each of these categories and the N/P ratio was at least 2. The PCA analysis identified three significant component for physical ability, seven for visual ability, three for general health condition, and three for cognitive ability. These PCs account for 73%, 74%, 59%, and 75% of total variability for each category. In this study, we labeled these components as Physical Component 1–3, Visual Component 1–7, General Health Component 1–3, and Cognitive Component 1–3. NB regression was used to model the relationship between CNC rate and each of 16 PCs individually. The results of the NB regression models are shown in Table 3. There is no evidence of lack-of-fit for all the models as the Chi-square/DF values are close to one. As shown in the table, the Visual-2 component is statistically significantly related to CNC rate with regression coefficient of −0.22 and p-value of 0.028. Table 3 Results of principal component regression analysis. PC

Regression Std. dev. 95% LCL 95% UCL p-Value Chi-square/DF coefficient estimation

Physical-1 Physical-2 Physical-3 General-1 General-2 General-3 Cognitive-1 Cognitive-2 Cognitive-3 Visual-1 Visual-2 Visual-3 Visual-4 Visual-5 Visual-6 Visual-7

0.11 −0.01 0.12 −0.07 −0.01 −0.18 0.11 0.13 −0.21 0.10 −0.22 0.03 0.10 −0.27 −0.27 −0.09

−0.16 −0.84 −0.36 −0.59 −0.59 −0.85 −0.45 −0.46 −0.57 −0.27 −0.42 −0.45 −0.32 −0.65 −0.55 −0.54

0.14 0.42 0.25 0.26 0.30 0.34 0.29 0.30 0.19 0.19 0.10 0.25 0.22 0.20 0.15 0.23

0.38 0.82 0.61 0.44 0.57 0.49 0.67 0.72 0.15 0.46 −0.02 0.52 0.53 0.12 0.02 0.36

0.417 0.981 0.622 0.778 0.977 0.591 0.708 0.665 0.259 0.611 0.028⁎ 0.901 0.633 0.174 0.069 0.683

1.01 1.05 1.03 1.04 1.05 1.13 1.03 1.01 1.05 1.02 1.30 1.04 1.05 1.06 1.24 1.07

⁎ Statistically significant at 0.05 level.

To identify which fitness metrics contributed significantly to the Visual-2 component, the factor loading pattern of the metrics for the Visual-2 was examined as show in Table 4. The value of factor loading ranges between 0 and 100 and indicates the relative contribution of each metric to the specific component. The positive or negative sign of factor loading indicates the positive or negative relationship between the metric and the PC. The metrics with factor loading greater than 40 are considered to have a substantial contribution to the PC (Stevens, 1986). As can be seen, five fitness metrics, CSR1.5, CSR3, CSR6, CSR12, and CSR18, contributed substantially to the Visual-2 component. The factor loadings for all five are positively associated with the PC. Combined with the NB regression results that Visual-2 is negatively associated with CNC risk, the five significant fitness metrics are thus negatively correlated with CNC risk. Thus, drivers with higher values of CSR1.5 to

Table 4 Factor loading pattern of Visual-2 component. Metric

Factor loadings Visual-2

Metric

Factor Loadings Visual-2

Metric

Factor loadings Visual-2

DVAC12 DVAC24 DVAC36 DGR Glare Acuity Glare CS 4 Glare CS 8 Glare CS 16

1 6 5 0 19 16 −13 −1

ACUITY CSL 1.5 CSL 3 CSL 6 CSL 12 CSL 18 CSR 1.5 CSR 3

−28 −3 21 13 16 5 86⁎ 84⁎

CSR 6 CSR 12 CSR 18 Color Sum OPT1 OPT2 OPT3 OPT4

84⁎ 72⁎ 70⁎

⁎ Significant metrics with factor loading value greater than 40.

13 5 16 24 13

F. Guo et al. / Journal of Safety Research 54 (2015) 49–54

53

Table 5 Negative binomial regression screening results. Metric

Estimate

Std. dev.

95% LCL

95% UCL

p-Value

Chi-square/DF

CSR 1.5 CSR 3 CSR 6 CSR 12 CSR 18

−0.73 −0.48 −0.56 −0.69 −0.51

0.35 0.25 0.29 0.28 0.33

−1.41 −0.98 −1.12 −1.24 −1.16

−0.06 0.02 0.00 −0.14 0.14

0.034⁎ 0.059 0.050⁎ 0.014⁎ 0.126

1.34 1.24 1.23 1.20 1.34

⁎ Statistically significant at 0.05 level.

CSR18 (i.e., better contrast sensitivity), tended to have lower risk of experiencing a crash-related event. 3.3. NB regression screening on 53 metrics As noted above, the NB regression analyses were performed on each of 53 individual metrics. Only three individual metrics (CSR 1.5, CSR 6, CSR 12) are individually significant with p-value smaller than 0.05. Table 5 shows the screening results for the five factors identified in the PCR analysis. Note that CSR 3 (p-value = 0.06) and CSR 18 (p-value = 0.13) are significant in the PCR analysis above but are not individually significant. The screening results are consistent with the PCR analysis in general, and confirm that no potential significant fitness metrics were omitted from the PCR analysis.

where αi is the standardized scoring coefficient for the ith metric; xi and si are the mean and standard deviation of the ith metric. Table 7 shows the coefficients for α i ; xi ; and si derived in the current study. The predicted CNC risk rate (i.e., the number of CNC events per 100 h driven) is computed as:

3.4. Prediction models for CNC risk

  CNC ratepred ¼ exp β0 þ β1  PC future ;

The PCR and NB screening analyses indicate that among the 53 fitness-to-drive matrices, five metrics related to the right-eye contrast sensitivity (CSR 1.5, 3, 6, 12, 18) significantly impact CNC risk. To provide a parsimonious model, we conducted a PCR analysis using only these five metrics. The PCA indicated that the first principal component (PC-1) accounts for 67% of total variability and it is the only component with eigenvalue greater than 1. The factor loading values for all five metrics are greater than 40. Therefore, we propose to use the first component for CNC risk prediction. The NB model estimation results for the proposed model are shown in Table 6. As can be seen, the PC is statistically significant (p b 0.05). The CNC risk rate ratio (i.e., the exponential of the regression coefficient for PC-1) is 0.768 (95% CI: 0.626–0.941). As the value of the PC increases by 1 unit, CNC risk rate decreases by 23%. The observed data, the curve of fitted mean CNC rates and the corresponding 95% pointwise confidence band (CB) obtained by NB regression are demonstrated in Fig. 2. Fig. 2 clearly indicates that the mean CNC rate decreases as the PC-1 score increases. One implementation of the above model is to predict the CNC risk based on CSR 1.5, 3, 6, 12, and 18 data observed in the future, denoted as x1,future to x5,future. The PC score for data collected in the future can be computed by summing the products of the standardized scoring coefficient αi and the standardized values of xi,future.

PC future ¼

5 X xi;future −xi αi  ; si i¼1

Fig. 2. Observed and fitted CNC rates.

where β0 and β1 are the estimated regression coefficients presented in Table 6. 4. Discussion To find a quantitative and efficient way to assess older drivers' fitness-to-drive and predict driving risk would potentially have a profound impact on driver license regulation practices, an individual's decision to restrict or cease driving, and the welfare of older drivers and all with whom they share the road. This study evaluated the relationship between senior drivers' fitness profiles and driving risk represented by crash and near-crash rate observed naturalistically. PCA was used for metric dimensionality reduction and group classification. Due to the moderate variance over-dispersion issue in the CNC data, an NB regression model was applied to model the relationship between CNC rates and participants' fitness profiles. The study indicated that right eye contrast sensitivity metrics (CSR 1.5–CSR 18) are significantly related to CNC risk. The better the contrast sensitivity, the lower the crash risk. Contrast sensitivity refers to the eye's ability to resolve information presented with limited contrast (i.e., where there is relatively little difference between the light and dark aspects of a stimulus—the lower the difference in contrast from which the information or content can reliably be retrieved, the greater the contrast sensitivity of that observer). A recent study sponsored by the National Highway Traffic Safety Administration (NHTSA) looked at the degree to which a variety of metrics of functional ability could be used to proscriptively predict crash and serious traffic violation rates (Staplin, Lococo, Gish, & Joyce, 2012).

Table 7 Coefficients for computing PC Score for CNC proposed model.

Table 6 Parameter estimation for proposed model. Parameter

DF

Estimate

Standard error

Wald 95% confidence limits

p-Value

Intercept PC-1 Dispersion

1 1 1

0.585 −0.265 0.087

0.148 0.104 0.145

0.296 −0.469 0.003

b0.001 0.011

0.875 −0.060 2.301

Metric

ai

xi

si

CSR 1.5 CSR 3 CSR 6 CSR 12 CSR 18

0.234 0.232 0.269 0.267 0.215

3.037 3.655 3.362 2.633 1.730

0.299 0.462 0.398 0.479 0.627

54

F. Guo et al. / Journal of Safety Research 54 (2015) 49–54

Results of that study showed greatest promise for a route-planning/ maze-solving assessment. In addition, a metric related to contrast sensitivity was also significantly related to crash involvement. Others using retrospective methods have also found positive results in terms of relating contrast sensitivity to crash rate (McGwin, Chapman, & Owsley, 2000; McNight & McNight, 1999; Owsley, Stalvey, Wells, Sloane, & McGwin, 2001). Casson and Racette (2000) reviewed visions standards in North America and concluded that binocular contrast sensitivity is as important as, if not more important than, good visual acuity for driving. The results of these studies and of the current study are not surprising as there are many driving scenarios where key visual targets (e.g., other vehicles or pedestrians) may have relatively low contrast with the background field, especially in low-light conditions. If the contrast between salient parts of the target image and the background visual field is below the driver's ability to detect them, then the potential for a crash exists and would be greater than that of a driver with greater sensitivity in that spatial frequency range. One important contribution of this study is confirming these previous findings related to the relationship between contrast sensitivity and senior driver risk using data collected via the naturalistic research paradigm. The main limitations of the current study are its small sample size and limited geographic coverage for the study area. Therefore caution should be used when extrapolating the results to general population. In this study, relatively complicated statistical approaches and models were applied to address the significant analysis challenge relating to the number of variables being greater than the sample size. Also, near crashes have been used as crash surrogates to indicate driving risk. As such, it is possible that other aspects of seniors' functional impairment may emerge as being predictive of safety related outcomes when a larger, more geographically diverse dataset can be analyzed. These issues can hopefully be addressed by conducting a similar analysis on data collected in the Strategic Highway Research Plan 2 (SHRP 2) Naturalistic Driving Study (Dingus et al., 2014), which includes data from more than 3000 participants collected within six geographically disperse and diverse locations, with several hundred of those participants being in the senior age range. The results of this study provide crucial information on the metrics and protocols that could someday be applied by motor vehicle departments, physicians, occupational therapists, Certified Driving Rehabilitation Specialists, and others for whom determining seniors' fitness to drive is an important component of their work. Acknowledgments This research was supported by a grant from the National Surface Transportation Safety Center for Excellence. References Anstey, K. J., Wood, J., Lord, S., & Walker, J. G. (2005). Cognitive, sensory and physical factors enabling driving safety in older adults. Clinical Psychology Review, 25(1), 45–65. Antin, J. F., Lockhart, T., Stanley, L. M., & Guo, F. (2012). Comparing the impairment profiles of older drivers and non-drivers: Toward the development of a fitness-todrive model. Safety Science, 50(2), 333–341. Casson, E. J., & Racette, L. (2000). Vision standards for driving in Canada and the United States. A review for the Canadian Ophthalmological Society. Canadian Journal of Ophthalmology, 35(4), 192–203.

De Raedt, R., & Ponjaert-Kristoffersen, I. (2000). Can strategic and tactical compensation reduce crash risk in older drivers? Age and Ageing, 29(6), 517–521. Dingus, T. A., et al. (2006). 100 Car naturalistic driving study—Phase II results of the 100 Car field experiment. National Highway Traffic Safety Administration. DOT HS 810 593. Dingus, T. A., et al. (2014). Naturalistic driving study: technical coordination and quality control. Project S06 draft final report. http://onlinepubs.trb.org/onlinepubs/shrp2/ SHRP2_S06Report.pdf (accessed 12/17/2014) Festa, E. K., Ott, B. R., Manning, K. J., Davis, J. D., & Heindel, W. C. (2013). Effect of cognitive status on self-regulatory driving behavior in older adults an assessment of naturalistic driving using in-car video recordings. Journal of Geriatric Psychiatry and Neurology, 26(1), 10–18. Fitch, G. M., et al. (2013). The impact of hand-held and hands-free cell phone use on driving performance and safety-critical event risk. National Highway Traffic Safety Administration. DOT HS 811 757. Guo, F., & Fang, Y. (2013). Individual driver risk assessment using naturalistic driving data. Accident Analysis and Prevention, 61(1), 3–9. Guo, F., Klauer, S. G., Hankey, J. M., & Dingus, T. A. (2010). Near crashes as crash surrogate for naturalistic driving studies. Transportation Research Record. Journal of the Transportation Research Board, 2147, 66–74. Jolliffe, I. T. (2002). Principal component analysis (2nd ed.). New York: Springer-Verlag. Kaiser, H. F. (1960). The application of electronic computers to factor analysis. Educational and Psychological Measurement, 20, 141–151. Klauer, S. G., Guo, F., Simons-Morton, B. G., Ouimet, M. C., Lee, S. E., & Dingus, T. A. (2014). Distracted driving and risk of road crashes among novice and experienced drivers. New England Journal of Medicine, 370(1), 54–59. Klauer, S. G., Guo, F., Sudweeks, J., & Dingus, T. A. (2010). An analysis of driver inattention using a case-crossover approach on 100-Car data: Final report. Washington, D.C: National Highway Traffic Safety Administration (DOT HS 811 334.). Marshall, S. C., & Gilbert, N. (1999). Saskatchewan physicians' attitudes and knowledge regarding assessment of medical fitness to drive. Canadian Medical Association Journal, 160(12), 1701–1704. McGwin, G., Jr., Chapman, V., & Owsley, C. (2000). Visual risk factors for driving difficulty among older drivers. Accident Analysis and Prevention, 32(6), 735–744. McNight, A. J., & McNight, A. S. (1999). Multivariate analysis of age-related driver ability and performance deficits. Accident Analysis and Prevention, 31(5), 445–454. Meyers, M., & Meyers, J. R. (2004). The influence of passengers on older drivers involved in fatal crashes. Experimental Aging Research, 30(2), 205–215. Olson, R. L., Hanowski, R. J., Hickman, J. S., & Bocanegra, J. (2009). Driver distraction in commercial vehicle operations (Document no FMCSA-RRT-09-042). USDOT: Federal Motor Carrier Safety Administration. Ouimet, M. C., et al. (2014). Higher crash and near-crash rates in teenaged drivers with lower cortisol response: An 18-month longitudinal, naturalistic study. Journal of the American Medical Association Pediatrics, 168(6), 517–522. Owsley, C., Stalvey, B., Wells, J., Sloane, M., & McGwin, G. (2001). Visual risk factors for crash involvement in older drivers with cataract. Archives of Ophthalmology, 119(6), 881–887. Staplin, L., Lococo, K. H., Gish, K. W., & Joyce, J. (2012). Functional assessments, safety outcomes, and driving exposure measures for older drivers (Report No. DOT HS 811 630). Washington, DC: National Highway Traffic Safety Administration. Stevens, J. (1986). Applied multivariate statistics for the social sciences. Hillsdale, NJ: Lawrence Erlbaum Associates. Stratford, M. (1959, February 14). Fitness to drive [Correspondence to the editor]. British Medical Journal, 442. Feng Guo, Ph.D. is an Associate Professor at the Department of Statistics at Virginia Tech with a joint appointment at the Virginia Tech Transportation Institute. He earned dual Ph.D. in Statistics (2007) and Transportation Engineering (2010) from the University of Connecticut. Youjia Fang, Ph.D. is a statistician/research associate at VTTI. Dr. Fang earned both his M.S. (2010) and Ph.D. (2014) degrees in Statistics at Virginia Tech. Jon Antin, Ph.D., CHFP is Director of the Center for Vulnerable Road User Safety (CVRUS) at VTTI. Dr. Antin also serves as the Vulnerable Road User subject matter expert for the National Surface Transportation Safety Center for Excellence at VTTI. Dr. Antin earned his B.S. in Psychology at L.S.U. then performed his graduate work in the Vehicle Analysis and Simulation Laboratory at Virginia Tech where he earned the M.S. and Ph.D. degrees in Industrial Engineering and Operations Research (Human Factors Option).