Factors Affecting Drivers' Cell Phone Use Behavior - SAGE Journals

0 downloads 0 Views 584KB Size Report
The purpose of this study was to examine drivers' cell phone use behavior as reflected in naturalistic driving data. Video data from 1 week's worth of driving for ...
Factors Affecting Drivers’ Cell Phone Use Behavior Implications from a Naturalistic Study Huimin Xiong, Shan Bao, and James R. Sayer same time frame (5). Many studies have been conducted to try to solve this puzzle. Simulator studies have clearly shown that performance, typically measured by reaction time, degraded when the driver was distracted by a variety of means, including cell phone conversations and dialing or texting (6–8). However, a field study, which by its very nature includes naturalistic behavior, has high external validity. Natural distractions have been observed through several field studies. The 100-Car study conducted around Washington, D.C., examined four categories of distractions caused by secondary task engagement, fatigue, driving-related inattention to the forward roadway, and nonspecific eye glance (9). The research analyzed how drivers taking their eyes off the forward roadway related to near-crash or crash risk and concluded that crash risk increased when the driver looked away from the road for more than 2 s, within a time span of 6 s. The Automotive Collision Avoidance System Field Operation Test, conducted by the University of Michigan Transportation Research Institute (UMTRI), examined whether the prevalence of secondary tasks increased with the presence of a driver assistance system (Adaptive Cruise Control and Forward Collision Warning). The study found that conversation with a passenger was the only increasing secondary behavior associated with the presence of driver assistance systems (10). Another naturalistic study conducted by UMTRI, the Road Departure Crash Warning Field Operation Test, found that 34% of the sample clips (1,440 video clips of 5 s for 36 drivers were sampled) included secondary task engagement. Handheld cell phone use was the third most frequently observed behavior among all the secondary behaviors (5.3%). The study also found that secondary task engagement led to significantly increased steering angle variance compared with driving without engaging in a secondary task. Phone use was associated with the highest steering angle variance (11). Flannagan et al. used data from the Integrated Vehicle-Based Safety System Field Operational Test (IVBSS FOT) to observe drivers’ selfregulation behavior with cell phone use while driving (12). Although these studies focused more on either overall driving performance or the effect of a new crash warning system on driver distraction in terms of all general secondary tasks, rather than cell phone use itself, the methodology and insights drawn from the studies could better represent the nature of drivers’ decisions to engage in secondary tasks and therefore provide a way to examine drivers’ behavior more realistically. Funkhouser and Sayer examined drivers’ cell phone use, including conversation and visual and manual tasks, through a naturalistic study (13). The frequency, duration, and circumstances under which drivers used their cell phone while driving were analyzed. The results showed that drivers tended to initiate a cell phone conversation more

The purpose of this study was to examine drivers’ cell phone use behavior as reflected in naturalistic driving data. Video data from 1 week’s worth of driving for 108 participants were visually scored for all instances of cell phone use, including conversation and visual or manual (VM) tasks. The frequency of cell phone use for each participant was used to classify drivers’ behavior. Three frequency groups (low, moderate, and high) were scored across all drivers for conversation and VM tasks separately. The regression tree method was used to classify drivers’ cell phone use behavior and identify associated factors. Drivers’ individual factors, including age, annual driving mileage, and education levels, as well as situational factors, including use duration, time of day, road type, lighting (day and night), traffic conditions, and speed when initiating cell phone use, impacted drivers’ cell phone use behavior. The impacts of these factors were different for cell phone conversation and VM tasks. Traffic conditions were identified as affecting drivers’ cell phone VM task use frequency but not cell phone conversation frequency. The study also looked at driver self-regulation behavior based on the frequency of cell phone use.

The use of a cell phone while driving, as an important contributing factor of driver distraction, has been extensively examined in the literature. Many studies have reported that cell phone use significantly degrades driving performance and increases crash risks (1, 2). Given the commonality of cell phone subscriptions and other entertainment devices available in the vehicle, the rate of distraction-related crashes could be expected to escalate. However, the Insurance Institute for Highway Safety reported the paradox of increasing cell phone subscription and use rates coupled with decreasing crash rates (3). The finding was estimated based on NHTSA data that the percentage of total distraction-related crashes that occurred in the United States remained at 17% from 2007 through 2010 and fell to 15% in 2011, while the percentage of cell phone use–related distraction crashes remained relatively constant between 0.8% and 0.9% (4). The puzzle surrounding cell phone use while driving is that cell phone use has been shown to decrease driving performance and increase crash risk, yet driver cell phone use rates have increased and crash rates and crash incidence have steadily declined in the University of Michigan Transportation Research Institute, 2901 Baxter Road, Ann Arbor, MI 48109-2150. Corresponding author: H. Xiong, [email protected]. Transportation Research Record: Journal of the Transportation Research Board, No. 2434, Transportation Research Board of the National Academies, Washington, D.C., 2014, pp. 72–79. DOI: 10.3141/2434-09 72

Xiong, Bao, and Sayer

frequently when they were driving at lower speeds and that afternoon was the peak time for cell phone use. To understand drivers’ cell phone use behaviors while driving, more research is needed to examine whether and how individual difference factors (age, gender, driving experience, and education), driving environment factors (traffic, road type, speed, and lighting conditions), and social norm factors (time of the day to call or text) influence cell phone use frequency and driving safety. The purpose of this study was to investigate how these factors relate to the use of a cell phone and the implications regarding self-regulation behaviors. Method Naturalistic Data The data used for this study were collected from the IVBSS FOT conducted by UMTRI in 2009 and 2010 (13). As part of the IVBSS FOT, 108 participants were randomly selected from a list of licensed drivers in Southeast Michigan. The participants drove instrumented vehicles with an integrated suite of collision warning systems for six weeks. The drivers were equally distributed in three age groups: younger (20–30 years old), middle aged (40–50 years old), and older (60–70 years old). The first two weeks of the FOT consisted of baseline driving with no warnings presented to the participants. However, in weeks three through six, warnings were presented to the participants. Data were recorded continuously, at 10 ∼ 50 Hz, during all six weeks, with more than 600 channels of numerical data and five on-board cameras recording video data inside and outside the car. There were no audio recordings of conversations with passengers or of cellular phone conversations. Participants were fully informed of the data recording and the camera locations, in addition to being trained in the operation of the integrated warning system. Cell Phone Event Video Review The study used data from the baseline period to examine drivers’ natural cell phone use behavior. Trained reviewers watched every moment of video from the first week of driving for all 108 participants. The video data were scored specifically for cell phone usage, including phone calls (referred to here as conversations) as well as any manipulation, or reading, of the cell phone while driving. A conversation was logged when the participant was holding the phone to his or her ear or clearly speaking into or listening to a handheld or hands-free device. Interactions with the phone that were not conversations are referred to as visual or manual (VM) tasks. VM tasks included manipulating the keys or the touch screen on the phone (to dial, text, email, and so forth) or looking away from the road at the phone screen when the phone was in hand. Conversations or VM tasks observed at the start or end of an ignition cycle were specifically scored as such; however, no distinction was made for the analysis.

73

Variables Response Variable: Drivers’ Cell Phone Use Behavior, Frequency Groups According to Funkhouser and Sayer, the number of cell phone use engagements while driving across all 108 drivers varied widely (13). Fourteen of the 108 participants engaged in zero conversations, but 12 of the 108 participants engaged in more than 30 conversations during the first week of driving. Similarly, for VM tasks, 14 of the 108 participants engaged in zero VM tasks, but six of the 108 participants engaged in more than 100 VM tasks during the first week of driving. The number of cell phone uses, either a conversation or VM task, for each driver indicates how frequently each driver used the cell phone while driving. Given such large differences across drivers, grouping drivers by use frequency could be a potential indicator of drivers’ cell phone engagement behaviors. Therefore, drivers were classified into three groups (low frequency, moderate frequency, and high frequency) based on the number of conversations or VM tasks each participant engaged in within the first week of driving (Table 1). The classifications were based on two criteria: the total number of events in each category and the clear cut lines. The cut lines were identified based on the distribution of the number of events. For example, to classify the moderate and high frequency groups for VM tasks, none of the participants had between 87 and 102 events, which was a big jump in the number of events for each participant. Therefore, the value 100 was used as a cut line. Participants with no conversation or VM tasks were excluded from the analysis. Covariates • Duration. The time between the initiation and termination of a conversation or VM task was measured in minutes. • Time of day. Cell phone initiation time was based on a 24-h clock. • Age group. Three age groups were used as designed for the FOT study: 1 (younger: 20–30 years old), 2 (middle aged: 40–50 years old), and 3 (older: 60–70 years old). • Gender. Males and females were included in the original data set. • Education group. Five education levels (high school, some college, bachelor’s degree, master’s degree, and doctor of medicine) were recorded in the original data set and coded as three levels: 1 (high school), 2 (some college or bachelor’s degree), and 3 (master’s degree or doctor of medicine).

TABLE 1   Driver Frequency Groups of Cell Phone Use: Conversations and VM Tasks

Event Type

Data Selection For cell phone use to be included in the data set, an event duration of a cell phone conversation or VM task could not occur when the car was in park. There were 1,381 conversations and 2,149 VM tasks observed and utilized in the analysis.

Conversation

VM tasks

Frequency Group

Number for Classification

Number of Drivers

Total Number of Events

Low Moderate High Low Moderate High

1–15 16–34 35+ 1–30 31–100 101+

65 19 10 75 13  6

437 462 482 660 681 808

74

• Self-reported annual driving mileage group. Driver self-reported annual driving mileage ranged from 1,000 mi to 70,000 mi and was coded in one of three groups: 1 (low mileage: 1,000–10,000 mi), 2 (moderate mileage: 10,001–20,000 mi), and 3 (high mileage: >20,000 mi). • Traffic. Three traffic conditions: 1 (sparse), 2 (moderate), and 3 (dense), were coded in the original data set. Sparse conditions were defined as on-road traffic with either zero or one vehicle observable by the forward radar, moderate traffic was defined as on-road traffic of between two and four vehicles, and dense traffic was defined as on-road traffic of more than four vehicles. • Road type. Six road types: 0 (unknown), 1 (highway), 3 (major surface), 4 (minor surface), 5 (local), and 6 (ramp) were identified in the original study. Unknown road types were generally driveways, parking lots, or very small residential streets. • Lighting (day and night). Day was defined as the period from morning civil twilight through evening civil twilight. • Start speed. The speed of the vehicle when the conversation or VM task was initiated was recorded in the original data set.

Data Analysis To examine the relationships between cell phone use frequency (categorical variable) and each covariate, the Fisher’s exact test (for a small sample size) and χ2 test for independence were used for categorical covariates. Each contingency table showed observed values and expected values (in parentheses). When the observed value was much greater than the expected value, the significantly greater impact of that category was applied and vice versa. The KolmogorovSmirnov test was used for continuous covariates to compare two distributions. It can be difficult for a single global regression model to capture the data features given that the data have many features that interact in complicated and nonlinear ways, even confounding each other. A regression tree is a nonlinear regression method to partition covariate space into smaller regions, where the interactions are more manageable. Recursive partitioning was used to form subspaces until chunks of the space were obtained that were more homogeneous for the response variable (14). Regression trees use binary trees to rep­resent the recursive partition. Each intermediate node has a condition associated with it. The observation goes to the left branch if it satisfies the condition; otherwise, it goes to the right branch. Each terminal node or leaf represents a predicted class. An observation belongs to a leaf if it falls in the corresponding cell of the partition (15). A regression tree provides the capability to distinguish the likelihood of drivers’ cell phone use behavior, which is a discrete outcome (three frequency groups: low, moderate, high), by partitioning the covariate space. In this study, the variables of interest were included as covariates (or predictors). High dimensional spaces of covariates were partitioned into regions with relatively homogeneous cell phone use frequency. Drivers’ frequency groups were taken at nodes of the tree. Each intermediate node represented a single attribute and had a true or false answer, such as age ≥ 1.5. The terminal nodes were the leaves and each contained a classification (1 representing low, 2 moderate, and 3 high frequency). The regression trees provided a hierarchical procedure to describe the partition with the level of each covariate placed to represent its priority to make a classification. Therefore, regression trees were used to identify factors that had an

Transportation Research Record 2434

impact on the driver’s cell phone use behavior. All the analyses were conducted using R statistical software. Results Conversation Duration Given the difference in conversation duration across all 1,381 cell phone conversation events, duration distributions were compared between frequency groups. All the distributions were highly skewed toward zero. However, the moderate frequency group had significantly longer durations compared with the low frequency group [Kolmogorov-Smirnov (KS) test, p = .001] and high frequency group (KS test, p < .001). No significant differences were observed between the low and high frequency groups (KS test, p = .12). Time of Day The percentage of cell phone conversations initiated during each hour of the day indicated to some degree what was deemed as socially acceptable behavior (Figure 1). Overall, 34.3% of the calls were initiated between 4:00 and 7:00 p.m. and only 1.3% were initiated between midnight and 7:00 a.m. The low frequency group had a consistent overall conversation distribution, but with a slightly higher proportion at the peak time. The distribution of the moderate frequency group showed a flatter shape compared with the overall distribution and had almost an equal chance to initiate a call in each hour from 2:00 to 10:00 p.m. However, the distribution of the high frequency group was a bi-norm shape with the peak value of 10% between 1:00 and 2:00 p.m. and between 5:00 and 6:00 p.m. This group also had a greater proportion between midnight and 1:00 a.m., between 5:00 and 6:00 a.m., between 12:00 and 2:00 p.m., and between 10:00 and 11:00 p.m. Individual Driver Differences Drivers’ age, gender, education level, and self-reported annual driving mileage were examined among the three frequency groups. As previous studies have shown, younger drivers had a significantly higher rate of cell phone conversations compared with older and middle-aged drivers (13). This study also confirmed that there is a significant difference in age groups among the three frequency groups (Fisher’s exact test, p = .02). In the low frequency group, there were more older drivers and fewer younger drivers observed when compared with the expected values (values in parentheses in Table 2). In the moderate frequency group, there were more observed middleaged drivers and fewer older drivers than the expected values. In the high frequency group, an extremely high proportion of younger drivers was observed (Table 2). No significant differences in gender or education affected which drivers were observed in the three groups. Significant differences in self-reported annual mileage among the three groups were observed (Fisher’s exact test, p = .04). In the low frequency group, a higher proportion of drivers with low annual mileage were observed compared with the other groups. In the moderate frequency group, a higher proportion of drivers with moderate annual reported mileage were observed compared with the other groups. In the high

Xiong, Bao, and Sayer

75

20 18 16 Percentage

14 12 10

Conversaon

8

Low Frequency

6

Moderate Frequency

4

High Frequency

2 23:00–24:00

22:00–23:00

21:00–22:00

20:00–21:00

19:00–20:00

18:00–19:00

17:00–18:00

16:00–17:00

15:00–16:00

14:00–15:00

13:00–14:00

12:00–13:00

11:00–12:00

10:00–11:00

09:00–10:00

08:00–09:00

07:00–08:00

06:00–07:00

05:00–06:00

04:00–05:00

03:00–04:00

02:00–03:00

01:00–02:00

00:00–01:00

0

Time of Day FIGURE 1   Percentage of cell phone conversations initiated during each hour of day by frequency group.

frequency group, a higher proportion of drivers with high annual mileage were observed compared with the other groups. The results indicate that low frequency in engaging in cell phone use was associated with less driving.

drivers engaged in a higher proportion of conversations on major surface roads and local roads compared with other conditions. Lighting

Traffic Conditions There was no significant difference in traffic conditions among the three groups. In other words, the proportions of sparse, moderate, and dense traffic conditions were similar in the three cell phone conversation frequency groups.

There was a significant difference between lighting among the three groups [χ2(df = 2) = 8.74, p = .013]. In the low frequency group, drivers engaged in a higher proportion of conversations during the day. However, in the two other groups, drivers engaged in a higher proportion of conversations at night. Start Speed

Road Type There was a significant difference in road type among the three groups (χ2 [degrees of freedom (df) = 10] = 23, p = .011). In the low frequency group, drivers engaged in a higher proportion of conversations on highways and major surface roads compared with other conditions. In the moderate frequency group, drivers engaged in a higher proportion of conversations on unknown and minor surface road types compared with other conditions. In the high frequency group,

TABLE 2   Contingency Table of Age and Frequency Groups for Conversation Observed Value (Expected Value), by Age Group Group

Younger

Middle Aged

Low Moderate High Total

19 (23.5)   7 (6.87)   8 (3.62) 34

24 (24.2)   9 (7.07)   2 (3.72) 35

Older 22 (17.3)   3 (5.05)   0 (2.66) 25

Total Observed 65 19 10 94

Speed distributions were compared between frequency groups. A significantly lower start speed for the high frequency group was observed compared with the low frequency group (KS test, p < .001). No significant differences were observed for the other groups.

Regression Tree The regression tree model showed the importance of each factor in identifying the frequency groups for cell phone use. If a driver was in the middle-aged or older driver group (age ≥ 1.5) and drove less than 10,000 miles per year (mileage < 1.5), the driver was more likely to be in the low frequency group. If a driver fell in the younger driver group (age < 1.5) and drove more than 20,000 miles per year (mileage ≥ 2.5), the driver was more likely to be in the high frequency group. Further partitions and factors for predicting the frequency group of cell phone conversations are shown in Figure 2. As criteria, traffic conditions were not recommended in making a determination about the frequency group of cell phone conversation. This finding was consistent with the previous marginal analysis in that there was no significant difference in traffic conditions among the three frequency groups.

76

Transportation Research Record 2434

FIGURE 2   Regression tree for cell phone conversations.

VM Tasks Duration Given the difference of VM task duration across all 2,149 events, duration distributions were compared between frequency groups. All distributions were highly skewed toward zero. However, the high frequency group had significantly longer durations compared with the low frequency group (KS test, p < .001) and the moderate frequency group (KS test, p < .001). No significant difference was observed between the low and moderate frequency groups (KS test, p = .06). Time of Day Overall, 33.8% of VM tasks were initiated between 4:00 and 8:00 p.m.; only 4.3% were initiated between 1:00 and 8:00 a.m. The low frequency group had a distribution consistent with the overall conversation distribution, but with a slightly higher proportion at the peak time as well as between 11:00 a.m. and 12:00 p.m. The distribution of the moderate frequency group showed a flatter shape with higher percentages between 1:00 and 2:00 p.m., 7:00 and 8:00 p.m., and 9:00 and 10:00 p.m. compared with the overall distribution. However, the peak time for the high frequency group was between 6:00 and 7:00 p.m. with 10% of their VM tasks conducted during that time period. They also had a greater proportion of VM tasks between midnight and 1:00 a.m., 2:00 and 3:00 p.m., and 11:00 p.m. and midnight (Figure 3). Individual Driver Differences Drivers’ age, gender, education level, and self-reported annual driving mileage were examined among the three frequency groups. As a previous study showed, younger drivers had a significantly higher rate of VM tasks compared with older and middle-aged drivers (13). This study also confirmed that there was a significant difference in

age groups (Fisher’s exact test, p = .014). In the low frequency group, there were more older drivers than expected and fewer younger drivers than expected. In the moderate and high frequency groups, there were more younger drivers than expected and fewer older drivers than expected. There were no significant differences in the effects of gender, education levels, or self-reported annual mileage among the three groups observed. Traffic There was a significant difference in traffic conditions among the three groups [χ2 (df = 4) = 16.6, p = .002]. In the moderate frequency group, drivers had a higher observed number of VM tasks (44) in dense traffic situations compared with the expected value (28.5). In the high frequency group, drivers had a higher observed number of VM tasks in sparse and moderate traffic conditions compared with the expected values (Table 3). Road Type There was a significant difference in road types among the three frequency groups [χ2 (df = 10) = 23.3, p = .01]. In the low frequency group, drivers had a higher proportion of VM tasks with unknown, highway, and major surface road conditions compared with other conditions. In the moderate frequency group, drivers had a higher proportion of VM tasks with unknown and local road conditions compared with other conditions. In the high frequency group, drivers had a higher proportion of VM tasks on major surface roads, minor surface roads, and ramps compared with other road conditions. Lighting There was a significant difference between lighting among the three frequency groups [χ2 (df = 2) = 40.6, p < .001]. In the low frequency group, drivers had a higher proportion of VM tasks during daylight

Xiong, Bao, and Sayer

77

16 14

Percentage

12 10 8

VM task

6

Low Moderate

4

High

2

00

:0 0 01 –01 :0 :0 0 0 02 –02 :0 :0 0 0 03 –03 :0 :0 0 0 04 –04 :0 :0 0 0 05 –05 :0 :0 0 0 06 –06 :0 :0 0 0 07 –07 :0 :0 0 0 08 –08 :0 :0 0 0 09 –09 :0 :0 0 0 10 –10 :0 :0 0 0 11 –11 :0 :0 0 0 12 –12 :0 :0 0 0 13 –13 :0 :0 0 0 14 –14 :0 :0 0 0 15 –15 :0 :0 0 0 16 –16 :0 :0 0 0 17 –17 :0 :0 0 0 18 –18 :0 :0 0 0 19 –19 :0 :0 0 0 20 –20 :0 :0 0 0 21 –21 :0 :0 0 0 22 –22 :0 :0 0 0 23 –23 :0 :0 0– 0 24 :0 0

0

Time of Day FIGURE 3   Percentage of cell phone VM tasks initiated during each hour of day by frequency group.

hours. However, in the other two groups, drivers had a higher proportion of VM tasks when it was dark outside. Start Speed Start speed distributions were compared for each frequency group. A significantly higher start speed for the high frequency group was observed compared with the low frequency group (KS test, p = .006) and the moderate frequency group (KS test, p < .001). A significantly lower start speed was observed for the moderate group compared with the low frequency group (KS test, p = .02).

driver was more likely to be in the low frequency group. If a driver drove more than 20,000 miles per year (mileage ≥ 2.5), was in the younger driver group (age < 1.5), and had a master’s level education or above (education ≥ 2.5), the driver was more likely to be in the high frequency group. Further partitions and factors to predict the frequency groups of cell phone VM tasks are shown in Figure 4. Traffic conditions were considered as criteria to determine a frequency group of cell phone VM tasks, which was also consistent with the previous marginal analysis in which there was significant difference in traffic conditions among the three groups. Discussion of Results

Regression Tree The regression tree model showed that self-reported annual mileage was on the top level in determining the frequency group for cell phone VM tasks. If a driver drove less than 20,000 miles per year (mileage < 2.5), had a high school or college education (education < 2.5), and was in the older driver group (age ≥ 2.5), the

TABLE 3   Contingency Table of Traffic Conditions and Frequency Groups for VM Tasks Observed Value (Expected Value), by Traffic Condition Frequency Group

Sparse

Moderate

Dense

Low Moderate High Total

  483 (482)   488 (498)   600 (591) 1,571

149 (150) 149 (155) 190 (183) 488

28 (27.6) 44 (28.5) 18 (33.8) 90

Total Observed 660 681 808 2,149

The goal of this study was to identify factors that influence drivers’ cell phone use behavior while driving. Great differences in the frequency of cell phone use were observed among individual drivers. There were 69.1% (65/94) drivers with low frequency cell phone conversations (≤15 times) and 79.8% (75/94) drivers with low frequency cell phone VM tasks (≤30 times) during one week of driving. The majority of these only contributed to about one-third of the cell phone use events. Therefore, cell phone use while driving was still not a highly frequent event for most drivers. This was consistent with findings from a previous study (16). In general, age and self-reported annual mileage played an important role in participants’ cell phone use frequency while driving. Older drivers with low mileage were more likely to be in the low frequency group. This group also tended to engage in cell phone use during the daytime with shorter duration compared with the other groups. The older drivers group had a higher proportion of drivers receiving a call (57%) rather than initiating a call (43%). Younger drivers had a significantly higher incidence of cell phone use compared with the two other age groups. Given that cell phone use is more prevalent among younger people in general, their greater experience and familiarity with cell phone

78

Transportation Research Record 2434

FIGURE 4   Regression tree for cell phone VM tasks.

use could be expected. Therefore, high frequency of cell use among younger drivers might not necessarily mean increased risk. In fact, in terms of conversation, significantly lower start speeds for the high frequency group were observed compared with the low frequency group. Among the high frequency group, there were 39.8% (192/482) events observed with low start speed (