Computers in Human Behavior 52 (2015) 115–123

Contents lists available at ScienceDirect

Computers in Human Behavior

journal homepage: www.elsevier .com/locate /comphumbeh

Mutual interferences of driving and texting performance

http://dx.doi.org/10.1016/j.chb.2015.05.0040747-5632/� 2015 Elsevier Ltd. All rights reserved.

⇑ Corresponding author at: Wichita State University, 1845 Fairmount St., Wichita,KS 67260, USA. Cell: +1 217 417 3830.

E-mail address: jibo.he@wichita.edu (J. He).

Jibo He a,⇑, Alex Chaparro a, Xiaohui Wu b, Joseph Crandall a, Jake Ellis a

a Department of Psychology, Wichita State University, Wichita, KS, USAb Department of Psychology, Tsinghua University, Beijing, China

a r t i c l e i n f o

Article history:

Keywords:Texting while drivingDriver distractionMobile devicesLane Change TaskCell phone

a b s t r a c t

Despite legislative and social campaigns to reduce texting while driving, drivers continue to text behindthe wheel. There is abundant evidence demonstrating that texting while driving impairs driving perfor-mance. While past driver distraction research has focused on how texting influences driving, the influ-ence of driving on texting behaviors has been overlooked. This study used a Lane Change Task and asmartphone texting application to study the mutual influences of driving and texting. Results showedthat concurrent texting impaired driving by increasing the lane deviation. Meanwhile, driving impairedtexting by increasing texting completion time, texting errors, and key entry time intervals, and reducedkey entry speed. In addition, we show that texting behavioral data collected can be used to distinguishtexting while driving from texting-only condition with an accuracy of 88.5%. The mutual interferencesof driving and texting inform the theory of dual-task performance and provide a scientific foundationto develop a smartphone-based technology to reduce the risky behavior of texting while driving.

� 2015 Elsevier Ltd. All rights reserved.

1. Introduction

Texting while driving has become a widespread risky behaviorand may impair driving performance more than talking on a cellphone (Caird, Johnston, Willness, Asbridge, & Steel, 2014; Drews,Yazdani, Godfrey, Cooper, & Strayer, 2009; He et al., 2014;Hosking, Young, & Regan, 2009; Klauer et al., 2014; Wilson &Stimpson, 2010). As many as 281,000 to 786,000 crashes in 2012may involve text messaging, according to the estimate of theUnited States National Safety Council (2012). Pickrell and Ye(2013) found 0.9% of drivers were visibly manipulatinghand-held devices while driving in 2010, and this percentageincreased to 1.3 percent in 2011. The risk and prevalence of textingwhile driving has attracted the attention of the general public, automanufacturers, legislators and safety researchers (Jacobson &Gostin, 2010; Owens, McLaughlin, & Sudweeks, 2011).

Concurrent texting impairs driving in various ways. For exam-ple, texting while driving increases hazard response time (Burge& Chaparro, 2012; Drews et al., 2009; He et al., 2014; Leung,Croft, Jackson, Howard, & Mckenzie, 2012), increases lane devia-tions (the difference between the center of the vehicle and the cen-ter of the appropriate lane) and lane excursions (leaving the laneunintentionally) (Alosco et al., 2012; Crandall & Chaparro, 2012;

Hosking et al., 2009; McKeever, Schultheis, Padmanaban, &Blasco, 2013; Rudin-Brown, Young, Patten, Lenné, & Ceci, 2013),increases mental demand (mental demands are psychologicaland mental stress experienced by an individual while completingone or more tasks) (Owens et al., 2011; Rudin-Brown et al.,2013), increases gaze-off-road durations (Hosking et al., 2009;Libby, Chaparro, & He, 2013; Owens et al., 2011), causes more col-lisions (Alosco et al., 2012; Drews et al., 2009), and raises the risksof traffic accident as many as 8–23 times (Olson, Hanowski,Hickman, & Bocanegra, 2009).

People have limited ability to perform two tasks simultane-ously, such as texting and driving and doing so results in deficitson one or both of the tasks being performed (Allport, Antonis, &Reynolds, 1972). According to the theories of dual-task perfor-mance, when two tasks are carried out concurrently, the perfor-mances of one or both tasks may be impaired, causing adual-task performance decrement (Wickens, 2002). For example,when performing a secondary auditory monitoring task (pressinga button when they hear a tone), drivers had slower reaction timewhen responding to a vehicle braking, compared to driving-onlyconditions, even when instructed to give the driving task priority(Levy & Pashler, 2008). While several studies have reported theeffects of texting on driving performance, how driving affects tex-ting performance has been ignored. Better understanding ofchanges in both driving and texting performance can inform theo-ries of dual-task performance and contribute to the efforts to mit-igate the risks of texting while driving.

116 J. He et al. / Computers in Human Behavior 52 (2015) 115–123

Researchers and practitioners have explored a variety ofapproaches to mitigate the risks of texting while driving, includinglegislation, social campaigns, and technological solutions.Legislative efforts have sought to discourage this behavior by mak-ing it illegal (McCartt & Geary, 2004; Mccartt, Hellinga, Strouse, &Farmer, 2010) and social campaigns have sought to educate driversabout the risks of texting while driving (Atchley, Hadlock, & Lane,2012; Nemme & White, 2010). In addition, telecommunicationcompanies including AT&T and T-Mobile have developed smart-phone applications to discourage texting while driving, such asDrive Mode and DriveSmart. Smartphone users can enable theseapplications to delay or block incoming calls and messages whenthey drive, limiting their exposure to the dangers of texting whiledriving. However, despite the associated risks of texting whiledriving and these legislative, social, and technological efforts, dri-vers continue to text while driving. Ninety-one percent of collegestudents reported having sent text messages while driving, eventhough they agreed or believed that texting while driving was dan-gerous and should be illegal (Atchley, Atwood, & Boulton, 2011;Harrison, 2011).

Is it possible to develop a smartphone application to monitortexting while driving, and prevent or discourage drivers fromengaging in such risk behavior? If an application can monitor tex-ting while driving, it can potentially be easier to implement, andcan complement current efforts in law enforcement or social cam-paigns. Researchers have attempted to detect drunk driving (Dai,Teng, Bai, Shen, & Xuan, 2010), cognitively distracted driving(Liang, Lee, & Reyes, 2007), aggressive driving (Johnson & Trivedi,2011; Zeeman & Booysen, 2013), and drowsy driving (Hammoud& Zhang, 2008; He et al., 2013). However, to our best knowledge,no application has been developed that detects texting while driv-ing. Thus, this study also explores the possibility using textingbehavioral data to identify whether a driver is texting whiledriving.

The popularity and risks of texting while driving highlight thecontinuing need for research and understanding of how textinginfluences driving and vice versa, and how drivers coordinate per-formance of both tasks (Atchley et al., 2011; Harrison, 2011). Mostdriving studies focus on driving performance, while texting perfor-mance is mostly ignored or less emphasized. Smartphones allowthe collection of detailed data on secondary texting performance.In this study, we utilize the capability of smartphones to capturetexting performance and describe the mutual interferences of tex-ting and driving. The goal of this paper is to discover the mutualinfluences of concurrent texting and driving, and sought the possi-bility to detect texting while driving and reduce its risks. It is pre-dicted that participants will show greater lane deviation whiledriving and texting compared to the drive-only condition. It is alsohypothesized that participants will take longer to complete thetexting task and make more errors when driving and texting thanwhen driving-only.

Fig. 1. Screenshot of driving simulator environment.

2. Materials and methods

2.1. Participants

Twenty-eight participants (12 men, 16 women) from auniversity community ages 18–35 years (M = 22.14 years,SD = 4.64 years) volunteered to participate in this driving experi-ment. All participants were screened prior to participation toensure normal or corrected-to-normal vision using the SnellenVisual Acuity chart (Ferris, Kassoff, Bresnick, & Bailey, 1982). Allparticipants completed a survey to ensure they wereright-handed, active drivers, with at least two years of drivingexperience (M = 6.29 years; SD = 4.51 years). They all owned a

touchscreen smartphone and on average, reported sending 83 textmessages per day (median = 70, SD = 86.06).

2.2. Apparatus and stimuli

Driving performance was assessed using a driving simulatorconsisting of a General Motors car seat and Logitech DrivingForce GT steering wheel and pedals. The Lane Change Task (LCT)version 1.2 software simulated the driving task using a60 inch-Sharp Aquos 3D HD LCD display.

A 4.300 HTC ThunderBolt touch-screen smartphone running theAndroid 2.3.4 operating system was used for the texting task.The buttons on the keyboard of the smartphone were arranged ina QWERTY layout.

2.3. Experimental design

There were five counterbalanced experimental conditions,including driving-only condition, two dual-task conditions, inwhich participants either drove while texting with one hand (dri-ve + text one hand) or two hands (drive + text two hands), andtwo texting-only conditions in which participants completed thetexting task with either one or two hands. We employed awithin-subject design. Participants finished all the task conditions.

The dependent variables measured for the driving task weremean lane deviation and standard deviation of lane deviation.Lane deviation refers to the difference between the center of thevehicle and the center of the appropriate lane. For the texting task,the dependent variables consisted of task completion time, keyentry per second, texting task completion time, texting errors,input time interval, standard deviation of input time interval,and device stability (He et al., 2014).

2.4. Experiment tasks

2.4.1. Lane Change Task (LCT)In the LCT, participants were required to drive down a straight

section of road with three lanes and were prompted to changelanes according to directions on signs, which appeared on bothsides of the roadway. An arrow on the sign, shown in Fig. 1, indi-cated which lane the driver supposed to maneuver into.Participants were instructed to change from their current lane ina deliberate manner, and to do so as quickly and efficiently aspossible.

Participants maintained a constant speed of 60 km/h and wereinstructed to make lane changes as quickly and accurately as pos-sible when they saw the lane change sign. The LCT was developedby Daimler Chrysler AG Research and Technology to test driver dis-traction (Hofmann, Rinkenauer, & Gude, 2008; Huemer & Vollrath,2010; Mattes & Hallén, 2008).

J. He et al. / Computers in Human Behavior 52 (2015) 115–123 117

2.4.2. Texting taskA texting task was used to assess texting performance with two

holding postures (texting with one hand vs. texting with twohands). An incoming text message sound was played at a randomtime interval, which followed a uniform distribution in the rangeof 40–60 s. Once the incoming sound was played, participantsopened the texting application to send text messages.Participants were presented with a text message and instructedto enter the text message into the phone manually. After the mes-sage was sent, the application was closed. The text message con-sisted of a 10-digit telephone number. The digit-entry textingtask was chosen to simulate the cognitive process and key entryrequirement of a cell phone dialing or texting task while alsoallowing the measurement of secondary texting performance.Similar digit-entry tasks are commonly used to study the effectof distraction on driving performance (He et al., 2014; Horrey &Wickens, 2004; Reed & Green, 1999; Salvucci & Macuga, 2002).The application saved timestamps, keyboard entries, andaccelerometer sensor values of the smartphone in log files for lateranalysis.

2.5. Procedure

Participants were informed of the purpose of the experimentand then signed a consent form to participate. They then com-pleted a questionnaire regarding driving experience and cell phoneuse, followed by a visual screening.

The session began with instructions on how to complete thetasks and a practice session to familiarize the participants withthe simulator and the cell phone. Participants received instructionon the LCT and how to correctly execute the task. Participants wereinstructed to prioritize the driving task during dual-task conditionsand text when they felt comfortable or believed it was safe. Thephone was used in portrait orientation for all texting conditions.In the two-hand texting conditions, participants were instructedto keep both hands on the phone while texting. Participants prac-ticed each of the text-only conditions for two minutes and each ofthe driving conditions for six minutes, for a total of 20 min of prac-tice. After the practice, participants began the experimental trials.Each trial consisted of four tracks lasting three minutes each, for atotal of 12 min. The time to complete the survey, practice time, andthe experiment took approximately 105 min. Participants couldwithdraw from the study at any time. After the study, participantswere given course credits for their participation.

2.6. Data analysis

Measures of driving performance included the mean and stan-dard deviation of lane deviation recorded by the driving simulator.Lane deviation was the difference between the driver’s actual laneposition (red line) and the ‘‘ideal model’’ of lane position (greenline) for the LCT, seen in Fig. 2. A large value of the standard

Fig. 2. The Lane Change Task. The green line represents the ideal path defined by the comthe top indicates the lane, which drivers should drive in (Mattes & Hallén, 2008). (For inteweb version of this article.)

deviation of lane deviation indicates poorer lane-keeping perfor-mance with a higher risk of lane departure and collision with vehi-cles in the adjacent lanes.

Texting performance measures included texting task comple-tion time, key entry per second (KES), mean and standard deviationof input time interval, mean numbers of inputs, and mean numberof corrections. Texting errors were measured using the Levenshteinedit distance of two strings (simply termed as edit distance) (Heet al., 2014; Levenshtein, 1966). The edit distance is the numberof discrete steps required to make the strings identical, includinginsertions, deletions, switching, and substitutions. Larger edit dis-tance values indicate a greater difference between the two stringsor more texting errors. For example, the numbers 5200251314 and200251319 have a Levenshtein distance of two. The pair of num-bers can be made identical by inserting a 5 at the beginning (inser-tion) and substituting the last 9 into 4 (substitution).

All the driving performance measures were submitted torepeated-measures one-way analyses of variance (ANOVA) withthe driving condition (drive-only, drive + texting with one hand,and drive + texting with two hands) as the only within-subject fac-tors. For the one-way ANOVA of driving performance, when themain effect is significant, pairwise comparisons were madebetween the driving conditions. Bonferroni correction was usedfor post hoc pairwise comparisons. The Bonferroni-corrected pvalue is 0.05/3 = 0.017. Texting performance variables were sub-mitted to two-way analyses of variance (ANOVA) with task load(text-only vs. drive + texting) and texting methods (one hand vs.two hands) as within-subject factors. Partial eta-squared (g2

p)was used to measure effect size. A g2

p value of 0.04 indicates a rec-ommended minimum effect size for a practically significant effect,0.25 indicates a moderate effect and 0.64 indicates a strong effect(Ferguson, 2009).

3. Results

3.1. Driving performance

The standard deviation of lane deviation was significantly dif-ferent across task conditions, F(2,54) = 29.14, p < .001, g2

p = .52.As shown in Fig. 3 pairwise comparisons revealed the standarddeviation of lane deviation in the drive-only condition(M = 1.09 m, SD = .20 m) was significantly smaller than that in thedrive + texting with one hand condition (M = 1.21 m, SD = .26 m,p = .001) and drive + texting with two hands (M = 1.27 m,SD = .28 m, p < .001). The standard deviation of lane deviation inthe drive + texting one hand condition was also significantly smal-ler than that in the drive + texting two hands condition.

The mean lane deviation also differed significantly across taskconditions, F(2,54) = 41.43, p < .001, g2

p = .61. As shown in Fig. 4,the mean lane deviation in the drive-only condition (M = .81 m,SD = .15 m) was significantly smaller than that in the drive + tex-ting with one hand condition (M = .94 m, SD = .21 m), and

puter and the red line represents the path taken by the driver. The arrow symbol atrpretation of the references to color in this figure legend, the reader is referred to the

Fig. 3. Standard deviation of lane deviation across task conditions. Error bars in allfigures indicate within-subject standard errors based on the main effect of task(Loftus & Masson, 1994).

Fig. 4. Mean lane deviation across task conditions.

Fig. 5. Texting task completion time across conditions.

118 J. He et al. / Computers in Human Behavior 52 (2015) 115–123

drive + texting with two hands condition (M = .98 m, SD = .24 m),t(27) = �6.86, p < .001 and t(27) = �7.66, p < .001 respectively.The mean lane deviation in the drive + texting with one hand con-dition was significantly different from the drive + texting with twohands condition, t(27) = �2.56, p = .02. Participants deviated theirtrajectory from the ‘‘ideal model’’ when texting with two handsmore than when texting using one hand or driving-only.

Fig. 6. Key entry per second across conditions.

3.2. Texting performance

A two-way within-subjects analyses of variance was conductedto evaluate the task completion time between driving while tex-ting with one or two hands and texting only with one or twohands. The texting task completion time (as shown in Fig. 5) pro-duced a significant main effect of task load, F(1,27) = 175.42,p < .001, g2

p = .87, with longer completion time in the drive + tex-ting conditions (texting using either one hand or two hands)(M = 16.73 s, SD = 3.64 s) than the texting-only conditions(M = 9.57 s, SD = 2.12 s), but the main effect of texting methodsand the interaction were not significant. Participants took longer

to complete the texting task while driving regardless of whetherthey texted with either one hand or two hands.

A two-way within subjects analysis of variance was conductedto evaluate the key entry per second between driving while textingwith one or two hands and texting only with one or two hands. Asshown in Fig. 6, the key entry per second (KES) produced a margin-ally significant interaction of task load and texting methods,F(1,27) = 3.91, p = .06, g2

p = .13. The main effect of task load wassignificant, F(1,27) = 226.31, p < .001, g2

p = .89. Simple main effectshowed that the KES in the texting-only conditions (M = 1.17,SD = .19) was significantly larger than that of drive + texting condi-tions (M = .67, SD = .13), t(27) = 15.04, p < .001. The main effect oftexting methods was not significant, F(1,27) = .07, p = .79,g2

p = .003. Participants inputted more keys per second whentexting-only compared to texting while driving.

A two-way within-subjects analyses of variance was conductedto evaluate the average texting errors between texting while driv-ing with one or two hands and texting only with one or two hands.The average texting errors (as shown in Fig. 7) differed significantlyacross task load conditions, F(1,27) = 215.13, p < .001, g2

p = .89.When texting while driving, drivers produced more texting errors(M = .48, SD = .66) than the texting-only condition (M = .11,SD = .14). The main effect of texting methods and interaction effect

Fig. 7. The texting error measured by Levenshtein edit distance (He et al., 2014;Levenshtein, 1966).

Fig. 9. Standard deviation of input time interval.

J. He et al. / Computers in Human Behavior 52 (2015) 115–123 119

were not significant, F(1,27) = .59, p = .45, g2p = .02 and

F(1,27) = 1.64, p = .21, g2p = .06 respectively.

A two-way within-subjects analyses of variance was conductedto evaluate the mean input time interval between driving whiletexting with one or two hands and texting only with one or twohands. The mean input time interval (shown in Fig. 8) also differedsignificantly across task conditions, F(1,27) = 184.19, p < .001,g2

p = .87. The mean input time interval was significantly higherin the drive + texting condition (M = 1.46 s, SD = .32 s) than thatof texting-only condition (M = .82 s, SD = .17 s). Concurrent textingand driving delayed key entry by .64 s (78%) compared to thetexting-only condition. The main effect of texting methods andinteraction effect were not significant, F(1,27) = .31, p = .58,g2

p = .01 and F(1,27) = 1.61, p = .22, g2p = .06 respectively.

Participants paused more between key stroke while driving andtexting compared to texting-only.

A two-way within-subjects analyses of variance was conductedto evaluate the standard deviation of input time interval betweendriving while texting with one or two hands and texting only withone or two hands. The standard deviation of input time interval(shown in Fig. 9) differed significantly across task conditions,F(1,27) = 46.69, p < .001, g2

p = .63 and texting methods,F(1,27) = 7.36, p = .01, g2

p = .21, but the interaction was not signif-icant, F(1,27) = .67, p = .42, g2

p = .02. The standard deviation of

Fig. 8. Mean input time interval.

input time interval was significantly higher in the drive + textingcondition (M = 2.12 s, SD = .78 s) than that of texting-only condi-tion (M = .91 s, SD = 1.01 s). The standard deviation of input timeinterval was also higher in the two-hand texting condition(M = 1.71 s, SD = 1.33 s) than the one-hand texting condition(M = 1.32 s, SD = .72 s).

Participant made 10.77 key entries on average with a standarddeviation of .79. Analysis of the number of key entries did not pro-duce any significant effects, all ps > .10.

Participants revised their entries by pressing the ‘Back’ buttonon average .26 times with a standard deviation of .25. Analysis ofthe number of times that the ‘Back’ button was pressed did notyield any significant main effect, all ps > .10.

3.3. Device stability

We also measured the stability of the Android smartphonewhen participants texted while driving. Device stability was mea-sured using the standard deviation of the accelerometer values forthe X, Y, and Z coordinates of the smartphone. The Android coordi-nate system (as shown in Fig. 10) is defined relative to the screen ofthe phone in the portrait orientation. The axes are not swappedwhen the device’s screen orientation changes. The X-axis is hori-zontal and points to the right, the Y-axis is vertical and points upand the Z-axis points toward the outside of the front face of thescreen. In this system, coordinates behind the screen have negativeZ values. The X, Y, and Z vectors of the accelerator sensor wereautomatically saved to a log file when participants sent text mes-sages using the smartphone. The accelerator’s values were sampledat a rate of 100 Hz only when participants used the textingapplication.

For the standard deviation of the accelerometer values in theX-axis, the main effects of task load and texting methods were bothsignificant, F(1,27) = 47.37, p < .001, g2

p = .64 and F(1,27) = 35.20,p < .001, g2

p = .57 respectively. The interaction was also significant,F(1,27) = 39.44, p < .001, g2

p = .59. Simple main effect test showedthat when texting using one hand, the standard deviation of theaccelerator values in the X-axis was not significantly different forthe drive + texting condition (M = .58, SD = .15) and thetexting-only condition (M = .53, SD = .19), t(27) = 1.26, p = .22.When texting using two hands, the standard deviation of the accel-erator values in the X-axis was significantly larger in thedrive + texting condition (M = 1.20, SD = .47) than thetexting-only condition (M = .52, SD = .27), t(27) = 7.16, p < .001.

For the standard deviation of the accelerometer values in theY-axis, a significant interaction was found, F(1,27) = 42.74,

Fig. 10. The Android coordinate system for accelerator sensors.

Fig. 11. Device stability as measured using the accelerometer sensor values in theX, Y, and Z-axis.

120 J. He et al. / Computers in Human Behavior 52 (2015) 115–123

p < .001, g2p = .61. The main effects of task load and texting meth-

ods were not significant, F(1,27) = .05, p = .82, g2p = .002 and

F(1,27) = 2.56, p = .12, g2p = .09 respectively. Simple main effect

tests showed that when texting using one hand, the standard devi-ation of the accelerometer values in the Y-axis was significantlysmaller in the drive + texting condition (M = .72, SD = .15) thanthe texting-only condition (M = .91, SD = .38), t(27) = 2.59, p = .02.When texting using two hands, the standard deviation of theaccelerometer values in the Y-axis was significantly larger in thedrive + texting condition (M = .99, SD = .34) than the texting-onlycondition (M = .78, SD = .31), t(27) = 3.02, p = .01.

For the standard deviation of the accelerometer values in theZ-axis, a significant interaction was found, F(1,27) = 50.46,p < .001, g2

p = .65. The main effects of task load and texting meth-ods were also significant, F(1,27) = 9.26, p = .01, g2

p = .26 andF(1,27) = 13.52, p = .001, g2

p = .33 respectively. Simple main effecttests showed that when texting using one hand, the standard devi-ation of the accelerometer values in the Z-axis was significantlysmaller in the drive + texting condition (M = .74, SD = .14) thanthe texting-only condition (M = .80, SD = .26), t(27) = 1.29, p = .21.When texting using two hands, the standard deviation of theaccelerometer values in the Z-axis was significantly larger in thedrive + texting condition (M = 1.07, SD = .33) than thetexting-only condition (M = .71, SD = .19), t(27) = 5.25, p < .001.See Fig. 11 for the device stability as measured using theaccelerometer sensor values in the X, Y, and Z-axis.

The interaction effects in the above analyses is because theinstability of the devices (as measured by the standard deviationof accelerator values in the X, Y, and Z-axis) is more pronouncedin the texting while driving using two hands condition than thetexting while driving using one hand condition. More instabilityto hold the smartphone in the condition when drivers texting usingtwo hands was because the inability to share manual resources(e.g. the hands). In contrast, drivers in the one hand texting condi-tion can hold the phone in one hand and control the steering wheelusing the other hand. Thus, for drivers who texted with two hands,texting while driving significantly increased the instability of thesmartphone, which indicates more manual interference of textingand driving and higher chances of dropping or even breaking thesmartphone when driving (see Fig. 11).

3.4. Detect texting while driving

To reduce the risks of texting while driving, it is important todetect the occurrence of such driving behaviors. We used a logisticregression to detect texting while driving, using texting data col-lected from the smartphone. The variables in the logistic regressioninclude the number of key entries, mean input time interval, stan-dard deviation of input time interval, the number of key entries,

and the number of times the ‘Back’ button was pressed. The textingtask completion time and texting errors were hard to define in nat-ural free-style texting, thus the two variables were not included inthe analysis. The dependent variable was the texting conditions,either drive + texting or texting-only.

To predict texting while driving we need a ‘‘training data set’’ todevelop the logistic regression model. The training data set can begeneric, that is based on the data from a group of participants or

Table 2Confusion matrix to predict texting while driving based on user-specific training data.

Predicted

Texting-only Drive + texting

Observed Texting-only 5.45 0.69Drive + texting 1.52 9.66

J. He et al. / Computers in Human Behavior 52 (2015) 115–123 121

individualized meaning that a participant’s data is used to derive atheir own user specific logistic regression model.

If a common training data set can be used to monitor textingwhile driving reliably for all subjects, the proposed application tomonitor texting while driving can be developed more easily, with-out the necessity to consider individual texting behaviors and doesnot require the collection of user specific data sets for model devel-opment. A user-specific training set may achieve higher predictionperformance than a model based on a generic data set. Weexplored the prediction performance using both generic trainingset and individualized training set.

3.4.1. Prediction based on generic training setThe experiment produced a total of 1748 texting instances.

1224 texting instances (70% of all instances) were randomlyselected as the training set. The other 524 instances (30%) wereused as the prediction set.

The best logistic regression model based on generic trainingdata set is:

Task load ¼ 2:31� :0030�mean input time interval� :0016

� standard deviation of input time interval

Where for task load, 0 indicates texting-only condition and 1 indi-cates drive + texting condition.

The classification accuracy for the generic training set is shownin Table 1. An accuracy as high as 88.5% was achieved using textingbehavior data to classify texting while driving from thetexting-only condition. The sensitivity of the model based on gen-eric training set is .88, and the specificity of the model is .94. Ouraccuracy in detecting texting while driving is comparable to orsometimes higher than other attempts to detect cognitive distrac-tion (Liang, Reyes, & Lee, 2007) and visual distraction (Kutila,Jokela, Markkula, & Rue, 2007). For example, Liang, Reyes et al.(2007) achieved an average accuracy of 81.1% in detecting drivercognitive distraction using Support Vector Machines. Our highdetection accuracy suggests it is possible to develop a smartphoneapplication to detect texting while driving using texting behavioraldata, such as mean and standard deviation of the input timeinterval.

3.4.2. Prediction based on user-specific training setUser-specific training set can potential improve model predic-

tion by learning the individual difference in texting behaviors. Toexplore to what extent user-specific training set can improve ourability to predict texting while driving, we also trainedtwenty-eight logistic regression models for each subject. Each dri-ver produced about 58.07 texting instances in total. 70% of the tex-ting instances for each subject were used as the training data setand the other 30% texting instances were used as the testing dataset.

The classification accuracy for the user-specific training data setis shown in Table 2. The average accuracy as high as 89.7% wasachieved using texting behavior data to classify texting while driv-ing from the texting-only condition. The sensitivity of the modelbased on generic training set is .94, and the specificity of the modelis .82. The high accuracy suggests it is possible to develop a smart-phone application to detect texting while driving using texting

Table 1Confusion matrix to predict texting while driving based on generic training data.

Predicted

Texting-only Drive + texting

Observed Texting-only 233 29Drive + texting 16 210

behavioral data, such as mean and standard deviation of the inputtime interval.

The above data suggest that we can use texting behavior data topredict texting while driving reliably with at least 88.5% accuracy.The accuracy using generic training data set (88.5%) is only slightlylower than the accuracy based on user-specific training data set(89.7%). The minor increase in the prediction accuracy ofuser-specific training data should be a result of the significant largeincrease in the mean and standard deviation of texting input inter-vals. This result suggests that we can just use a generic logisticregression model to predict texting while driving, which is easierto implement than a user-specific training data set.

4. Discussion

Using a classic Lane Change Task and smartphone technology,we found evidence of mutual interference between driving andtexting. Texting impaired driving by increasing lane deviations,which resonates with previous studies showing that textingimpaired driving performance (Drews et al., 2009; He et al.,2014; Hosking et al., 2009). To our best knowledge, existing studieshave often ignored the other aspect of dual-task performance, thatis, how driving influences texting? In addition to the finding thattexting impairs driving, our study also showed that drivingimpaired texting by slowing text entry and increasing textingerrors. The mutual interference of texting and driving contributesto our understanding of dual-task performance. Interestingly, wecan use a logistic regression model based on texting behaviors todetect texting while driving with accuracy as high as 88.5%.

Texting can increase a driver’s crash risks and endanger theirlives. Drivers may continue to text for a variety of reasons.Drivers may read or send texts out of habit, regardless their subjec-tive perception of the risks of such behaviors (Bayer & Campbell,2012) and are more likely to take a phone call if they believe the callis important (Nelson, Atchley, & Little, 2009). They may send textsto reduce unpleasant emotions (Feldman, Greeson, Renna, &Robbins-Monteith, 2011) or maintain a social relationship by reply-ing promptly to a text message. For these drivers, a clear messagethat ‘‘driving and texting interfere mutually’’ should be deliveredin social campaigns to further discourage texting while driving.

Practically, a better understanding of texting performance mayinform the development of smartphone applications that candetect and perhaps reduce texting while driving. Several research-ers are beginning to explore smartphone applications that canmonitor impaired driving performance (Dai et al., 2010;Hammoud & Zhang, 2008; He et al., 2013; Johnson & Trivedi,2011; You et al., 2013; Zeeman & Booysen, 2013;), such as drunk,distracted, aggressive and drowsy driving. However, no attemptshave been made to detect texting while driving yet. Our studydemonstrated that a logistic regression model using texting behav-ioral data (such as mean and standard deviation of input timeinterval) can detect texting while driving with accuracy as highas 88.5%. This finding provides a theoretical support to develop asmartphone application, which can detect and prevent textingwhile driving. A smartphone-based application to monitor textingwhile driving has the advantage to be easier to implement (Eren,Makinist, Akin, & Yilmaz, 2012), while it is very costly for lawenforcement or social campaigns to prevent texting while driving.

122 J. He et al. / Computers in Human Behavior 52 (2015) 115–123

More studies are needed to improve our training data set and algo-rithm to detect texting while driving. For example, an on-roaddriving study using a natural texting task is needed to collect valu-able training data set to test and improve our algorithm.

Many cell phone carriers and cell phone manufacturers offer‘‘Drive Mode’’ applications, which allow users to block incomingcalls and messages or send automatic responses. These applica-tions often require users to initiate the drive mode before drivingand turn it off after driving, which is a bothersome process andlimits the use of these applications. Automatic detection of textingwhile driving can improve existing ‘‘Drive Mode’’ applications. Ifsmartphone applications incorporate automatic detection of tex-ting while driving and stop or delay messages from distracting dri-vers, these applications can be more effective in reducing textingwhile driving than existing applications. Our study demonstratedthat we can reliably detect texting while driving using generictraining set based on a group of users. Our finding provides newknowledge to improve existing ‘‘Drive Mode’’ applications. Futurestudies shall investigate whether other information, includingdriving dynamics, can be used to develop smartphone-based tech-nology to detect texting while driving. For example, GPS informa-tion can be used to calculate travelling velocity, which could thenbe used as a precondition for identifying whether the user is tex-ting while driving. The smartphone application will detect textingwhile driving only when users travel at higher speeds, above20 mph, which would greatly improve the performance of theapplication. Researchers have focused on how to use smartphonecapabilities to improve driving safety (He et al., 2012, 2013,2014; Ren, Wang, & He, 2013; Wang, Cardone, Corradi, Torresani,& Campbell, 2012). To our best knowledge, this is one of the firstpapers to demonstrate the possibility of using smartphone tech-nology to detect texting while driving. This technology is an inno-vative way to mitigate the risks of texting while driving, whichcomplements the existing social and legal approaches.

With a better understanding of texting and driving behaviorsand the rapid advancement of smartphone technology, smart-phones can be used to make driving safer, instead of more danger-ous (He et al., 2013; Ren et al., 2013). As Dr. Adesman stated,‘‘Technological solutions will likely need to be developed to signifi-cantly reduce the frequency of texting while driving’’ (AmericanAcademy of Pediatrics, 2013; Lindqvist & Hong, 2011).

Acknowledgements

We appreciate the valuable suggestions from Frank Schieber,Yulan Liang, Chun Wang, and Evan Palmer. We also acknowledgethe help in data collection from the undergraduate research assis-tants Colton Turner and Kirsten Turner. The authors wouldacknowledge to the National Natural Science Foundation of Chinawith Grant 71401004 and the U.S. Department of Transportation(DOT) through the University Transportation Centersprogram sponsored by Research and Innovative TechnologyAdministration (RITA).

References

Allport, D. A., Antonis, B., & Reynolds, P. (1972). On the division of attention: Adisproof of the single channel hypothesis. The Quarterly Journal of ExperimentalPsychology, 24(2), 225–235.

Alosco, M. L., Spitznagel, M. B., Fischer, K. H., Miller, L. A., Pillai, V., Hughes, J., et al.(2012). Both texting and eating are associated with impaired simulated drivingperformance. Traffic Injury Prevention, 13(5), 468–475. http://dx.doi.org/10.1080/15389588.2012.676697.

American Academy of Pediatrics (2013). Don’t txt n drive: Teens not getting msg:43 percent of youths admit to texting while driving. Science Daily. 4 May 2013.

Atchley, P., Atwood, S., & Boulton, A. (2011). The choice to text and drive in youngerdrivers: Behavior may shape attitude. Accident Analysis & Prevention, 43(1),134–142. http://dx.doi.org/10.1016/j.aap.2010.08.003.

Atchley, P., Hadlock, C., & Lane, S. (2012). Stuck in the 70s: The role of social normsin distracted driving. Accident Analysis & Prevention, 48, 279–284. http://dx.doi.org/10.1016/j.aap.2012.01.026.

Bayer, J. B., & Campbell, S. W. (2012). Texting while driving onautomatic: Considering the frequency-independent side of habit. Computers inHuman Behavior, 28(6), 2083–2090. http://dx.doi.org/10.1016/j.chb.2012.06.012.

Burge, R., & Chaparro, A. (2012). The effects of texting and driving on hazardperception. In Proceedings of the human factors and ergonomics society annualmeeting (Vol. 56, No. 1, pp. 715–719) doi: http://dx.doi.org/10.1177/1071181312561149.

Caird, J. K., Johnston, K. A., Willness, C. R., Asbridge, M., & Steel, P. (2014). A meta-analysis of the effects of texting on driving. Accident Analysis & Prevention, 71,311–318. http://dx.doi.org/10.1016/j.aap.2014.06.005.

Crandall, J. M., & Chaparro, A. (2012). Driver distraction: Effects of text entrymethods on driving performance. In Proceedings of the human factors andergonomics society annual meeting (Vol. 56, No. 1, pp. 1693–1697). doi: http://dx.doi.org/10.1177/1071181312561339.

Dai, J., Teng, J., Bai, X., Shen, Z., & Xuan, D. (2010). Mobile phone based drunk drivingdetection. In Pervasive computing technologies for healthcare (PervasiveHealth).doi: http://dx.doi.org/10.4108/ICST.PERVASIVEHEALTH2010.8901.

Drews, F. A., Yazdani, H., Godfrey, C. N., Cooper, J. M., & Strayer, D. L. (2009). Textmessaging during simulated driving. Human Factors, 51(5), 762–770.

Eren, H., Makinist, S., Akin, E., & Yilmaz, A. (2012). Estimating driving behavior by asmartphone. In 2012 IEEE intelligent vehicles symposium (IV) (pp. 234–239). doi:http://dx.doi.org/10.1109/IVS.2012.6232298.

Feldman, G., Greeson, J., Renna, M., & Robbins-Monteith, K. (2011). Mindfulnesspredicts less texting while driving among young adults: Examining attention-and emotion-regulation motives as potential mediators. Personality andIndividual Differences, 51(7), 856–861. http://dx.doi.org/10.1016/j.paid.2011.07.020.

Ferguson, C. J. (2009). An effect size primer: A guide for clinicians and researchers.Professional Psychology: Research and Practice, 40(5), 532–538.

Ferris, F. L., Kassoff, A., Bresnick, G. H., & Bailey, I. (1982). New visual acuity chartsfor clinical research. American Journal of Ophthalmology, 94(1), 91–96.

Hammoud, R. I., & Zhang, H. (2008). Alertometer: Detecting and mitigating driverdrowsiness and fatigue using an integrated human factors and computer visionapproach. In R. I. Hammoud (Ed.), Passive Eye Monitoring (pp. 301–321). BerlinHeidelberg: Springer. Retrieved from http://link.springer.com/chapter/10.1007/978-3-540-75412-1_14.

Harrison, M. A. (2011). College students’ prevalence and perceptions of textmessaging while driving. Accident Analysis & Prevention, 43(4), 1516–1520.http://dx.doi.org/10.1016/j.aap.2011.03.003.

He, J., Chaparro, A., Nguyen, B., Burge, R. J., Crandall, J., Chaparro, B., et al. (2014).Texting while driving: Is speech-based text entry less risky than handheld textentry? Accident Analysis & Prevention, 72, 287–295. http://dx.doi.org/10.1016/j.aap.2014.07.014.

He, J., Fields, B.M., Roberson, S., Cielocha, S., Peng, J., & Coltea, J. (2012). ‘‘System andmethod to monitor and reduce vehicle operator impairment.’’ U.S. PatentApplication 13/717,506, filed December 17, 2012.

He, J., Roberson, S., Fields, B., Peng, J., Cielocha, S., & Coltea, J. (2013). Fatiguedetection using smartphones. Journal of Ergonomics, 03(03). http://dx.doi.org/10.4172/2165-7556.1000120.

Hofmann, P., Rinkenauer, G., & Gude, D. (2008). Head-up-displays support responsepreparation in a Lane Change Task. In Proceedings of the human factors andergonomics society annual meeting (Vol. 52, No. 18, pp. 1233–1237). doi: http://dx.doi.org/10.1177/154193120805201817.

Horrey, W. J., & Wickens, C. D. (2004). Driving and side task performance: Theeffects of display clutter, separation, and modality. Human Factors, 46(4),611–624.

Hosking, S. G., Young, K. L., & Regan, M. A. (2009). The effects of text messaging onyoung drivers. Human Factors, 51(4), 582–592.

Huemer, A. K., & Vollrath, M. (2010). Alcohol-related impairment in the LaneChange Task. Accident Analysis & Prevention, 42(6), 1983–1988. http://dx.doi.org/10.1016/j.aap.2010.06.005.

Jacobson, P. D., & Gostin, L. O. (2010). Reducing distracted driving: regulationand education to avert traffic injuries and fatalities. JAMA, 303(14), 1419–1420.

Johnson, D. A., & Trivedi, M. M. (2011). Driving style recognition using a smartphoneas a sensor platform. In 2011 14th International IEEE Conference on IntelligentTransportation Systems (ITSC) (pp. 1609–1615). doi: http://dx.doi.org/10.1109/ITSC.2011.6083078.

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 andexperienced drivers. New England Journal of Medicine, 370(1), 54–59.

Kutila, M., Jokela, M., Markkula, G., & Rue, M. R. (2007). Driver distraction detectionwith a camera vision system. IEEE International Conference on Image Processing,6, 201–204.

Leung, S., Croft, R. J., Jackson, M. L., Howard, M. E., & Mckenzie, R. J. (2012). Acomparison of the effect of mobile phone use and alcohol consumption ondriving simulation performance. Traffic Injury Prevention, 13(6), 566–574.http://dx.doi.org/10.1080/15389588.2012.683118.

Levenshtein, V. I. (1966). Binary codes capable of correcting deletions, insertionsand reversals. Soviet Physics Doklady, Johnson, 10, 707.

Levy, J., & Pashler, H. (2008). Task prioritization in multitasking during driving:Opportunity to abort a concurrent task does not insulate braking responsesfrom dual-task slowing. Applied Cognitive Psychology, 22, 507–525.

J. He et al. / Computers in Human Behavior 52 (2015) 115–123 123

Liang, Y., Lee, J., & Reyes, M. (2007). Nonintrusive detection of driver cognitivedistraction in real time using Bayesian Networks. Transportation ResearchRecord: Journal of the Transportation Research Board, 2018, 1–8. http://dx.doi.org/10.3141/2018-01.

Liang, Y., Reyes, M. L., & Lee, J. D. (2007). Real-time detection of driver cognitivedistraction using support vector machines. IEEE Transactions on IntelligentTransportation Systems, 8(2), 340–350.

Libby, D., Chaparro, A., & He, J. (2013). Distracted while driving: A comparison of theeffects of texting and talking on a cell phone. In Proceedings of the human factorsand ergonomics society annual meeting, (Vol. 57, No. 1, pp. 1874–1878) doi:http://dx.doi.org/10.1177/1541931213571418.

Lindqvist, J., & Hong, J. (2011). Undistracted driving: A mobile phone that doesn’tdistract. In Proceedings of the 12th workshop on mobile computing systems andapplications (pp. 70–75). New York, NY, USA: ACM. http://dx.doi.org/10.1145/2184489.2184504.

Loftus, G. R., & Masson, M. E. (1994). Using confidence intervals in within-subjectdesigns. Psychonomic Bulletin & Review, 1(4), 476–490.

Mattes, S., & Hallén, A. (2008). Surrogate distraction measurement techniques. In K.Young, J. D. Lee, & M. A. Regan (Eds.), Driver distraction: Theory, effects, andmitigation. CRC Press.

McCartt, A. T., & Geary, L. L. (2004). Longer term effects of New York State’s law ondrivers’ handheld cell phone use. Injury Prevention, 10(1), 11–15. http://dx.doi.org/10.1136/ip.2003.003731.

Mccartt, A. T., Hellinga, L. A., Strouse, L. M., & Farmer, C. M. (2010). Long-term effectsof handheld cell phone laws on driver handheld cell phone use. Traffic InjuryPrevention, 11(2), 133–141.

McKeever, J. D., Schultheis, M. T., Padmanaban, V., & Blasco, A. (2013). Driverperformance while texting: Even a little is too much. Traffic Injury Prevention,14(2), 132–137. http://dx.doi.org/10.1080/15389588.2012.699695.

National Safety Council (2012). Annual estimate of cell phone crashes 2012.Retrieved on 20.05.14 from <http://www.nsc.org/DistractedDrivingDocuments/Attributable-Risk-Estimate.pdf>.

Nelson, E., Atchley, P., & Little, T. D. (2009). The effects of perception of risk andimportance of answering and initiating a cellular phone call while driving.Accident Analysis & Prevention, 41(3), 438–444. http://dx.doi.org/10.1016/j.aap.2009.01.006.

Nemme, H. E., & White, K. M. (2010). Texting while driving: Psychosocial influenceson young people’s texting intentions and behaviour. Accident Ana⁄⁄⁄lysis &Prevention, 42(4), 1257–1265. http://dx.doi.org/10.1016/j.aap.2010.01.019.

Olson, R. L., Hanowski, R. J., Hickman, J. S., & Bocanegra, J. L. (2009). Driver distractionin commercial vehicle operations (No. FMCSA-RRR-09-042). Federal MotorCarrier Safety Administration, Washington, DC.

Owens, J. M., McLaughlin, S. B., & Sudweeks, J. (2011). Driver performance while textmessaging using handheld and in-vehicle systems. Accident Analysis &Prevention, 43(3), 939–947. http://dx.doi.org/10.1016/j.aap.2010.11.019.

Pickrell, T. M., & Ye, T. J. (2013). Driver electronic device use in 2011. (Report No.DOT HS 811 719). Washington, DC: National Highway Traffic SafetyAdministration.

Reed, M. P., & Green, P. A. (1999). Comparison of driving performance on-road andin a low-cost simulator using a concurrent telephone dialling task. Ergonomics,42(8), 1015–1037.

Ren, Z., Wang, C., & He, J. (2013). Vehicle detection using Android smartphones. The7th international driving symposium on human factors in driver assessment,training, and vehicle design. Retrieved from <http://trid.trb.org/view.aspx?id=1261925>.

Rudin-Brown, C. M., Young, K. L., Patten, C., Lenné, M. G., & Ceci, R. (2013). Driverdistraction in an unusual environment: Effects of text-messaging in tunnels.Accident Analysis & Prevention, 50, 122–129. http://dx.doi.org/10.1016/j.aap.2012.04.002.

Salvucci, D. D., & Macuga, K. L. (2002). Predicting the effects ofcellular-phone dialing on driver performance. Cognitive Systems Research, 3(1),95–102.

Wang, T., Cardone, G., Corradi, A., Torresani, L., & Campbell, A. T. (2012). WalkSafe: Apedestrian safety app for mobile phone users who walk and talk while crossingroads. In Proceedings of the Twelfth Workshop on Mobile Computing Systems&#38; Applications (pp. 5:1–5:6). New York, NY, USA: ACM. http://dx.doi.org/10.1145/2162081.2162089.

Wickens, C. D. (2002). Multiple resources and performance prediction. TheoreticalIssues in Ergonomics Science, 3(2), 159–177. http://dx.doi.org/10.1080/14639220210123806.

Wilson, F., & Stimpson, J. (2010). Trends in fatalities from distracted driving in theUnited States, 1999 to 2008. American Journal of Public Health, 100(11),2213–2219. http://dx.doi.org/10.2105/AJPH.2009.187179.

You, C.-W., Lane, N. D., Chen, F., Wang, R., Chen, Z., Bao, T. J., et al. (2013). CarSafeapp: Alerting drowsy and distracted drivers using dual cameras onsmartphones. In Proceeding of the 11th Annual International Conference onMobile Systems, Applications, and Services (pp. 13–26). New York, NY, USA: ACM.http://dx.doi.org/10.1145/2462456.2465428.

Zeeman, A. S., & Booysen, M. J. (2013). Combining speed and acceleration to detectreckless driving in the informal public transport industry. Retrieved from<http://scholar.sun.ac.za/handle/10019.1/86105>.