Naturalistic study on the usage of smartphone applications among Finnish drivers

This is a self-archived version of an original article. This version may differ from the original in pagination and typographic details. Author(s): Title: Year: Version: Copyright: Rights: Rights url: Please cite the original version: CC BY-NC-ND 4.0 https://creativecommons.org/licenses/by-nc-nd/4.0/ Naturalistic study on the usage of smartphone applications among Finnish drivers © 2018 Elsevier Ltd. Accepted version (Final draft) Kujala, Tuomo; Mäkelä, Jakke Kujala, T., & Mäkelä, J. (2018). Naturalistic study on the usage of smartphone applications among Finnish drivers. Accident Analysis and Prevention, 115, 53-61. https://doi.org/10.1016/j.aap.2018.03.011 2018 1 Naturalistic study on the usage of smartphone applications among Finnish 1 drivers 2 3 Tuomo Kujala 1* and Jakke Mäkelä 1 4 5 1 University of Jyväskylä, Jyväskylä, Finland 6 *Corresponding author: tuomo.kujal[email protected], P.O. Box 35, 40014 University of Jyväskylä, 7 Finland 8 9 Abstract 10 11 We present results from a naturalistic study that tracked how Finnish drivers use their smartphones while 12 driving. We monitored 30 heavy-user drivers in Finland in June-September 2016, and recorded the times 13 when they used their phones, the application used at the time of touch, and the location and speed of the 14 car. Touches were used as a proxy for estimating visual distraction due to visual-manual tasks. Our data 15 set allows us to determine whether drivers use their phones differently on different road types (highway, 16 main road, local rural road, urban road). We found that the road type has an effect on phone use but the 17 effect is to the opposite direction than expected. The drivers produced more touches per hour on urban 18 roads but the instances of use tend to be shorter than on the highway or on main roads. We also 19 collected statistics on the applications that were used. By far the largest amount of distraction is caused by 20 the WhatsApp messaging service, used by a majority of the drivers. An instance of WhatsApp use 21 included a median of 8 touches, and had a median duration of 35 seconds. By contrast, navigation 22 application use included a median of 3 touches and lasted for 11 seconds. The findings suggest that the 23 Finnish smartphone heavy-users do not actively modulate their phone use based on the demands of the 24 traffic conditions and that the greatest risk from smartphone use may be currently caused by messaging 25 applications. 26 27 Keywords: distraction; smartphone; application; road type; attentional demand; behavioural adaptation 28 2 1. Introduction 29 Driver distraction by mobile phone use has been associated with increased safety-critical incident risk in 30 traffic (e.g., [7][15]). Although it is well known that many drivers use their smartphones to do various 31 tasks while driving (e.g., [7][15]), little to nothing is known about the actual applications they use or the 32 exact traffic conditions in which they use those applications. This information is important for determining 33 the actual risk level caused by the distraction. A given application, such as texting, may cause a very 34 serious risk of accident on a busy and congested urban road. The risk may be considerably smaller when 35 driving along a straight and nearly empty highway. 36 In this paper, we present findings from naturalistic driving data that enabled us to determine what 37 applications drivers use and whether they use applications differently in different driving scenarios. Ideally, 38 we would be able to compare what applications are used on highways, major roads, minor roads, and 39 urban traffic. The greatest limitation is that no data on the congestion or real-time traffic density is 40 available. However, the road type [8] and the associated photo material recorded during the phone use 41 provide us at least a rough indication of the visual demands of the driving scenario. 42 Although there is no exact way to determine the visual-manual distraction caused by any given 43 application, we used the number of touches on the smartphone as a proxy. A touch on a touch screen is 44 almost always accompanied by a glance on the touch screen due to the limited haptic feedback of the 45 device [2][14]. 46 Among others, Victor et al. [15] have suggested that a series of glances away from the road in a 47 short period of time may lead to safety-critical uncertainty of the task-relevant road events, even if the 48 glances off road were brief. Therefore, an application that requires a large number of touches in a short 49 time period can be considered to cause high cognitive demand on the driver. On the other hand, an 50 instance of use, of which duration is long, means that the driver is distracted for a longer time period. 51 Longer in-car tasks have been associated with increased probability of increased individual glance 52 durations off road [10][11]. 53 3 There is evidence from naturalistic driving (e.g. [16]) as well as from on-road (e.g. [17]) and 54 simulator studies (e.g. [12]) suggesting that drivers adapt their off-road glancing behaviours according to 55 the dynamic demands of the driving task. Drivers tend to decrease their off-road glance durations and the 56 number of off-road glances when the demands of the driving task increase. The naturalistic field study of 57 Metz et al. [13] (maneuvering, German drivers) and the video-clip based study of Hancox et al. [4] also 58 suggest that drivers tend to attend to distracting activities in a situationally aware manner. 59 Based on the previous research findings, our preliminary hypothesis was that at least experienced 60 drivers should have developed a sense of acceptable risk levels, and would be able to adapt their 61 smartphone usage to the demands of the given driving conditions. We would thus expect to see differences 62 in smartphone usage between road types. In particular, we hypothesized that there would be less touches 63 on the phone and the applications would be used less often in high-demand driving scenarios. It was our 64 expectation that urban roads would present the need for most vigilance due to intersections, traffic lights 65 and other traffic (including cars, bicycles and pedestrians), and hence we would see less phone use in 66 urban conditions than on highway or main roads, even though the highways and main roads have higher 67 nominal speed limits. In order to analyse if different levels of distraction can be associated with different 68 smartphone applications, we studied the most used applications on the road as well as the number of 69 touches and the durations of application use instances. 70 4 2. Materials and Methods 71 2.1 Hardware and Software 72 The results presented here are a subset of a larger experiment that studied the effects of distraction 73 warnings on smartphone use while driving. The presented data are solely from the control part of the 74 experiments, when phone usage data was simply collected without any interventions. The control phase 75 was finished by September 2016, but the experimental phase continued until December 2016, ending to a 76 web questionnaire (self-reported car and phone data presented here). 77 The data were collected using custom software developed by Ficonic Solutions Ltd, located in 78 Jyväskylä, Finland. The software consisted of two parts: a “Watcher” application running on Samsung 79 XCover 3 smartphones that were installed in the volunteers’ cars dashboards with a double-suction cup 80 windshield car holder (see Figure 1), and a small “Observer” application that was installed on all Android 81 phones and other Android devices which the drivers reported to use while driving. The Watcher 82 application created a Wi-Fi hotspot, onto which the Observer phones connected when within range. 83 84 Fig. 1. The position of the dashboard smartphone with the Watcher application in a participant's car. 85 5 The Watcher phones had a continuous cellular network connection to allow enhanced GPS 86 positioning. Data were uploaded to the remote server via 3G or 4G connection, depending on the 87 connectivity at the area, whenever the Watcher application was on standby. The Watcher application 88 recorded the GPS position at one-second intervals whenever the car was in motion. The in-built power89 savings system in the Android version meant that whenever the car was stationary, no GPS positions were 90 recorded. 91 In order to enable the location-based warnings in the second phase of the experiment, the Watcher 92 application constantly mapped the position of the car against the Finnish national Digiroad map data set 93 (http://www.liikennevirasto.fi/web/en/open-data/digiroad), and determined the road on which the car was 94 for any given GPS fix. The application also collected acceleration data on three axles but this data is not 95 analyzed in this paper. 96 The Observer background application in the drivers’ Android devices worked by creating a 97 transparent layer over the other applications. A touch on the phone was thus recorded by the Observer 98 application. A flag about every touch was immediately sent over the Wi-Fi network to the Watcher phone, 99 including information about the Android front application (FrontApp) that was running at the moment. 100 The Watcher application took a photo with the back camera of the Android phone by each touch on the 101 driver's phone. The camera was positioned and secured to the windshield holder and dashboard in a way 102 that the camera had clear visibility to the road environment in front of the vehicle (see Figure 1). 103 The car models used by the participants during the study are listed in Table 1 and their Android 104 devices with the Observer application in Table 2 (based on self-reporting). One participant could have up 105 to three cars and three Android devices in use during the research. If the car was changed in the middle of 106 the study, the research equipment was moved to the new car. Most of the participants' cars were equipped 107 with manual transmission (29/40, 72.5 %). 108 109 6 Table 1 Participants' car models in the study. 110 111 Make Model Year (where available) Transmission Audi A3 NA manual Audi A4 NA manual Audi A4 NA manual Audi A5 NA automatic Audi A6 2004 automatic BMW 518 2015 automatic BMW 320 2013 automatic Chevrolet Suburban 1996 automatic Citroen C5 2008 automatic Ford Fiesta NA manual Ford Focus NA manual Ford Focus 2000 manual Ford Ranger 2016 automatic Nissan Qashqai 2012 manual Nissan Qashqai 2015 manual Nissan Primera 1998 manual Opel Astra 2004 manual Peugeot 206 SW 2004 manual Peugeot 308 2008 manual Seat Altea 2005 automatic Skoda Octavia 2006 manual Skoda Octavia 2012 automatic Skoda Octavia SW 2015 manual Skoda Rapid 2014 manual Smart ForFour 2006 manual Subaru Forester 2010 manual Toyota Avensis 1998 manual Toyota Corolla NA manual Toyota Corolla NA manual Volkswagen Caddy Maxi NA manual Volkswagen Golf 2000 manual Volkswagen Golf 2000 manual Volkswagen Golf SW 2003 manual Volkswagen Passat 1998 manual Volkswagen Passat 2000 manual Volkswagen Polo NA manual Volkswagen Transporter 1996 manual Volvo S80 1999 manual Volvo V70 2002 manual Volvo V70 2012 automatic 112 7 Table 2 Participants' Android devices in the study. 113 114 Make Model Count CAT B15Q 1 HTC Desire 1 Huawei Honor Holly 1 Huawei Honor 7 3 Huawei Honor 8 1 Huawei Nexus 6P 1 LG GFlex 1 LG Nexus 5x 2 Samsung Galaxy S4 4 Samsung Galaxy S5 3 Samsung Galaxy S5 Mini 1 Samsung Galaxy S6 Edge 2 Samsung Galaxy S7 1 Samsung Galaxy XCover 2 2 Samsung Galaxy J5 1 Samsung Galaxy Alpha 1 Sony Xperia Z2 1 Sony Xperia Z3 1 Sony Xperia Z3+ 3 Sony Xperia Z5 1 Sony Xperia X Performance 1 115 2.2 Participants 116 The number of volunteers recruited in the study via convenience sampling was initially 31, starting in June 117 2016. One participant dropped out of the study before sufficient control data could be collected. The total 118 number of drivers in this study is therefore N=30 (22 M, 8 F; median age 37, mean age 39, SD 12.2, range 119 18-64). Participants were recruited via ads in newspapers and social media. The recruitment ad required 120 volunteers to drive “a lot” and to “regularly” use their smartphones while driving, but no quantitative 121 criteria were given. Well over 200 applications were eventually received. The final participants were 122 selected based on multiple criteria, including how much they reported themselves to drive per year. Those 123 drivers were favored who reported driving in both urban and rural areas. An attempt was made to choose 124 as many women as possible, and to achieve a wide spread in ages. The participants got to keep the 125 Android smartphone, the car charger and the car holder as a reward for participation after the study. All 126 the participants signed a written informed consent document prior to participation and they were allowed 127 8 to withdraw from the study at any point. The University of Jyväskylä Ethical Committee was enquired for 128 the need to have an ethical review for the study and the study was approved without a formal review. 129 2.3 Data Collection and Analysis 130 The initially planned number of days for the drivers to spend in the control stage was 63 days. However, 131 this was increased to 84 days (12 weeks) for the great majority of the drivers. In practice, there was a large 132 variation in the number of days that drivers actually drove during the control period, enhanced by technical 133 problems, which caused periods of data blackout. The mean number of days was 36.7 (SD 18.3), with 134 median of 39 and range from 7 to 68. 135 The data used in this paper are only for cases where the car was traveling with a speed of at least 2 136 m/s (7.2 kmh). This was the lower speed limit that could be reliably detected by the GPS. Data were 137 collected separately in cases where the car was stationary, but this data is not included in this paper. In 138 general, there was no significant difference between the number of touches while driving or while 139 stationary (p = .870). The number of hours spent driving varied from 7 to 167 hours, with a median of 50 140 hours (mean 56 hours). The total number of touches recorded per driver while driving also varied 141 enormously. The range was 132 to 22,337, with three drivers recording over 10,000 touches during the 142 control period. The median for touches was 1,538. 143 In order to analyze the reliability of the road classification system used in the analyses, we analyzed 144 the vehicle speeds at touch on the different road types and compared these to the roads' nominal speed 145 limits. In addition, for analyzing if the traffic densities differed by road type during the phone use, we 146 made an analysis of traffic densities in the photos taken by each touch on the drivers' phones. A random 147 sample of 310 photos collected during the control stage in between June and September 2016 was 148 manually scored by a research assistant for the number of vehicles and light traffic (pedestrians, cyclists, 149 small motorized vehicles) visible in the photo. Only those road users visible in the photos who had an 150 access to the road the participant was driving, were included in the data (e.g., on-coming cars which were 151 on an adjacent lane with a fence are excluded). Only photos with a clear view of the road environment 152 15 3.4 Touches per Hour by Road Type 244 Statistics for a given road type were compiled only when there were at least 50 touches by a driver for a 245 given road type. The number of touches in in each road can be normalized to the amount of time spent on 246 the given road type. Whenever the car is in motion, the location is updated once a second. Touches per 247 hour can thus be estimated by dividing the number of touches by the number of location fixes and 248 multiplying by 3600.. 249 As implied by Figure 6, the touches per hour are very strongly non-Gaussian. For the highway data, 250 the Shapiro-Wilks test gives W=0.62, which implies non-normality with p<0.01. The Wilcoxon rank sum 251 test shows no statistically significant differences between the medians for the various road types; the 252 highway versus local has p=0.21, while the rest are much higher (Table 5). 253 254 Fig. 6. Touches per hour by the four different road types (data was included only if the driver had at least 255 50 touches on the given road type). 256 257 258 16 Table 5 Touches per hour on the different road types 259 260 Road type Touches per hour, median 15th percentile 85th percentile Highway 55 28 148 Main 63 16 131 Local 49 17 164 Urban 67 35 153 261 Within-subject variations were controlled by dividing each driver’s touch density (i.e. touches per 262 hour) by the driver’s mean touch density over the whole experiment (Figure 7, Table 6). Wilcoxon rank 263 sum tests (Table 7) show that the normalized median of the urban roads is larger than the other road types, 264 the effect sizes being large. Thus, drivers tended to touch their phones more while driving on urban roads. 265 266 Fig. 7. Normalized touches per hour by road type. 267 268 17 Table 6 Normalized touches per hour on the different road types 269 270 Road type Norm touches per hour, median 15th percentile 85th percentile Highway 0.80 0.58 1.25 Main 0.89 0.54 1.16 Local 0.87 0.67 1,17 Urban 1.31 0.86 1.692 271 272 Table 7. Results of Wilcoxon rank sum tests for normalized touches per hour by road type 273 274 Road type Main Local Urban Highway p=0.62 p=0.78 p<0.001* (d=1.05) Main p=0.25 p=<0.001* (d=1.08) Local p<0.001* (d=0.94) *Significant at p<0.05. Cohen’s d is calculated for the statistically significant cases. 275 276 3.5 Analysis of Application Use Instances 277 We defined application use instances by clustering the touches. If two touches were separated by less than 278 30 seconds, they were considered to be part of the same cluster. The 30-second threshold is based on the 279 on-road data of Blanco et al. [1], in which the longest and most complicated in-car tasks lasted for 30 280 seconds on average, and the naturalistic driving data of Christoph and van Nes [3], in which the average 281 duration of manual interactions with a mobile phone was 31.0 seconds. The most frequently occurring 282 FrontApp tag in the cluster is considered to be the main application used in the cluster. Especially on 283 urban roads, drivers may perform part of the task while stationary or braking to stationary. In this analysis, 284 we have included only instances, which start when the car is moving with a speed of at least 2 m/s during 285 18 at least one touch. The data are normalized to uses per hour by the same method that was used to derive 286 touches per hour (Figure 8). 287 The data are again non-Gaussian, with a Shapiro-Wilkes test for the highway data giving W=0.92, 288 implying non-normality with p=0.038. Non-parametric tests thus need to be performed again. The median 289 is 4.1 and 15-85 percentile limits are 2.2 and 9.5. A typical driver in our data set thus uses some 290 application every fifteen minutes or so, while the heaviest users are using their phones almost every five 291 minutes. There were clear differences between road types (Figure 9, Tables 8 and 9), application usage per 292 hour being significantly more frequent on urban roads compared to the other road types (medium to large 293 effect). 294 295 Fig. 8. Application use instances per hour, N=30. 296 297 19 298 Fig. 9. Application use instances per hour by road type. 299 300 Table 8 Application use instances per hour on the different road types 301 302 Road type Application use instances per hour, median 15th percentile 85th percentile Highway 7.0 3.4 13.7 Main 7.2 3.8 13.5 Local 8.8 4.4 16.2 Urban 11.5 8.4 19.0 303 20 Table 9 Results of Wilcoxon rank tests for application use instances per hour by road type 304 305 Road type Main Local Urban Highway p=0.99 p=0.21 p<0.001* (d=0.90) Main p=0.19 p<0.001* (d=0.93) Local p=0.045* (d=0.48) *Significant at p<0.05. Cohen’s d is calculated for the statistically significant cases. 306 The duration of each application use instance can also be estimated as the time between the first and 307 the last touches in a cluster. There are statistically significant differences between the road types (Figure 308 10 and Table 10). According to the Shapiro-Wilks test, the data are too skewed to make an ANOVA 309 comparison. However, a Wilcoxon rank test (Table 11) shows that task durations are significantly longer 310 on highways, main, or local roads than on urban roads. The effect size is medium to strong (Cohen’s d up 311 to 0.84). Application use instance durations tended to be shorter on urban roads than on the other road 312 types. 313 314 Fig. 10. Duration of application use instances by road type (s). 315 21 Table 10 Duration of application use instances on the different road types 316 317 Road type Median duration (s) 15th percentile 85th percentile Highway 28.4 16.7 56.2 Main 25.7 13.9 42.8 Local 22.7 15.4 34.1 Urban 15.6 10.0 23.6 318 Table 11. Results of Wilcoxon rank tests for mean duration of application use instances by road type 319 320 Road type Main Local Urban Highway p=0.39 p=0.17 p<0.001* (d=0.84) Main p=0.59 p=0.002* (d=0.68) Local p=0.002* (d=0.46) *Significant at p<0.05. Cohen’s d is calculated for the statistically significant cases. 321 322 3.6 Frequently Used Applications 323 The application analysis is complicated somewhat by the fact that on different Android phone models, the 324 FrontApp listed may be different for the same application, or in some cases the same for different 325 applications. For example, Spotify use can be marked by at least three different FrontApp values. 326 The actual number of application use instances per application per user is also relatively small. Thus, 327 it is not possible to make reliable comparisons between different drivers or road types. However, accurate 328 aggregate statistics can be collected, that is, combined statistics from all users over the entire control 329 period (see Table 12). 330 The most used applications were also used at all speeds. PokémonGo was an exception; it was 331 played mostly at lower speeds, as could be expected from the game mechanics. The median speed during 332 an instance of PokémonGo usage is just 21 kmh; that is, it was very often played at crawling speeds. 333 However, there are also some instances of use at higher speeds (19 above 60 kmh). By contrast, for 334 22 instance, WhatsApp was used almost evenly at all speeds (median 57 kmh). All of the frequently used 335 downloadable applications are also among the 100 most downloaded applications in Google Play in 336 Finland, which suggests the applications drivers use in the car are the same applications they use in 337 general. However, naturally the frequency of use for driving-related applications, such as navigation 338 applications, could be higher than outside the car. 339 340 23 Table 12 Statistics for the most frequently used applications 341 342 Application N instances Touches/instance median (15%-85%) Duration (s) median (15%- 85%) Speed kmh median (15%- 85%) N drivers using Time distracted (h) Contacts! 818 3 (1-15) 17 (4-57)! 56 (20-94)! 27 7.5 Whatsapp 622 8 (1-76 35 (9-112)! 57 (16-95) 23! 10.1 Music 366 4 (1-17-) 17 (3-48)! 68 (25-92) 12! 2.9 Maps 277 3 (1-13) 11 (1-42)! 60 (18-96)! 19 2.2 Facebook 215 7(2-39) 31 (10-96)! 68 (21-99) 15 3.5 PokémonGo 175 26 (5-152)! 141 (30-637)! 21 (10-48) 4! 13.9 Search 169 9 (3-37) 34 (12-83)! 52 (10-94) 11! 2.5 Dialer 118 3 (1-15) 19 (4-65)! 53 (15-99) 12! 1.0 Browser 104 8 (2-29) 37 (5-93) 67 (27-94) 15 1.6 MMS 92 5 (1-36) 28 (6-97) 49 (18-92)! 12 1.5 YouTube 63 8 (1-32) 24 (1-78)! 82 (29-89)! 3 0.7 Email 59 8 (2-35) 45 (14-104)! 74 (24-109) 14 1.2 Banking 36 6 (2-19) 28 (6-64)! 56 (19-93)! 12 0.5 Email 59 8 (2-35) 45 (13-130)! 74 (24-109) 14 1.2 Netflix 33 3 (1-10)! 10 (1-41) 83 (59-86) 3! 0.2 Calendar 32 8 (2-14) 37 (18-134) 61 (8-84)! 12 0.6 Instagram 24 8 (2-72) 34 (7-128) 73 (31-100) 6 0.4 Camera 20 5 (1-34) 33 (3-65) 50 (18-91) 12! 0.3 Snapchat 19 8 (3-33) 28 (7-69)! 41 (20-59) 2! 0.2 Gallery 18 4 (2-39) 31 (6-148) 40 (14-68) 6! 0.3 Fonecta* 20 9 2-22) 33 (9-101) 54 (15-83) 6 0.3 News 9 5 (2-10) 16 (1-40) 57 (28121) 3! 0.1 Outlook 8 5 (3-22) 35 (17-102) 86 (64-98) 2! 0.2 Twitter 7 15 (6-76) 75 (25-320)! 79 (33-98) 2 0.4 Tinder 4 12 (5-26) 32 (25-78) 45 (41-49) 2 0.1 *Fonecta is a commonly used Finnish directory service. 343 24 The ratio of time spent distracted can − in principle − be estimated by adding the durations of the 344 application use instances and dividing by the time spent driving (Tables 12 and 13, see Figure 11 for 345 percentages per road type). Note that this estimate may be somewhat distorted, since the task duration may 346 include some cases in which the car has been stopped during part of the task execution. In any case, 347 Shapiro-Wilks tests show no significant differences between the road types in the ratio of time spent 348 distracted. 349 Table 13 Approximate ratio of time spent distracted on the different road types 350 351 Road type Ratio median 15th percentile 85th percentile Highway 0.063 0.027 0.17 Main 0.052 0.018 0.11 Local 0.053 0.018 0.15 Urban 0.062 0.026 0.091 352 353 Fig. 11. Percentage of time spent distracted by application use by road type, N=30. 354 31 One instance of WhatsApp also required a large number of touches (median 8). By contrast, Contacts 474 required and another commonly used application, Maps, required only a median of 3 touches per use 475 instance. This suggests that not all applications are equally risky in terms of distraction. A surprising find 476 was that PokémonGo was one of the most heavily used applications in traffic as measured by the total 477 number of touches (a small minority of drivers using it very heavily). 478 The findings indicate, unexpectedly, that Finnish drivers who use their smartphones while on the 479 move use their phones more frequently while driving in urban conditions. The usage pattern may however 480 be slightly different, with application uses being performed more rapidly in the urban conditions. 481 Acknowledgments 482 We thank Ficonic Solutions Ltd of Jyväskylä, Finland, for designing the software used in this experiment 483 and Paula Paananen for the help with the analysis of the photo material. This work was supported by the 484 VisGuard project, funded by TEKES (Finnish Funding Agency for Technology and Innovation; Grant 485 TEKES Dnro 813/31/2015) and the University of Jyväskylä. 486 References 487 488 [1] Blanco, M., Biever, W. J., Gallagher, J. P., & Dingus, T. A. (2006). The impact of secondary task cognitive 489 processing demand on driving performance. Accident Analysis & Prevention, 38, 895-906. 490 491 [2] Burnett, G.E. & Porter, J.M. (2001). Ubiquitous computing within cars: designing controls for non-visual use. 492 International Journal of Human Computer Studies, 55, 521-531. 493 494 [3] Christoph, M. & van Nes, N. (2013). When do drivers use their mobile phone?An analysis on the context of 495 mobile phone use based on naturalistic driving data. In Proceedings of the 3rd International Conference on Driver 496 Distraction and Inattention (oral presentation 81-O). 497 498 [4] Hancox, G., Richardson, J., & Morris, A. (2013). Drivers' willingness to engage with their mobile phone: The 499 influence of phone function and road demand. IET Intelligent Transport Systems, 7(2), 215-222. 500 501 [5] Horrey, W.J., & Lesch, M.F. (2009). Driver-initiated distractions: Examining strategic adaptation for in-vehicle 502 task initiation. Acc. Anal. Prev. 41, 115-122. 503 504 [6] Huth, V., Tattegrain, H., Sanchez, Y., & Brusque, C. (2013). Deciding to be distracted: Drivers’ strategic 505 choices to interact with the mobile phone. In Proceedings of the 3rd International Conference on Driver Distraction 506 and Inattention (oral presentation 41-O). 507 508 32 [7] Klauer, S.G., Dingus, T.A., Neale, V.L., et al. (2006). The Impact of Driver Inattention On Near-Crash/Crash 509 Risk: An Analysis Using the 100-Car Naturalistic Driving Study Data. Washington, DC: National Highway Traffic 510 Safety Administration. 511 512 [8] Kujala, T., Mäkelä, J., Kotilainen, I., & Tokkonen, T. (2016). The attentional demand of automobile driving 513 revisited –Occlusion distance as a function of task-relevant event density in realistic driving scenarios. Human 514 Factors, 58, 163-180. 515 516 [9] Kujala, T., Karvonen, H., & Mäkelä, J. (2016). Context-sensitive distraction warnings – Effects on drivers׳ 517 visual behavior and acceptance. International Journal of Human-Computer Studies, 90, 39-52. 518 519 [10] Kujala, T. & Salvucci, D.D. (2015). Modeling visual search on in-car displays: The challenge of modeling 520 safety-critical lapses of control. International Journal of Human-Computer Studies, 79, 66-78. 521 522 [11] Lee, J.D., Roberts, S.C., Hoffman, J.D., & Angell, L.S. (2012). Scrolling and driving: how an MP3 player and 523 its aftermarket controller affect driving performance and visual behavior. Human Factors 54, 250-263. 524 525 [12] Metz, B., Schömig, N., & Krüger, H. P. (2011). Attention during visual secondary tasks in driving: Adaptation 526 to the demands of the driving task. Transportation Research Part F: Traffic Psychology and Behaviour, 14(5), 369527 380. 528 529 [13] Metz, B., Landau, A., Hargutt, V., & Neukum, A. (2013). Exposure to secondary tasks in germany: Results 530 from naturalistic driving data. In Proceedings of the 3rd International Conference on Driver Distraction and 531 Inattention (oral presentation 35-O). 532 533 [14] Salvucci, D.D. & Kujala, T. (2016). Balancing structural and temporal constraints in multitasking contexts. In 534 Proceedings of the 38th Annual Meeting of the Cognitive Science Society (CogSci 2016), 2465-2470. 535 536 [15] Victor, T., Dozza, M., Bärgman, J., et al. (2015). Analysis of Naturalistic Driving Study Data: Safer Glances, 537 Driver Inattention, and Crash Risk (No.S2-S08A-RW-1). Washington, DC: Transportation Research Board. 538 539 [16] Wierwille, W.W. (1993). An initial model of visual sampling of in-car displays and controls. In: Gale, A.G., 540 Brown, I.D., Haslegrave, C.M., Kruysse, H.W., Taylor, S.P. (Eds.), Vision in Vehicles IV (pp. 271–279). 541 Amsterdam, NL: Elsevier Science. 542 543 [17] Wikman, A. S., Nieminen, T., & Summala, H. (1998). Driving experience and time-sharing during in-car tasks 544 on roads of different width. Ergonomics, 41(3), 358-372. 545 546 [18] Young, K.L. & Salmon, P.M. (2012). Examining the relationship between driver distraction and driving errors: 547 a discussion of theory, studies and methods. Saf. Sci. 50, 165–174. 548 549 [19] Özkan, T., Lajunen, T., Chliaoutakis, et al. (2006). Cross-cultural differences in driving behaviours: A 550 comparison of six countries. Transportation Research Part F: Traffic Psychology and Behaviour, 9(3), 227-242. 551