Are drivers aware of sleepiness and increasing crash risk while driving?

Page 1

Aw

Aa

b

a

ARRA

KFSDCR

Fni2(suoArwtphc

h0

Accident Analysis and Prevention 70 (2014) 225–234

Contents lists available at ScienceDirect

Accident Analysis and Prevention

jo u r n al homepage: www.elsev ier .com/ locate /aap

re drivers aware of sleepiness and increasing crash riskhile driving?

nn Williamsona,∗, Rena Friswell a, Jake Olivierb, Raphael Grzebietaa

Transport and Road Safety (TARS) Research, School of Aviation, University of New South Wales, Sydney, NSW, AustraliaSchool of Mathemetics and Statistics, University of New South Wales, Sydney, NSW, Australia

r t i c l e i n f o

rticle history:eceived 6 November 2013eceived in revised form 10 April 2014ccepted 10 April 2014

eywords:atigueleepinessrivingrash riskoad safety

a b s t r a c t

Drivers are advised to take breaks when they feel too tired to drive, but there is question over whether theyare able to detect increasing fatigue and sleepiness sufficiently to decide when to take a break. The aim ofthis study was to investigate the extent to which drivers have access to cognitive information about theircurrent state of sleepiness, likelihood of falling asleep, and the implications for driving performance andthe likelihood of crashing. Ninety drivers were recruited to do a 2 h drive in a driving simulator. They weredivided into three groups: one made ratings of their sleepiness, likelihood of falling asleep and likelihoodof crashing over the next few minutes at prompts occurring at 200 s intervals throughout the drive, thesecond rated sleepiness and likelihood of falling asleep at prompts but pressed a button on the steeringwheel at any time if they felt they were near to crashing and the third made no ratings and only useda button-press if they felt a crash was likely. Fatigue and sleepiness was encouraged by monotonousdriving conditions, an imposed shorter than usual sleep on the night before and by afternoon testing.Drivers who reported that they were possibly, likely or very likely to fall asleep in the next few minutes,were more than four times more likely to crash subsequently. Those who rated themselves as sleepy or

likely to fall asleep had a more than 9-fold increase in the hazards of a centerline crossing compared tothose who rated themselves as alert. The research shows clearly that drivers can detect changes in theirlevels of sleepiness sufficiently to make a safe decision to stop driving due to sleepiness. Therefore, roadsafety policy needs to move from reminding drivers of the signs of sleepiness and focus on encouragingdrivers to respond to obvious indicators of fatigue and sleepiness and consequent increased crash risk.

It is well-known that fatigue affects our ability to perform.atigue is an acknowledged road safety hazard of a similar mag-itude to alcohol while driving (Transport for NSW, 2011) and is

nvolved in around 19% of fatal crashes in NSW (Transport for NSW,011) and 31% of fatal crashes where three or more people are killedRoads and Traffic Authority (RTA), 2001). Other countries showimilar statistics including the UK where fatigue is attributed top to 20% of crashes (Jackson et al., 2011) and the USA with 16.5%f fatal crashes involving drowsy driving (American Automobilessociation Foundation for Traffic Safety, 2010). Surveys of driverseport that the experience of sleepiness while driving is common,ith more than half of French (57.3%) and US (64%) drivers ques-

ioned reporting drowsiness or sleepiness at the wheel over the

ast 12 months (Philip et al., 2010; Swanson et al., 2012) whichas been associated with higher risk of self-reported sleep-relatedrashes (Connor et al., 2002; Sagaspe et al., 2010).

∗ Corresponding author. Tel.: +61 2 93854599; fax: +61 2 93856040.E-mail address: [email protected] (A. Williamson).

ttp://dx.doi.org/10.1016/j.aap.2014.04.007001-4575/© 2014 Elsevier Ltd. All rights reserved.

© 2014 Elsevier Ltd. All rights reserved.

Fatigue and sleepiness are related states that in many studiesof driving performance are not differentiated; the term fatigueoften encompasses sleepiness. In fact, the term fatigue is oftenused to describe an overarching category that includes sleepinessand other mental fatigue phenomena such as task-related fatigue,fatigue resulting from illness, etc. There is considerable debate overthe definition and conceptualization of fatigue (Noy et al., 2011)and the extent to which it should be distinguished from sleepiness(Balkin and Wesensten, 2011). Part of the problem is that althoughthe antecedents of fatigue and sleepiness may be argued to be dif-ferent, their effects on subjective feelings of loss of alertness andtiredness and on performance are similar. The causes of sleepinessuniquely relate to sleep (i.e., amount, quality, time since awakeningand time of day effects) whereas the causes of fatigue can relate totask-related factors (i.e., duration and workload) as well as sleep-related factors. In this paper, both terms are used but the primary

focus of the study is on understanding awareness of sleepiness.

Managing fatigue (including sleepiness) is not a simple matterfor road or workplace safety. Unlike other road or work safety prob-lems, there are no clear exposure limits and fatigue management

Page 2

2 lysis and Prevention 70 (2014) 225–234

asolmhfsta

atfw(Cstpthscwtfaai

certsadEwswtesn22e

beitetatapstb

bt

Table 1Details of the characteristics of the three conditions assessed in this study.

Condition Characteristics

Condition 1: buttonpress only

Unprompted button press for crash likelihood,no subjective ratings

Condition 2: threeratings only

Prompted subjective ratings for Karolinskasleepiness scale (KSS), likelihood of fallingasleep, likelihood of crashing

Condition 3: tworatings, button press

Prompted subjective ratings for Karolinskasleepiness scale (KSS) and likelihood of falling

26 A. Williamson et al. / Accident Ana

pproaches often take the form of guidance rather than prescribingpecific actions through regulation. Fatigue management strategiesn the road and in the workplace involve driving or working hoursimits and advice that people take breaks when they feel tired. A

ajor assumption inherent in this advisory approach is that driversave access to cognitive information about their personal levels of

atigue and drowsiness that allow them to make the decision totop and rest before their performance is too adversely affected thatheir safety becomes compromised. There is considerable debatebout the validity of this assumption.

Research suggests that people can detect decreasing alertnessnd increasing fatigue and sleepiness. Many studies have shownhe expected decreases in alertness and increases in self-ratedatigue and sleepiness when sleep deprived (Dinges et al., 1997),hen required to work at vulnerable times in the circadian rhythm

Monk, 1991) or for prolonged periods without a break (Rosa andolligan, 1988). One study (Nilsson et al., 1997), showed that underimulation conditions drivers can make a judgment about whenhey should stop driving due to fatigue, apparently based on theirhysical symptoms. Interestingly, driver ratings of fatigue at theime they stopped were very similar, no matter whether the drivead been for only 40 min, or for as long as 180 min. This studyuggests that drivers can detect changes in fatigue but it is notlear when performance effects begin for a fatigued person andhether fatigued people have the capacity to detect the effect of

hese changes in state on performance. These are critical questionsor safety. It may not be enough simply to be aware of changes inlertness or feelings of fatigue and sleepiness. Drivers need to beble to detect and, importantly, to respond to changes which havempact on their capacity to drive safely.

Unfortunately, the evidence on the relationship betweenhanges in alertness and sleepiness-related states and performanceffects is equivocal. There is some evidence that increasing self-eported sleepiness is related to poorer performance in drivingasks. For example, Reyner and Horne (1998) showed in a drivingimulator that increasing subjective sleepiness was significantlyssociated with an increase in the number of safety-related inci-ents. Horne and Baulk (2004) also found that subjective sleepiness,EG-recorded sleepiness and lane deviations in a driving simulatorere highly correlated. In contrast, some laboratory studies have

hown that self-rated alertness or fatigue is significantly correlatedith self-rated performance but that the correlation of changes in

hese attributes with changes in actual performance is only mod-rate at best (Dorrian et al., 2000, 2003). In addition, some on-roadtudies found no association between self-assessed fatigue and aumber of non-driving performance measures (Williamson et al.,000) or a set of driving-related performance measures (Belz et al.,004). Further research is needed to clarify when performanceffects begin to occur and become noticeable for a fatigued person.

A number of studies have highlighted the differentiationetween detecting fatigue and sleepiness and deciding when thesexperiences might lead to falling asleep and potentially to crash-ng. Horne and Reyner (1999) found that drivers underestimatedhe probability of falling asleep when sleepy and seemed to under-stimate their likelihood of crashing. There is also recent evidencehat even partially sleep-deprived people who are sitting quietly in

darkened room doing a task requiring them to predict how closehey are to falling asleep have limited ability to predict when theyre going to first fall asleep (Kaplan et al., 2008). In fact it seems thateople may not be able to tell when they are in the early stages ofleep. There is evidence that people overestimate the time they takeo fall asleep and they can be in the early stages of sleep without

eing aware of it Baker et al. (1999).

It seems that drivers can detect that they are increasinglyecoming fatigued or drowsy, but may be less able to respond tohese sensations at the appropriate time by discontinuing what

asleep, unprompted button press for crashlikelihood

they are doing and taking a rest break. Current solutions to theproblem of managing driver fatigue for road safety rely on ill-defined relations between drivers’ judgments of their subjectivestate and their behavioural capacity. If we are to make an impacton driver fatigue, we need to know whether the current advi-sory approach can be successful. Unlike the issue of drink drivingwhich can be addressed at least in part by proscribing drivers’alcohol consumption and for which there is a clearly defined dose-response relationship between alcohol use and performance effects(Holloway, 1995), for fatigue, the problem is not so clear cut. Whilewe know a considerable amount about what causes fatigue and canmake some predictions about when it might occur, our predictionsare not perfect. This means that it is not sufficient to simply telldrivers that they must not drive during vulnerable periods such asthe middle of the night, or when they have not had enough sleepor have been driving for too long. Such prescriptions need moreresearch before they could be implemented as limits that will havethe desired effect of keeping tired drivers off the road and allowingalert drivers to drive. In the meantime, drivers are advised to takea break from driving when tired and to sleep or nap before fatigueand sleepiness begins to affect their driving skills. This approachassumes that drivers have access to cognitive information abouttheir levels of fatigue and sleepiness and are able to make thedecision to stop and rest before their performance is sufficientlyadversely affected that their risk of crashing becomes too high.As discussed above the validity of this assumption is extremelyquestionable.

Clearly, a study is needed that explores the relationshipsbetween driver awareness of fatigue and sleepiness and theperceived risk of crashing and their likelihood of actually crash-ing. This was the aim of the current study: to investigate the extentto which we have access to cognitive information about our currentfatigue state and levels of sleepiness, and the implications of hav-ing access for detection of changes in driving performance and thelikelihood of crashes. This study looked at the relationship betweendriver ratings of sleepiness, likelihood of falling asleep and likeli-hood of crashing measured at intervals and at driving performancethroughout a 2 h simulator drive. The study design extended earlierwork by Reyner and Horne (1998) by adding an additional conditionto investigate whether drivers can detect changes in crash likeli-hood as well as sleepiness state and sleep propensity and looked atthe relationship between these subjective ratings and driving sim-ulator performance. The study also examined whether the needto make subjective ratings across the drive influenced reportedexperiences of sleepiness and driving performance.

1. Methods

1.1. Study design

The study involved three conditions (see Table 1), each com-pleted by a separate group of 30 participants. Condition 1(Unprompted button press) was designed to determine whether

Page 3

lysis a

dltwmaaado2wl(tetptafwctt

a1(tip(t

espod

1

ahpwr

1

1

ildriprttiIi

A. Williamson et al. / Accident Ana

rivers are able to make unprompted judgments that they areikely to crash by pressing a button on the steering wheel whenhey felt they were likely to crash. Condition 2 (Prompted ratings)as designed to determine whether drivers are able to make judg-ents about their level of sleepiness, likelihood of falling asleep

nd the likelihood of crashing when prompted by an audible beepround every 200 s during the drive. Condition 3 (Prompted ratingsnd unprompted button press) was designed to determine whetherrivers were able to make ratings of sleepiness and the likelihoodf falling asleep when prompted by the audible beep around every00 s but they were not prompted about crash ratings. Rather, theyere asked to press the button if at any time they felt they were

ikely to crash. In this way, the effects of each of the manipulationsbutton press and prompted ratings) were counterbalanced. Condi-ion 2 looked at changes in prompted ratings on sleepiness-relatedxperiences and driving performance across the drive and coun-erbalanced condition 1 that involved no prompting, but buttonresses made if the driver felt he/she was close to crashing. Condi-ion 3 counterbalanced prompted ratings for sleepiness and fallingsleep but included unprompted button presses when the driverelt it appropriate. The aim of the prompt-no prompt comparisonsere to look at the effect of regular reminders of the need to assess

rash risk (prompting) which are needed to test the main aims ofhis study but that are less realistic than leaving drivers to makehe judgment with no reminding.

In the prompted conditions, driver ratings were requiredpproximately every 200 s intervals over a 2 h drive starting at4:30 h. Driver performance measures, reflecting safety outcomesrun-off-road crashes and centreline crossings) were taken duringhe drive and the timing of these measures and the driver rat-ngs were examined. Driver sleepiness was enhanced by askingarticipants to shorten their night sleep just before testing to 5 hactigraph validated) and by testing in the period coinciding withhe mid-afternoon circadian low-point.

In order to maximize the likelihood that participants wouldxperience feelings and effects of fatigue and sleepiness during thetudy, they were asked to reduce their sleep to 5 h over the nightrior to testing, testing occurred during the afternoon which is rec-gnized to be a vulnerable period for fatigue and sleepiness and therive involved 2 h of monotonous, rural driving.

.2. Participants

Participants were recruited by advertisements in newspapers,round the university and in the local community which soughtealthy, experienced drivers who did not have any diagnosed sleeproblems. Each condition involved separate groups of 30 driversho reported they were normal night sleepers. Participants were

andomly allocated to a study condition.

.3. Measures

.3.1. Driving simulatorA beta version of the STISIM Drive personal computer based

nteractive driving simulator was used. This is a low fidelity simu-ator involving a computer monitor, steering wheel and pedals. Therive for this study was 2 h long and designed to be a monotonous,ural simulation with few curves and minimal roadside scenery. Itnvolved variable speed limits to which drivers were asked to com-ly. The simulator was programmed to make an audible ‘beep’ ategular intervals of around every 200 s. Prompts were programmedo occur at particular locations (distances) rather than particular

imes. The track distances used were those corresponding to 200 sf the driver was travelling at the speed limits for the entire drive.n the prompted conditions (2 and 3) there were 35 rating promptsn the 2 h session. The simulator included a button on the steering

nd Prevention 70 (2014) 225–234 227

wheel that participants were asked to use in conditions 1 and 3 toindicate that they felt they were likely to crash.

1.3.2. Sleep validationActigraphs (MiniMitter Actiwatch 64) were used to validate the

overnight sleep restriction for all participants.

1.3.3. Sleepiness detectionThe 9-point Karolinska sleepiness scale, a validated (Akerstedt

and Gillberg, 1990) and widely used subjective measure of sleepi-ness was used. The scale is constructed such that 1 = extremelyalert, 2 = very alert, 3 = alert, 4 = rather alert, 5 = neither, 6 = somesigns of sleepiness, 7 = sleepy, but no effort to stay awake, 8 = sleepy,but some effort to stay awake, 9 = very sleepy, great effort to stayawake.

1.3.4. Likelihood of falling asleepParticipants in conditions 2 and 3 were asked to answer the

question ‘What is the likelihood of you falling asleep during the nextfew minutes’ immediately following the beep, with the responseoptions of A = very unlikely, B = unlikely, C = possibly, D = likely,E = very likely.

1.3.5. Likelihood of crashingParticipants in condition 2 were asked to answer the question

‘What is the likelihood of you crashing during the next few min-utes’ immediately following the beep, with the response options ofA = very unlikely, B = unlikely, C = possibly, D = likely, E = very likely.

The relevant rating scales were displayed just above the simula-tor monitor to allow participants to check the options, if necessary,without losing vision of the road. Ratings were made verbally andwere audio recorded via a small desktop microphone.

1.3.6. Driving performancemeasures were taken of lateral deviation in lane (centreline

excursions) and crashes where the vehicle left the road completelyto either the left or right.

1.4. Procedure

All participants were given a 30 min practice drive on a daypreceding the test day in order to reduce practice effects on per-formance and to enhance the monotony of driving during thetest session. Participants were then given actigraphs to be wornovernight and instructed in their use. They were asked to shortentheir sleep period to around 5 h by going to bed later than usual.In all other respects, participants were asked to maintain theirnormal practices with respect to alcohol, caffeine and medicationuse. On the practice day, participants completed a questionnaireon demographic characteristics, daily routine and current healthstatus. The questionnaire also included questions on height andweight from which body mass index (BMI = kg/m2) was calculated,the multivariable apnea probability index (Maislin et al., 1995;Gurubhagavatula et al., 2004) and the Epworth sleepiness scale(Johns, 1991, 1992) to assess sleep problems. Participants wereasked to keep a brief diary of their activity in the evening beforethe test, including bed time and rising time.

On the test day, participants were asked to arrive at the lab-oratory at around 2.00 pm. They were asked to complete a shortquestionnaire regarding their sleep quality and activities betweenwaking and arriving at the laboratory, and returned their actigraphsand sleep diaries. They were reminded of their task and instructedto obey the road rules and to remain in the far left hand lane for the

entire drive. Participants were asked to give up all timing devices(e.g., watches, phones etc.). They were then asked to start the 2 htest drive at 2:30 pm. In conditions 2 and 3, rating prompts beganapproximately 5–6 min into the drive to allow time for participants

Page 4

2 lysis a

ti(

1

uwiimoaooow

tcocdrqm

R

pocblatctgtFp2Lidswmwlowtwfi

2

2

i

28 A. Williamson et al. / Accident Ana

o reacquaint themselves with the simulator controls. The task var-ed according to which condition they were assigned to: promptedconditions 2 and 3) or unprompted (condition 1).

.5. Statistical analysis

The data was analyzed in three stages: the first was a manip-lation check to evaluate the extent to which study participantsere similar across conditions and complied with the study design

n terms of activities and sleep in the 24 h before the test and sonvolved participants from all three conditions; the second was a

anipulation check to determine whether the need to make ratingsr not throughout the drive influenced crash-related outcomes andgain included all three conditions; and the third was an analysisf the relationships between driver ratings of sleepiness, likelihoodf falling asleep and of crashing and crash-related performanceutcomes and included only participants from conditions 2 and 3hich actually made ratings.

The analysis of the first stage compared demographic charac-eristics, daily routine and recent experiences between the threeonditions and involved ANOVA and chi-square tests dependingn the nature of the measure used. The second stage, comparinghanges in the three types of ratings over the drive for each con-ition involved a linear mixed model approach. For this analysis,atings data were fitted with models containing both linear anduadratic time terms to adequately capture the data. The followingodel was fitted for each type of rating for participant i:

atingi = ˇ0 + ˇ1Time + ˇ2Time2 + ˇ3Factor + ˇ4Time × Factor

+ ˇ5Time2 × Factor + ai + biTime

In these models, Time was measured as consecutive ratingrompt number (1–35) so that Time = 0 corresponds to the startf the driver simulation, Factor was a dichotomous variable indi-ating whether the participant was in condition 2 or 3 and ai andi were random intercept and slope terms respectively. In all theinear mixed models, tests of the Time and Time2 coefficients (ˇ1nd ˇ2, respectively) indicate whether the slope and curvature ofhe ratings over time are significant. The Factor coefficient (ˇ3) indi-ates whether the Factor groups differ initially (at time = 0). Tests ofhe interaction coefficients (ˇ4 and ˇ5) indicate whether the Factorroups have different slopes initially and whether they differ overime. At baseline, the ratings are identical between the levels ofactor when ˇ3 = 0 and ˇ4 > 0 indicates the ratings for condition 3articipants are increasing at a greater rate than those in condition. The inclusion/exclusion of random effects was determined usingikelihood ratio tests. The third stage involved a comparison of rat-ngs for participants who had safety-related outcomes during therive (crashed or crossed the centreline or not). This involved twoeparate analyses. The first looked at differences between thoseith and without each type of safety outcome using the linearixed model approach described above with Factor now indicatinghether the participant crashed (1) or not (0), or crossed the centre-

ine (1) or not (0), depending on the analysis. The second analysisf safety-related outcomes looked at whether participant ratingsas associated with the hazards of crashing or centreline crossings

hrough Survival analysis using Cox proportional hazards models inhich the three types of ratings were modelled against time to therst crash or to the first centre-line crossing (near crash) outcomes.

. Results

.1. Study condition and routine leading up to test session

The final study sample included 85 participants. No partic-pants were excluded due to chronic sleep problems. Loss of

nd Prevention 70 (2014) 225–234

participants from the study sample was even across conditionsand was mainly due to four participants having considerably moresleep than intended (>6 h) and one who had a very unusual sleeppattern due to illness. Participants in the three conditions werevery similar. There was no significant age difference betweenthe groups completing the three experimental conditions (over-all mean age = 45.8 yr, univariate ANOVA F(2,82) = 0.32, p = 0.73).Just over half of the participants (56.5%) were female and therewere no significant differences in gender distribution between thethree conditions (X2

(2) = 0.74, p = 0.69). Regardless of experimen-tal condition, most drivers had held a driver’s license for morethan 10 years (83.5%). Almost all the participants usually droveat least once a week (90.6%), averaging around 140 km in a typi-cal week (median = 100 km) with no difference between conditions(F(2,74) = 1.06, p = 0.35).

BMI was categorized according to World Health Organization(1995) guidelines into underweight (<18.5), normal [18.5, 25),overweight [25, 30) and obese (30+). Underweight and normal cat-egories were combined for analysis because there was only oneunderweight participant. The three experimental conditions didnot differ on the distribution of BMI (X2

(4) = 1.06, p = 0.90), whichwas similar to the Australian adult population overall (AustralianBureau of Statistics, 2010). They also did not differ on their nor-mal sleep patterns, when sleep occurred nor the amount obtained(mean bedtime = 11 pm, mean wake time = 6:45 am, mean dura-tion = 7:20 h of sleep; Fall asleep F(2,82) = 0.04, p = 0.96; Wake upF(2,82) = 0.54, p = 0.59: Hours sleep F(2,82) = 0.42, p = 0.66). MeanEpworth sleepiness scale scores were within the normal range (10or less) with 88.2% of participants scoring in this range and therewere no significant differences between conditions (F(2,82) = 2.56,p = 0.08). Multivariable Apnea Probability index was typically lowand there was no evidence of group differences (F(2,76) = 0.81,p = 0.45).

On the day of testing, most participants consumed a caffeine-containing drink (71.1%), just under one-third (30.1%) reportedtaking their usual medication and 16.9% reported drinking alcoholin the 24 h before the start of the test session, however there wereno differences between the three conditions (caffeine, medicationsand alcohol respectively, X2

(2) = 0.26, p = 0.88, X2(2) = 2.56, p = 0.28,

X2(2) = 0.21, p = 0.90). Of the participants who consumed caffeine

drinks (89.1%) on the test day the amount was one or two drinks,and it was last consumed on average around 4 h before the start ofthe test session. Those taking medications did so around breakfasttime on the morning of the test day (approximately 6–7 h beforetesting). Participants who drank alcohol in the 24 h before the testsession, drank no more than one or two drinks on the evening beforethe test day, a minimum of 14 h before testing.

Actigraph validated sleep showed that on average, participantsfell asleep around 1:45 am and woke around 6:30 am on the testday with around 4:30 h sleep on average and around 8 h betweenwaking and commencement of testing. Most participants (91.8%)had actiwatch data recorded for their sleep on the night before thetest session. There were no significant differences between condi-tions on any diary sleep variables (p > 0.05 in each case), includingrated quality and feeling refreshed on waking nor on any actigraphmeasures.

2.2. Patterns of ratings and crash-related outcomes across thedrive

The majority of participants reported some degree of sleepinessduring the test drive (85.1% scored at least 6 (some signs of sleepi-

ness) on the Karolinska scale), around half reported it being at leastlikely that they would fall asleep (56.1% scored 4 or 5 on the Sleeplikelihood scale) and 43.5% reported that it was at least likely theymight crash (scored 4 or 5 on the Crash likelihood scale). The two

Page 5

A. Williamson et al. / Accident Analysis and Prevention 70 (2014) 225–234 229

Table 2Results of the linear fixed effects models for two rating scales, Karolinska sleepinessscale and likelihood of falling asleep by time in the drive and study condition.

Source ̌ SE p Value

Karolinska sleepiness scaleIntercept 1.63 0.36 0.0001Time 0.37 0.02 0.0001Time2 −0.006 0.0003 0.0001Condition 0.93 0.57 0.11Time × condition −0.09 0.03 0.001Time2 × condition 0.001 0.0005 0.01

Likelihood of falling asleepIntercept 1.01 0.15 0.0001Time 0.15 0.01 0.0001Time2 −0.003 0.0002 0.0001Condition 0.44 0.25 0.09

cwaufpwtwo(rbcotsso

iatc(pcwshat

otftdis

fhwsrhp

Table 3Results of the linear fixed effects models for each of the three rating scales, KarolinskaSleepiness scale, likelihood of falling asleep and likelihood of crashing, by time inthe drive and whether or not a crash occurred.

Source ̌ SE p Value

Karolinska sleepiness scaleIntercept 1.88 0.32 0.0001Time 0.28 0.02 0.0001Time2 −0.004 0.002 0.0001Crash 0.68 0.58 0.25Time × crash 0.16 0.03 0.0001Time2 × crash −0.005 0.0005 0.0001

Likelihood of falling asleepIntercept 1.04 0.12 0.0001Time 0.12 0.009 0.0001Time2 −0.002 0.0001 0.0001Crash 0.62 0.28 0.03Time × crash 0.06 0.02 0.0003Time2 × crash −0.002 0.0003 0.0001

Likelihood of crashingIntercept 0.94 0.18 0.0001Time 0.11 0.013 0.0001Time2 −0.001 0.0002 0.0001Crash 0.17 0.25 0.51

slope, Likelihood ratio = 850.77, p < 0.0001, intercept, Likelihoodratio = 1251.2, p < 0.0001; Likelihood of falling asleep, slope, Likeli-hood ratio = 787.44, p < 0.0001, intercept, Likelihood ratio = 759.54,p < 0.0001; Likelihood of crashing, slope, Likelihood ratio = 327.04,

Table 4Results of the linear fixed effects models for each of the three rating scales, Karolinskasleepiness scale, likelihood of falling asleep and likelihood of crashing, by time inthe drive and whether or not a centreline crossing occurred.

Source ̌ SE p Value

Karolinska sleepiness scaleIntercept 2.00 0.3 0.0001Time 0.28 0.17 0.0001Time2 −0.005 0.0003 0.0001Centreline 0.25 0.63 0.70Time × centreline 0.11 0.03 0.0001Time2 × centreline −0.003 0.0005 0.0001

Likelihood of falling asleepIntercept 1.03 0.12 0.0001Time 0.13 0.009 0.0001Time2 −0.003 0.0002 0.0001Centreline 0.52 0.28 0.07Time × centreline 0.02 0.02 0.15Time2 × centreline −0.0003 0.00003 0.25

Likelihood of crashingIntercept 0.82 0.16 0.0001Time 0.14 0.01 0.0001

2

Time × condition −0.01 0.02 0.44Time2 × condition 0.000007 0.0003 0.98

onditions in which ratings occurred (condition 2 and condition 3)ere very similar in ratings of sleepiness and likelihood of falling

sleep (see Table 2). Comparison of sleepiness ratings by conditionsing linear mixed modelling showed significant random effectsor both intercepts and slopes (slope, Likelihood ratio = 785.09,

< 0.0001; intercept, Likelihood ratio = 1243.5, p < 0.0001) and soere included in the final model. Fixed effects were significant for

he time variables but not for condition, but both interaction effectsere statistically significant. The same modelling of likelihood

f falling asleep ratings also showed significant random effectsslope, Likelihood ratio = 788.81, p < 0.0001; intercept, Likelihoodatio = 785.97, p < 0.0001) and fixed effects for the time variables,ut neither the condition effect nor the interactions were statisti-ally significant. These findings show that condition had no effectn the likelihood of falling asleep but for sleepiness ratings condi-ion and time interacted. Examination of the predicted curves forleepiness ratings over time shows that the curve for condition 2 islightly flattened and ratings are somewhat lower in the last thirdf the drive compared to condition 3.

Just over one-third of participants (34.5%) had at least one crashnvolving the vehicle leaving the road to either the left or right andpproaching half (43.5%) had at least one centreline crossing duringhe test drive. There was no significant difference between the threeonditions, however, in the number of participants who crashedX2

(2) = 1.37, p = 0.50) or had centre line crossings (X2(2) = 2.12,

= 0.35). There was a significant relationship between centrelinerossings and crashing as participants who crossed the centrelineere also likely to crash (X2

(1) = 28.79, p < 0.001). There were notatistically significant differences on any demographic or recentistory variables for participants who crashed during the test drivend those who did not or for those who crossed the centreline andhose who did not.

The analysis of the relationships between ratings and safetyutcomes combined data from conditions 2 and 3 as the condi-ions were not different on most measures. The only difference wasound for sleepiness ratings in the form of the significant interac-ion between condition and time. The average magnitude of thisifference was only 0.17 sleepiness scale units, so was very small

n practice, and the overall pattern of sleepiness ratings was veryimilar between conditions.

For crashes, separate linear mixed modelling was conductedor ratings of sleepiness, likelihood of falling asleep and likeli-ood of crashing (see Table 3 and Fig. 1) Significant random effectsere found in each case for both intercepts and slopes (Sleepiness:

lope, Likelihood ratio = 883.98, p < 0.0001, intercept, Likelihood

atio = 1160.13, p < 0.0001; Likelihood of falling asleep, slope, Likeli-ood ratio = 813.87, p < 0.0001, intercept, Likelihood ratio = 665.32,

< 0.0001; Likelihood of crashing, slope, Likelihood ratio = 343.66,

Time × crash 0.10 0.02 0.0001Time2 × crash −0.002 0.0004 0.0001

p < 0.0001, intercept, Likelihood ratio = 239.15, p < 0.0001) and sowere all included in the final model. For each type of rating, fixedeffects were significant for the time variables and for both timeby crash interactions, but the crash variable was only significantfor likelihood of falling asleep. These findings and examination ofpredicted curves show that ratings increased more rapidly andto higher levels in the first half of the drive for participants whocrashed compared to those who did not. This pattern was seen foreach type of rating, although ratings of Likelihood of falling asleepwere significantly higher for drivers who crashed across the drive.

For centreline crossings, similar linear mixed modelling wasconducted for each type of rating (see Table 4 and Fig. 2).The final models for each rating type included random effectsagain as intercepts and slopes were significant (Sleepiness:

Time −0.003 0.0003 0.0001Centreline 0.4 0.31 0.21Time × centreline −0.009 0.02 0.68Time2 × centreline 0.0007 0.0004 0.08

Page 6

230 A. Williamson et al. / Accident Analysis and Prevention 70 (2014) 225–234

0 5 10 15 20 25 30 35

24

68

Rating Prompts

Kar

olin

ska

Sca

le

No Crash

Crash

0 5 10 15 20 25 30 35

12

34

5

Rating Prompts

Like

lihoo

d of

Sle

ep

No Crash

Crash

0 5 10 15 20 25 30 35

12

34

5

Rating Prompts

Like

lihoo

d of

Cra

sh No Crash

Crash

F ood oc ults in

pecbs(tmp

boptrt

ig. 1. Changes in driver ratings on the Karolinska sleepiness scale (graph A), likelihrashed or did not crash on the 2 h drive showing raw and modeled (solid lines) res

< 0.0001, intercept, Likelihood ratio = 258.49, p < 0.0001). Fixedffects were significant for time for each type of rating, but not forentreline crossings. For sleepiness ratings the centreline crossingy time effects were significant, but neither of the other rating typeshowed significant interactions. As shown in the model curvesFig. 2), while ratings increased over time for all participants, forhose who had centreline crossings Sleepiness ratings increased

ore rapidly and were higher especially earlier in the drive thanarticipants who had no centreline crossings.

Survival analysis using Cox modelling assessed the relationshipetween each type of rating prior to the first crash and the time toccurrence of the crash. The same analysis was also used for ratings

rior to the first centreline crossing. Kaplan–Meier survival func-ion plots of the predictors of crashing are shown in Fig. 3 and theesults for the Cox regressions are shown in Table 5. Tests of each ofhese models showed that the first crash was significantly predicted

f falling asleep scale (graph B) and likelihood of crashing (graph C) for drivers who each case.

by the preceding rating of likelihood of falling asleep (X2(2) = 6.24,

p < 0.01) with hazard ratio statistics showing that rating the likeli-hood of falling asleep in the next couple of minutes as possible,likely or very likely increased the hazard of a crash 4.3 timescompared to participants rating falling asleep as unlikely or veryunlikely. Neither sleepiness ratings nor crash likelihood ratings sig-nificantly predicted the first crash (X2

(2) = 4.19, p = 0.12; X2(2) = 0.94,

p = 0.33, respectively). The results for predictors of the first centre-line crossing (see Table 5 and Fig. 4) showed significant predictorsof ratings of sleepiness (X2

(2) = 11.14, p < 0.004) and of falling asleep(X2

(2) = 13.46, p < 0.001), but not for ratings of likelihood of crashing(X2

(2) = 3.05, p = 0.08). Participants with high ratings of sleepiness

(sleepy requiring effort to stay awake) increased the hazard of acentreline crashing more than 10-fold compared to those ratingthemselves as alert to neutral. Participants who rated themselvesas more likely to fall asleep (possible, likely or very likely) had more

Page 7

A. Williamson et al. / Accident Analysis and Prevention 70 (2014) 225–234 231

0 5 10 15 20 25 30 35

24

68

Rating Prompts

Kar

olin

ska

Sca

le No Centreline Crossing

Centreline Crossing

0 5 10 15 20 25 30 35

12

34

5

Rating Prompts

Like

lihoo

d of

Sle

ep No Centreline Crossing

Centreline Crossing

0 5 10 15 20 25 30 35

12

34

5

Rating Prompts

Like

lihoo

d of

Cra

sh No Centreline Crossing

Centreline Crossing

F ood oc lines)

tw

3

midoamrfddoo

tfa

ig. 2. Changes in driver ratings on the Karolinska sleepiness scale (graph A), likelihrossed the centerline or not during the 2 h drive showing raw and modeled (solid

han a 9-fold increased risk of centreline crashes compared to thoseho rated falling asleep as unlikely or very unlikely.

. Discussion

This study demonstrated that under conditions in which driversight be expected to experience fatigue and sleepiness, driv-

ng in the afternoon after reduced prior sleep, the majority ofrivers reported increasing sleepiness levels, increasing likelihoodf falling asleep and of crashing. The study also demonstrated thatdverse safety outcomes (crashes and centreline crossings) wereore likely for drivers who reported high levels of sleepiness and

ated the likelihood of falling asleep and of crashing as higher. Inact, only sleep-related and crash likelihood ratings differentiatedrivers who crashed or had a centreline crossing from those whoid not indicating that these outcomes were unlikely to be due tother factors like age, gender, driving experience, amount of sleepr recent activities.

Most significantly for the aims of this study, the results showedhat drivers who rated the likelihood of falling asleep in the nextew minutes as positive (very likely, likely or even possible) crashedt more than four times the rate per unit time (i.e., a hazard ratio

f falling asleep scale (graph B) and likelihood of crashing (graph C) for drivers whoresults in each case.

of over 4) and crossed the centerline at more than ten times therate per unit time (i.e., a hazard ratio of more than 10) comparedto those with lower levels of sleepiness. Furthermore, drivers whorated themselves as more likely to fall asleep in the next few min-utes crossed the centerline more than nine times more frequently(hazard ratio over 9) than those who rated their risk of falling asleepas low. It seems that awareness of the likelihood of falling asleepis the best predictor of unsafe outcomes: centreline crossings andcrashing. Awareness of sleepiness significantly predicted centrelinecrossings, but not crashes although the same pattern of results wasseen in both outcomes. This apparent anomaly may be due simplyto fewer crashes occurring across the 2 h drive.

These results show that drivers are clearly aware of experiencesrelating to fatigue and sleepiness and are able to detect when theseexperiences might lead to falling asleep. The fact that drivers haveaccess to this cognitive information makes it possible for themto respond to increasing fatigue and sleepiness in order to avoidadverse road safety consequences of crashing. These results con-

firm that fatigue and sleepiness and even the risk of falling asleepdo not just take drivers unaware. Rather they show that drivers doknow that they are fatigued and sufficiently sleepy to possibly fallasleep while driving. Most importantly, this study demonstrated

Page 8

232 A. Williamson et al. / Accident Analysis and Prevention 70 (2014) 225–234

0 2000 4000 6000

0.0

0.2

0.4

0.6

0.8

1.0

Time (sec)

Kap

lan−

Mei

er E

stim

ate

Karolinska Score

1−56−78−9

0 2000 4000 6000

0.0

0.2

0.4

0.6

0.8

1.0

Time (sec)

Kap

lan−

Mei

er E

stim

ate

Sleep Likelihood

Very unlikely/unlikelyPossibly/Likely/Very Likely

0 2000 4000 6000

0.0

0.2

0.4

0.6

0.8

1.0

Time (sec)

Kap

lan−

Mei

er E

stim

ate

Crash Likelihood

Very unlikely/unlikelyPossibly/Likely/Very Likely

Fig. 3. Kaplan–Meier survival plots for sleepiness, likelihood of falling asleep and likelihood of crashing as predictors of crashes while driving.

0 2000 4000 6000

0.0

0.2

0.4

0.6

0.8

1.0

Time (sec)

Kap

lan−

Mei

er E

stim

ate

Karolinska Score

1−56−78−9

0 2000 4000 6000

0.0

0.2

0.4

0.6

0.8

1.0

Time (sec)

Kap

lan−

Mei

er E

stim

ate

Sleep Likelihood

Very unlikely/unlikelyPossibly/Likely/Very Likely

0 2000 4000 6000

0.0

0.2

0.4

0.6

0.8

1.0

Time (sec)

Kap

lan−

Mei

er E

stim

ate

Crash Likelihood

Very unlikely/unlikelyPossibly/Likely/Very Likely

Fig. 4. Kaplan–Meier survival plots for sleepiness, likelihood of falling asleep and likelihood of crashing as predictors of centreline crossings while driving.

Page 9

A. Williamson et al. / Accident Analysis a

Table 5Cox regression model of the relationship between time to first driving incident(offroad accident or centreline crossing) and participant ratings of sleepiness, like-lihood of falling asleep and likelihood of crashing.

Driving incident preceding ratingpredictor

Hazardratio

95.0% CI forhazard ratio

p

First offroad accidentKSS sleepiness

Not sleepy (1–5)* 1.00Sleepy (6–7) 0.42 0.08 2.27 0.31Sleepy/effort to stay awake (8–9) 1.71 0.54 5.39 0.36

Likelihood of falling asleepUnlikely/very unlikely (A and B)* 1.00Possibly/likely/very likely (C–E) 4.31 1.24 15.02 0.02

Likelihood of crashingUnlikely/very unlikely (A and B)* 1.00Possibly/likely/very likely (C–E) 2.00 0.48 8.41 0.34

First centreline crossingKSS sleepiness

Not sleepy (1–5)* 1.00Sleepy (6–7) 3.59 0.43 30.11 0.25Sleepy/effort to stay awake (8–9) 10.40 1.37 78.97 0.025

Likelihood of falling asleepUnlikely/very unlikely (A and B)* 1.00Possibly/likely/very likely (C–E) 9.43 2.19 40.64 0.003

Likelihood of crashingUnlikely/very unlikely (A and B)* 1.00Possibly/likely/very likely (C–E) 2.84 0.84 9.59 0.1

* Reference category: sleepiness rating = alert-neutral (1–5), likelihood of fallingau

ttoocltmRo

thn1mttcnftidtc

ornfaosaaf

reaction speed while driving in the simulator environment (Philip

sleep = very unlikely or unlikely (A and B), likelihood of crashing = very unlikely ornlikely (A and B).

hat not only do drivers who have adverse safety outcomes ratehemselves as higher on sleepiness, likelihood of falling asleep andf crashing, the Cox modelling results show that higher ratingsf likelihood of falling asleep precede crashes and that centrelinerossings are preceded by higher ratings of sleepiness and of like-ihood of falling asleep. This is the first study to demonstrate theemporal relationship between subjective experiences and perfor-

ance outcomes. The results support the contention of Horne andeyner (1999) who argued that no driver falls asleep at the wheelf a vehicle without experiencing fatigue and sleepiness first.

A unique aspect of this study was the requirement for driverso make ratings of the likelihood of crashing. Previous studiesave showed significant relationships between ratings of sleepi-ess and falling asleep and performance (e.g., Reyner and Horne,998; Horne and Burley, 2010), but did not look at subjective judg-ents about the likelihood of safety outcomes. This study was able

o show that drivers make judgments of crash likelihood differentlyo judgments about sleepiness and falling asleep. While drivers whorashed had higher ratings of crash likelihood than drivers who didot crash, consistent with the results for ratings of sleepiness and

alling asleep, higher ratings of crash likelihood did not precedehe first crash or centreline crossing. These results suggest thatncreasing awareness of sleepiness and the risk of falling asleepoes not transfer to awareness of the risk of crashing and point tohe need for interventions to focus on sleep likelihood rather thanrash likelihood.

The results of this study help to direct much-needed policyn fatigue risk management especially for driving. While furtheresearch is needed to accurately predict when fatigue and sleepi-ess and related performance impairment is most likely to occur

or all drivers, in the meantime, this study supports an advisorypproach based upon subjective sleepiness and assessed likelihoodf falling asleep but not assessed crash risk. This study demon-trated that drivers do indeed have access to cognitive information

bout their levels of fatigue, sleepiness and whether they might fallsleep that they need to avoid adverse effects on performance by,or example, stopping for a break, a nap or a caffeine-containing

nd Prevention 70 (2014) 225–234 233

drink. These results indicate, however, that drivers do not appearto translate this heightened awareness of sleepiness and even offalling asleep into judgments that their risk of crashing is also ele-vated. It seems that, as others have pointed out (Horne and Burley,2010), drivers may ignore or overlook feelings of sleepiness despitethe obvious risk of adverse safety consequences.

This study is significant as the findings clarify targets for actionto manage driver fatigue and sleepiness. The research indicates thatdrivers are able to help themselves to manage driver fatigue andthat the consequences of fatigue and sleepiness while driving arenot necessarily inevitable without secondary safety countermea-sures like lane departure or eye closure warnings. These resultsshow that drivers have access to cognitive information about theirlevel of sleepiness and the likelihood of falling asleep before crash-ing that should inform judgments about when to take breaks fromdriving and, in fact, when to drive at all. This means that decid-ing to drive, or continue to drive, when fatigued, sleepy or feelingthat sleep is likely is a road safety judgment equivalent to the deci-sion to exceed the speed limit or to drive after drinking alcohol.Drivers can choose to behave safely when fatigued or not. The find-ing that crash likelihood was downplayed even though drivers wereaware of sleepiness levels that could compromise safety also high-lights poor road safety judgment. Drivers can choose to continueto drive, despite clear warnings that their capacity to drive safelyis diminishing, or not.

Countermeasures for driver fatigue, therefore, need to encour-age and motivate drivers to avoid driving when fatigued and sleepy.This means that driver education is not useful if it only describesthe signs and symptoms of fatigue. This research shows that driversalready know this. Educational approaches need to reinforce theneed to respond appropriately to the early signs and symptoms offatigue and sleepiness on the basis that they are important indica-tors of increasing safety risk that should not be ignored. Motivationto make safe decisions about driving when fatigued and sleepy maybe enhanced with appropriate enforcement. For example, althoughthe lack of standard or reliable measures of fatigue and sleepinesshas traditionally made enforcement difficult, road safety authori-ties currently identify fatigue-related crashes, often single vehiclecrashes, and it would be possible to implement significantly higherpenalties for these types of crashes in order increase the salience ofthe message about the perils of failing to respond to signs and symp-toms of fatigue. This combination of education and enforcement isclearly needed to target safe decision-making around driving whilefatigued or sleepy.

This study focused on healthy drivers without self-reportedsleep problems, so the findings may not apply to drivers withchronic sleep problems (Smolensky et al., 2011). It might be pre-dicted, however, that since this study demonstrated effects with atemporary sleep deprivation in healthy drivers that effects mightbe more pronounced for drivers who experience sleep loss on amore frequent basis.

This study was conducted in a driving simulator in order tomaintain control over the driving conditions being tested. The find-ings may therefore not reflect real world conditions. The monotonyof the simulator drive may have enhanced the feelings of fatigueand sleepiness and the lack of consequences of crashing and centre-line crossing may have made them more likely. As the focus of thestudy was on the temporal relationships between these variablesand especially on the first crash or centreline crossing, greater num-bers of crashes are less of concern. Research comparing fatigue andsleepiness in a simulator or real driving showed similar patterns ofeffects but more pronounced effects on both sleepiness ratings and

et al., 2005). Nevertheless, this study should be repeated on-roadusing less dangerous outcome measures in order to confirm thefindings.

Page 10

2 lysis a

tldmomtiTotfihttbefrslasstltis

spbmtwcadm

A

cfMW

R

A

A

A

B

B

34 A. Williamson et al. / Accident Ana

Another potential limitation of the study design was the needo ask drivers to make ratings of their current sleepiness level andikelihood of falling asleep or crashing in order to find out howrivers were feeling during the drive. The act of making ratingsay have enhanced their alertness or made drivers more aware

f safer driving which may in turn have enhanced driver perfor-ance leading to lower numbers of crashes. It was possible to test

his hypothesis as the study included a condition requiring no rat-ngs and one in which only sleepiness-related ratings were made.he results showed no significant effects of making ratings or notn the number of crashes or centerline crossings. The only statis-ically significant difference between conditions was that in thenal third of the drive, drivers who made regular crash risk ratingsad slightly lower sleepiness ratings than those who only neededo press a button if they felt they were going to crash. It may behat making drivers think actively about crash risk when they areecoming most tired might make drivers less sensitive to high lev-ls of sleepiness. It is notable that judgments of the likelihood ofalling asleep were very similar between conditions that rated crashisk or not. Kaida et al. (2007) compared the effects of sleep deprivedtudy participants making verbal ratings on subjective and physio-ogically assessed sleepiness and performance on a vigilance task,nd found some evidence that the act of making ratings decreasedleepiness measured subjectively and physiologically but, as in thistudy, had no effect on performance. The differences between thewo studies in effects of ratings on sleepiness may be due to theength of the test period; with Kaida et al. (2007) study involvingwo test periods of 40 min separated by a break, whereas the driven the current study was 2 h in duration. The longer duration in thistudy may have increased sleepiness.

In conclusion, by showing that drivers are aware of increasingleepiness and the risk of falling asleep while driving, this researchrovides some much-needed evidence on which new policy cane developed for managing driver fatigue, rather than the currentake-do approach. Through a systematic examination of the rela-

ionships between awareness of increasing fatigue and sleepinesshile driving and their implications for performance effects and

rashing, this study has shown that relying on subjective fatiguend sleepiness states will provide a valid estimate of safety-relevantriving performance effects. Our task therefore is to encourage andotivate drivers to use this information responsibly.

cknowledgements

We are grateful to Jerome Favand for his contribution to dataollection. Thanks to all study participants who did without sleepor this project. This project was funded by a National Health and

edical Research Council (NHMRC) Project grant (ID568855). A.illiamson is supported by an NHMRC Senior Research Fellowship.

eferences

kerstedt, T., Gillberg, M., 1990. Subjective and objective sleepiness in the activeindividual. Int. J. Neurosci. 52 (1–2), 29–37.

merican Automobile Association Foundation for Traffic Safety, 2010. Asleep atthe Wheel: the Prevalence and Impact of Drowsy Driving. American Automo-bile Association Foundation for Traffic Safety, Washington DC, 〈http://www.aaafoundation.org/pdf/2010DrowsyDrivingReport.pdf〉, accessed 31/10/2013.

ustralian Bureau of Statistics, 2010. National Health Survey: Summary ofResults, 2007–2008 (Reissue), Data Cube, Table 1.1 Self-reported Body MassIndex—2007–08, 2004–05 and 2001(a), Persons. Australian Bureau of Statistics,〈http://www.abs.gov.au/AUSSTATS/[email protected]/DetailsPage/4364.02007-2008%20(Reissue)?OpenDocument〉 accessed 17/10/2012.

aker, F.C., Maloney, S., Driver, H.S., 1999. A comparison of selective estimates ofsleep with objective polysomnographic data in healthy men and women. J.Psychosom. Res. 47 (4), 335–341.

alkin, T.T.J., Wesensten, J.J., 2011. Differentiation of sleepiness and mental fatigueeffects. In: Ackerman, P.L. (Ed.), Cognitive Fatigue. Multidisciplinary Perspectives

nd Prevention 70 (2014) 225–234

on Current Research and Future Applications. American Psychological Associa-tion, Washington, DC, pp. 47–60.

Belz, S.M., Robinson, G.S., Casali, J.G., 2004. Temporal separation and self-rating ofalertness as indicators of driver fatigue in commercial motor vehicle operators.Hum. Factors 46 (1), 154–169.

Connor, J., Norton, R., Ameratunga, S., Robinson, E., Civil, I., Dunn, R., Bailey, J., Jackson,R., 2002. Driver sleepiness and risk of serious injury to car occupants: populationbased case control study. Br. Med. J. 324 (7346), 1125–1128.

Dinges, D.F., Pack, F., Williams, K., Gillen, K.A., Powell, J.W., Ott, G.E., Aptowicz, C.,Pack, A.I., 1997. Cumulative sleepiness, mood disturbance, and psychomotorvigiliance performance decrements during a week of sleep restricted to 4–5 hper night. Sleep 20 (4), 267–277.

Dorrian, J., Lamond, N., Dawson, D., 2000. The ability to self-monitor performancewhen fatigued. J. Sleep Res. 9, 137–144.

Dorrian, J., Lamond, N., Holmes, A.L., Burgess, H.J., Roach, G.D., Fletcher, A., Dawson,D., 2003. The ability to self-monitor performance during a week of simulatednight shifts. Sleep 26 (7), 871–877.

Gurubhagavatula, I., Maislin, G., Nkwuo, J.E., Pack, A.I., 2004. Occupational screeningfor obstructive sleep apnea in commercial drivers. Am. J. Respir. Crit. Care Med.170 (4), 371–376.

Holloway, F.A., 1995. Low-dose alcohol effects on human behavior and performance.Alcohol, Drugs Driv. 11 (1), 39–54.

Horne, J.A., Baulk, S.D., 2004. Awareness of sleepiness when driving. Psychophysio-logy 41, 161–165.

Horne, J.A., Burley, C.V., 2010. We know when we are sleepy: subjective versus objec-tive measurements of moderate sleepiness in healthy adults. Biol. Psychol. 83(3), 266–268.

Horne, J., Reyner, L., 1999. Vehicle accidents related to sleep: a review. Occup. Envi-ron. Med. 56, 289–294.

Jackson, P., Hilditch, C., Holmes, A., Reed, N., Merat, N., Smith, L., 2011. Fatigue androad safety: a critical analysis of recent evidence. In: Road Safety Web Publica-tion No. 21. Department for Transport, London, UK.

Johns, M.W., 1991. A new method for measuring daytime sleepiness: the Epworthsleepiness scale. Sleep 14 (6), 540–545.

Johns, M.W., 1992. Reliability and factor analysis of the Epworth sleepiness scale.Sleep 15 (4), 376–381.

Kaida, K., Akerstedt, T., Kecklund, G., Nilsson, J.P., Axelsson, J., 2007. The effects ofasking for verbal ratings of sleepiness on sleepiness and its masking effects onperformance. Clin. Neurophysiol. 118 (6), 1324–1331.

Kaplan, K.A., Itoi, A., Dement, W.C., 2008. Awareness of sleepiness and ability topredict sleep onset: can drivers avoid falling asleep at the wheel? Sleep Med. 9,71–79.

Maislin, G., Pack, A.I., Kribbs, N.B., Smith, P.L., Schwartz, A.R., Kline, L.R., Schwab,R.J., Dinges, D.F., 1995. A survey screen for prediction of apnea. Sleep 18 (3),158–166.

Monk, T.H., 1991. Circadian aspects of subjective sleepiness: a behavioural messen-ger? In: Monk, T.H. (Ed.), Sleep, Sleepiness and Performance. John Wiley & Sons,Chichester, pp. 39–63.

Nilsson, T., Nelson, T.M., Carlson, D., 1997. Development of fatigue symptoms duringsimulated driving. Accid. Anal. Prev. 29 (4), 479–488.

Noy, Y.I., Horrey, W.J., Popkin, S.M., Folkard, S., Howarth, H.D., Courtney, T.K.,2011. Future directions in fatigue and safety research. Accid. Anal. Prev. 43 (2),495–594.

Philip, P., Sagaspe, P., Taillard, J., Valtat, C., Moore, N., Åkerstedt, T., Charles, A.,Bioulac, B., 2005. Fatigue, sleepiness, and performance in simulated versus realdriving conditions. Sleep 28 (12), 1511–1516.

Philip, P., Sagaspe, P., Lagarde, E., Leger, D., Ohayon, M.M., Bioulac, B., Boussuge, J.,Taillard, J., 2010. Sleep disorders and accidental risk in a large group of regularregistered highway drivers. Sleep Med. 11 (10), 973–979.

Reyner, L.A., Horne, J.A., 1998. Falling asleep whilst driving: are drivers aware ofprior sleepiness? Int. J. Legal Med. 111, 120–123.

Roads and Traffic Authority, 2001. Driver Fatigue: Problem Definition and Counter-measure Summary. Roads and Traffic Authority, Sydney, Australia.

Rosa, R.R., Colligan, M.J., 1988. Long workdays versus restdays: assessing fatigueand alertness with a portable performance battery. Hum. Factors 30 (3),305–317.

Sagaspe, P., Taillard, J., Bayon, V., Lagarde, E., Moore, N., Boussuge, J., Chaumet,G., Bioulac, B., Philip, P., 2010. Sleepiness, near-misses and driving accidentsamong a representative population of French drivers. J. Sleep Res. 19 (4),578–584.

Smolensky, M.H., Di Milia, L., Ohayon, M.M., Philip, P., 2011. Sleep disorders, medicalconditions, and road accident risk. Accid. Anal. Prev. 43 (2), 533–548.

Swanson, L.M., Drake, C., Arnedt, J.T., 2012. Employment and drowsy driving: asurvey of American workers. Behav. Sleep Med. 10 (4), 250–257.

Transport for NSW, 2011. Road traffic crashes in New South Wales. In: StatisticalStatement for the Year ended 31 December 2011Author. Transport for NSW,Sydney, Australia.

Williamson, A.M., Feyer, A.-M., Friswell, R., Finlay-Brown, S., 2000. Demon-stration Project for Fatigue Management Programs in the Road Transport

Industry—Summary of Findings, CR 192. Australian Transport Safety Bureau,Canberra.

World Health Organization, 1995. Physical status: the use and interpretation ofanthropometry. In: Report of a WHO Expert Committee. WHO Technical ReportSeries 854. World Health Organization, Geneva (Author).