Road layout design standards and driver behaviour

Transcription

1 TRANSPORT RESEARCH LABORATORY Road layout design standards and driver behaviour Prepared for Quality Services (Traffic, Safety and Environment) Division, Highways Agency G Maycock (TRL), P J Brocklebank (Thorburn Colquhoun) and R D Hall (University of Southampton) TRL REPORT 332

2 First Published 1998 ISSN Copyright Transport Research Laboratory This report has been produced by the Transport Research Laboratory, under/as part of a Contract placed by the Highways Agency. Any views expressed are not necessarily those of the Agency. Transport Research Foundation Group of Companies Transport Research Foundation (a company limited by guarantee) trading as Transport Research Laboratory. Registered in England, Number TRL Limited. Registered in England, Number Registered Offices: Old Wokingham Road, Crowthorne, Berkshire, RG45 6AU.

3 CONTENTS Page Executive Summary 1 1 Introduction 3 2 Driver stratification Introduction Speed: A key behavioural parameter Stratification survey: Methodology General Site selection Speed and registration numbers The questionnaire survey The stratification survey: results Characteristics of sample Mean observed speed by age, sex and mileage Speed band distribution by age, sex, mileage and self reported speed Speed band distribution: psychological variables Speed differences between various driver sub-groups Multiple regression analysis of speed Introduction Site to site variation in speed Age, sex and annual mileage effects Speed differences between various driver sub-groups Analysis of drivers by speed group Segmentation analysis (CHAID) The proposed stratification system Accident analysis Introduction Accident tabulations Accident models Model results Speed and accidents 14 3 Emergency stopping behaviour Introduction Previous studies Study methodology The driving simulator Experimental protocol Survey sample Reaction times Introduction Simple reaction times to STOP signs 17 iii

4 Page Reaction times to the four simulated driving hazards Determinants of reaction times Analysis of mean reaction times Discussion of models of mean reaction times Drivers with long reaction times Discriminant analysis 22 4 Overtaking behaviour Introduction Previous studies Study methodology Introduction Simulator study Public road observational study Results: preliminary analyses Simulator study On-road study Statistical modelling The form of the models The on-road data model The simulator data model Combined data analysis 31 5 Summary and discussion Introduction Driver stratification Driver speeds Analysis of speed data Accident frequencies Driver stratification Speed and accidents Emergency stopping performance Study methodology Reaction times Determinants of reaction times Analysis of drivers with long reactions Overtaking behaviour Study methodology Results of the analyses 34 6 Relevance to design standards 34 7 Acknowledgements 36 iv

5 Page 8 References 36 Appendix A: Survey sites 39 Appendix B: The questionnaire survey 40 Appendix C 42 Abstract 43 Related publications 43 v

6

7 Executive Summary Introduction The Highways Agency, who have responsibility for the design and maintenance of trunk roads in England, have commissioned a review of certain aspects of the current standards to take account of the considerable amount of research into driver behaviour that has been conducted in recent years. A major part of this consultancy exercise was to undertake experimental studies of the role of driver characteristics in three specific aspects of the driving task - speed choice, emergency braking and overtaking. The objective of these studies was to provide design engineers with more detailed insights into the types of driver using the roads and their behaviour in critical manoeuvres. This work has been undertaken by a consortium consisting of the Transport Research Laboratory (Speed choice), Thorburn Colquhoun, (Emergency stopping) and University of Southampton (Overtaking). The three studies are described in this report. Driver stratification (speed choice) The objective of the study of speed choice was to identify those characteristics of the driving population which are influential in determining the speed at which drivers choose to drive on various types of road. From these characteristics a simple classification or stratification of driver types has been derived. In order to obtain the necessary data, measurements of vehicle speed were made on a sample of trunk roads. These speed observations were subsequently linked by means of a follow-up postal questionnaire survey to the demographic and psychological characteristics of the drivers, and to selfreported accident data. Responses were available for analysis from 6,435 drivers observed on 43 sections of single and dual carriageways and motorways. Using multivariate methods, those factors which influence both speed choice and accident involvement were identified. In addition a hierarchical segmentation analysis procedure was used to produce a simple classification of drivers. The analysis of the speed data showed that speed choice strongly depends on age, annual mileage, and trip frequency. It also depends on some of the observed trip characteristics, and to a lesser extent on gender, road type and the psychological variables measured. In terms of the speed of an individual driver relative to the speed of all drivers, faster drivers tend to be younger rather than older and to drive high annual mileages in company cars; they were also likely to be in the managerial, administrative or professional occupational groups and to be travelling without passengers for business purposes. In terms of relative speed, the influence of road type (single carriageways compared to dual carriageways and motorways) was small. A simple driver stratification was developed in which drivers were grouped into two mileage categories (<10,000 and 10,000+ miles annually) and two age categories (<44 and 45+ years of age) giving a four cell stratification. Accident involvement was related to age, annual mileage and driving experience. By using the predicted speed as an explanatory variable in the accident model a relationship between drivers speeds and accidents was developed which suggested that a 1 per cent change in an individual driver s choice of speed is associated with a 13.1 per cent change in that individual s accident liability. Further work is required to show to what extent this association is causal. Emergency stopping Emergency stops are necessary when something unexpected happens, such as inappropriate driving behaviour by someone in front, or an obstruction on the road ahead. The objectives of the study of emergency stopping behaviour were to explore the determinants of emergency stopping and to obtain estimates of emergency stopping times. In order to measure emergency stopping behaviour, a portable driving simulator was developed which consisted of a driving rig made up of a car seat, a steering wheel and normal driving pedals. Foot pedal sensors were linked to a computer which also controlled a video recorder and television monitor. Simulated road scenes in which six emergency stop situations were incorporated were presented to subjects using this simulator. The emergency stop hazards comprised two simple reaction tests in which drivers had to respond as quickly as possible to the word STOP superimposed on the video image, and four real driving hazards. In total, some 2,700 emergency stops undertaken by 558 subjects were measured. In addition, subjects provided demographic and psychological information using the same questionnaire used in the driver stratification project. The mean brake reaction time from the first and second simple reaction (STOP) tests combined, was 0.58 seconds. The characteristics of the four real hazards presented varied considerably and elicited different responses from subjects. For hazards 1 and 2 (vehicles emerging reasonably rapidly from side roads) the mean brake response times were 1.05 and 1.01 seconds after the time at which the hazard vehicle first appeared in the visual scene. Hazard 3 developed much more slowly, and the mean brake response to this hazard was 1.31 seconds. By contrast, in hazard 4 the vehicle encroached on the carriageway very rapidly, the corresponding mean response time after the vehicle first appeared was only 0.67 seconds. Analysis of the data shows that reaction times fell as the test progressed. Taking this effect into account the ranges for the mean unalerted brake reaction times are estimated to be seconds for the four hazards; the simple averages over the four hazards are respectively 1.21 for the mean and 1.61 for the 85 percentile value. Regression models were used to explore the relationships between the various reaction times and the individual characteristics of drivers. Reaction times would appear to be dependent on age, occupational group, the 1

8 frequency of driving, the number of accidents the driver has reported in the previous 3 years, and some of the psychological measures. Overtaking The study of overtaking behaviour investigated accelerative overtaking on rural single carriageway roads. Of principal interest was the distance required to complete such an overtaking manoeuvre safely and how this distance was related to the overtaking opportunity provided and to the characteristics of the driver. Data was obtained in two ways - using a simulator, and making observations of overtaking in real on-road situations. The TRL driving simulator was used to present a range of overtaking opportunities to a sample of 40 driver subjects. The overtaking opportunities were presented as straight sections of single carriageway road of 5 different lengths for each of two design speeds (85 and 100 kph). The vehicle to be overtaken travelled at a constant speed either 15 or 30 kph below the design speed of the road; in half of the overtaking situations a vehicle was approaching in the opposing lane. Observations of drivers in real on-road overtaking opportunities were made by driving a vehicle at constant speed along a section of road which included an overtaking section. Any member of the driving public following the experimental vehicle at the start of the overtaking section was thus presented with an overtaking opportunity similar to that provided in the driving simulator. The behaviour of the following vehicle was timed and recorded. Such measurements were made at 18 sites, and included a total of 2,228 potential overtaking opportunities presented to 624 drivers. The questionnaire used in the driver stratification project was used to collect the personal characteristics of the drivers observed on the road. The analysis used logistic modelling to relate the probability of a driver accepting an overtaking opportunity to the explanatory variables of road, situation and driver characteristics. The analysis showed that for both the simulator data and the on-road data, the key variables were those associated with the length of the overtaking gap offered and the speed of the vehicle to be overtaken. Age was also a significant variable in both the simulator and the on-road overtaking with older drivers requiring longer overtaking gaps than younger drivers. From the on-road data, it would also appear that the probability of overtaking is higher for high mileage drivers than for low mileage drivers and lower for female drivers than for males. 2

9 1 Introduction The Transport Research Laboratory (TRL) was commissioned by the Highways Agency to conduct a review of certain aspects of trunk road design standards, taking account of current knowledge about driver characteristics, behaviour and safety. The overall aim of the programme of work was to provide highway designers with insights into the impact of driver characteristics and behaviour on some aspects of trunk road design. A major part of this consultancy was to undertake experimental studies of the role of driver characteristics in three specific driving tasks - speed choice, emergency braking and overtaking. The objective of these studies was to provide design engineers with more detailed information about the types of driver using the roads and their behaviour in critical manoeuvres. This work has been undertaken by a consortium consisting of the Transport Research Laboratory (Speed choice), Thorburn Colquhoun, (Emergency stopping) and the University of Southampton (Overtaking). This report describes these three experimental studies. Following this introduction, sections 2, 3 and 4 describe respectively a simple approach to classifying drivers on the basis of their speed choice (driver stratification), and the studies of emergency stopping and overtaking. Section 5 provides a summary and conclusions. Appendix A lists the sites surveyed in this study and Appendix B gives details of the questionnaire used in the stratification study. Appendix C gives a brief description of the statistical modelling process used in the analysis of accidents and of overtaking probabilities. 2 Driver stratification 2.1 Introduction The driver stratification study was intended to provide a broad behavioural classification of drivers using trunk roads in Great Britain. The intention in developing such a stratification was that it should be meaningful in terms of the highway design process and that it would act as the basis for classifying drivers in the emergency stopping and overtaking studies. It was proposed that the classification should be based on the speed at which drivers choose to drive, since speed is a key behavioural characteristic which is readily observable. Thus, the principal objective of the driver stratification study was to identify those features of the driving population which are of importance to individual speed choice on various types of trunk road, and which could be of practical importance when making design decisions. The study was also designed to identify the accident liability of the groups of drivers defined by the resulting stratification. 2.2 Speed: A key behavioural parameter In the present highway design standards philosophy, design speed, the 85th percentile speed of traffic on the road, is the key parameter which links many aspects of the design together - notably horizontal and vertical curvature, sight distances and superelevation. However, the concept of the 85th percentile driver though simple and convenient, may not be entirely realistic. It is possible therefore that a more meaningful classification of drivers based on the speed at which individual drivers choose to drive might provide designers with greater insights into the needs of road users than they currently have. There are potentially many factors that can influence a driver s speed and safety, and a considerable amount of research has been undertaken exploring these factors. Some studies have been based on the actual driving speeds of individual drivers observed at a specific site on a public highway at one point in time. Such studies depend on the assumption that the speed selected by a driver on one particular occasion at one particular location reflects their speed at other times and on other roads. Wasielewski (1984) has provided some evidence that individual speed choice is consistent over time. Moreover, in this type of study, valid measurements of a driver s speed choice can only be made if the driver is free to set his or her own preferred speed. Evans and Wasielewski (1983) have suggested that vehicles travelling with a headway of 2.5 seconds can be considered to be driving at their own preferred speed. The driver stratification study being reported here has adopted a criterion of 3 seconds for such free speeds. Based on speed measurements in four European countries, Smeed (1973) found that both driver age and driver sex were good predictors of speed choice, with younger drivers driving faster than older drivers and men driving faster than women. These findings are supported by results reported by Herberg (1978) and Galin (1981). Wasielewski (1984) in a study of risk taking behaviour also found that younger drivers drive faster, but that when age differences were allowed for using a multivariate model, there were no speed differences between the sexes. Fildes et al (1991) in a study of fast drivers also found that driver age is a significant factor in speed choice but that driver sex was not a significant factor; he also found, among other things, that business travellers were likely to be over-represented among the faster drivers. In addition to the results obtained from roadside surveys of speed, there is a sizeable body of research that has explored driver behaviour - including speed choice - using in-depth laboratory tests or self-report questionnaire data (see Elander el al, 1993, for a recent review). There have been numerous attempts over the years to relate driving behaviour to personality factors and cognitive skills such as hazard perception, with varying degrees of success (Quimby and Watts, 1981; Wilson and Greensmith, 1983), and there is recent evidence to show that characteristics such as driving style or decision making style - whether measured by laboratory tests (Quimby et al, 1986) or by means of questionnaires (French et al, 1993) do reflect those individual characteristics which influence driving performance. Moreover, a driver s willingness to break traffic laws - including speed limits - or commit driving violations has been found to be an important behavioural characteristic (Reason et al, 1991, Parker et al, 1992) which is related to accident liability. 3

10 It would seem clear therefore that the speed at which an individual chooses to drive is influenced by a wide variety of demographic, situational, cognitive and motivational factors. The questionnaire used in the present survey as well as collecting the basic demographic and trip specific information, includes in addition, questions designed to obtain measures of driving style, decision making style, hazard detection ability and tendency to violate traffic rules. 2.3 Stratification survey: Methodology General In order to relate an individual driver s choice of speed to his or her personal characteristics, it is necessary to collect both observed on-road speed data and personal information from the same group of subjects. In the present study this was achieved by collecting the information in two stages. The first stage consisted of making speed measurements of a sample of drivers/vehicles on a selection of trunk roads, and at the same time noting the car s registration number. The second stage involved identifying the owner of the vehicle from the records held by the Driver and Vehicle Licensing Agency. The owners of the sampled vehicles were then sent an individual letter informing them of the location, date and time the vehicle was observed and asking them to complete and return a questionnaire (enclosed with the letter) if they were driving on that occasion, or to pass the questionnaire to whoever was driving at that time. Sections briefly describe the selection of road sites, the strategy adopted for the sampling of driver speeds and registration numbers and the questionnaire survey Site selection Speed surveys were carried out on single and dual carriageway trunk road sites and motorways in both rural and urban areas. The sites were sampled so as to be representative of roads in Great Britain, and within each road type sites were chosen to give a range of curvatures and/or sight distances arising from both horizontal and vertical curvature. In all, 43 sites were used - 22 on single carriageways, 10 on dual carriageways and 10 on motorways; they are listed in Appendix A together with an indication of the horizontal and vertical curvatures existing at the selected sites Speed and registration numbers In general spot speeds were obtained by the use of a radar gun, although at some motorway sites time lapse photography was used. The surveys were carried out in two phases, the first undertaken by the TRL photographic unit and the second phase sub-contracted to Babtie Ltd. In the first phase, a continuous record of vehicles passing a survey point was recorded on video tape with the speed of each free flow vehicle electronically superimposed onto the video record. The vehicle speeds and the registration numbers were later transcribed in the Laboratory from the video record. The second phase of the survey used radar guns for speed measurement and registration numbers were recorded manually on site. Because free flow speeds were required (defined as vehicles with time headways of at least 3 seconds between them and the vehicle in front), speed measurements were taken in relatively uncongested conditions during off-peak periods or in the evening. Attempts to obtain registration numbers in the dark were generally unsuccessful, and as a result, no day/night comparisons have been possible. Only cars (not HGVs, LGVs or motorcycles) in a single specific lane were sampled at any one time. In the case of dual carriageways and motorways each lane was sampled in turn for a similar period of time (typically 10 minute intervals) so that the number of vehicles eligible for sampling was proportional to the traffic flow in each lane. The speed distribution obtained therefore represented the free flow speeds of all vehicles passing a particular point on the road. Once the vehicle speeds had been transcribed (from the video), and the whole speed distribution at a site was known, the drivers who were to be included in the questionnaire survey were selected according to their speed relative to the characteristics of the speed distribution at the particular site and time they were observed. Drivers were sampled in five speed bands as shown in Table 1. Table 1 Driver sampling strategy in relation to observed speeds Speed range Sample Proportion of respondents Band 1 (Fast) Speed > 85th %ile All 18% Band 2 70th %ile < speed < 85th %ile All 19% Band 3 Around the median speed 3 out of 8 20% Band 4 15th %ile < speed < 30th %ile All 21% Band 5 (Slow) Speed < 15th %ile All 23% The drivers of particular interest in relation to individual speed choice are those in the upper and lower tails of the speed distribution. To maximise the efficiency of the subsequent analysis therefore, all drivers in bands 1, 2, 4 and 5 were included in the sample, but only a proportion of drivers in the 40 percentile band around the mean; the latter sample amounted to 15 per cent of the total - ie the same number of drivers as in each of the other 4 bands. The final column in Table 2 shows the actual proportion of drivers responding to the survey and available for analysis. It will be seen that the response rate is rather higher from the low speed drivers than it is from the high speed drivers. Nevertheless, the sample represents a reasonable distribution of drivers across all 5 speed bands The questionnaire survey The questionnaire survey is described in detail in Appendix B. The questionnaire asked about factors such as trip purpose, passengers carried and whether the car was company owned or privately owned - factors which would be relevant to the particular journey being made when the driver s speed was measured. The questionnaire then asked respondents to rate their own driving style, and included groups of questions which formed psychological scales 4

11 Table 2 Distribution of drivers ages Percentage of drivers Age group Males Females Both years years years years years and over Average age Total Numbers Table 3 Miles driven per year Percentage of drivers Annual mileage band Males Females Both Up to 5, ,001-10, ,001-15, ,001-25, ,001 and over Average mileage 19,680 10,480 17,170 Total Numbers designed to measure hazard involvement, willingness to violate a range of traffic rules (formal and informal), and decision making style. These four psychological scales are detailed in Appendix B which gives a Table giving the mean value for each scale, its range and some reliability statistics. The questionnaire also asked drivers how many accidents they had been involved in as a driver over the last three years and whether these accidents had involved injury. Accidents were defined as any incident which occurred on a public road and which involved injury or damage to property; the accident data was thus expected to be mainly damage only accidents. Finally, drivers were asked to provide the following personal details: age, how long they had held a full driving licence (driving experience), the frequency with which they made driving trips, annual mileage, sex and occupational group. Questionnaires were sent to 14,050 drivers in all, and the overall response rate for valid questionnaires was 46 per cent (see Appendix B). Because of missing data, the number of respondents available for inclusion in the analyses to be described in the following sections, will vary from analysis to analysis. 2.4 The stratification survey: results Characteristics of sample Tables 2 and 3 show how the sample of respondents is distributed by age, sex, and annual mileage. It can be seen that overall the sample included over twice as many men as women drivers. There is a fairly even spread of ages, though the female drivers are on average rather younger than the male drivers. Table 3 shows that annual mileage driven by women is on average about half that driven by men. Overall, the drivers in this sample are high mileage drivers, driving about twice the national average mileage Mean observed speed by age, sex and mileage Table 4 shows the mean speed of the sampled drivers by age group for male and female drivers driving on single carriageways and dual carriageways (including motorways). There is a clear age effect for male drivers; year old male drivers drive 3-5 mph faster than the over 60s. Women drivers drive slightly slower than men and the age effect is not so pronounced. Table 4 Mean speed (mph) by age group, sex and road type Single Dual carriageways carriageways and motorways Age group Males Females Males Females years years years years years and over All age groups Table 5 shows the mean speeds by annual mileage groups. The table shows that high mean speeds are strongly associated with high mileage drivers for both men and women, though it remains to be seen from the results of the multivariate modelling whether or not this apparent mileage effect is due to a correlation between age and mileage - young higher speed drivers tending to drive higher mileages. Table 5 Mean speed (mph) by annual mileage group, sex and road type Single carriageways Dual carriageways and motorways Annual mileage Males Females Males Females Up to 5,000 miles ,001-10,000 miles ,001-15,000 miles ,001-25,000 miles ,001 and over All mileage bands Speed band distribution by age, sex, mileage and self reported speed As an alternative way of viewing the speed effects in relation to age, Table 6 shows the proportion of drivers (male and female) in the five speed bands as a function of age (grouped). Focusing on the band 5 (slow) drivers, it will be seen that only 15 per cent of young male drivers feature in this band compared with 32 per cent of drivers 5

12 over 60. Conversely, 28 per cent of year old male drivers feature in band 1 (fast), compared with only 9 per cent of over 60s. A similar effect can be seen in relation to female drivers in the lower half of the Table. Table 6 Distribution of sample by age group and speed band Percentage of drivers Speed Speed Speed Speed Speed Number Age group band 5 band 4 band 3 band 2 band 1 of cases (Slow) (Fast) Males years years years years years and over All ages Females years years years years years and over All ages A similar Table (not reproduced here) reflecting the speed-mileage effect, shows that low mileage drivers are more likely to be found in speed band 5 (slow) and high mileage drivers in band 1 (fast). For example, for all drivers in the sample, slightly over 30 per cent of drivers who drive less than 5,000 miles annually feature in band 5 (slow) compared with about 15 per cent of those who drive over 25,000 miles annually. Conversely, only 10 per cent of those drivers who drive less than 5,000 miles per year feature in band 1 (fast) compared with 26 per cent of those who drive over 25,000 miles annually Speed band distribution: psychological variables As a preliminary indication of the usefulness of the four psychological variables in predicting speed effects (see Appendix B for definitions), simple one-way analyses of variance were carried out using each of the psychological variables in turn as the dependent variable and with the speed band as the grouping variable; to allow for age and mileage effects, these terms were included in the analysis as covariates. The results are shown in Table 7. With the exception of the analysis using hazard involvement, age and annual mileage were always massively significant statistically. In the analysis of hazard involvement, annual mileage was significant but age was not. With the effects of these variables allowed for, Table 7 shows that two of the four psychological variables - driving style and violation score - showed significant differences between the speed bands. The two nonsignificant scales were decision making and hazard involvement. Table 7 Analyses of variance exploring the effect of the psychological variables on speeds within the five speed bands with age and annual mileage as covariables Scale score (Speed band Psychological 5 (slow) - variable band 1 (fast)) Significance F Driving style Significant (F=4.4, p<0.001) Hazard involvement Not significant (F=1.6) Violations Significant (F=19.3, p<0.001) Decision making Not significant (F=1.5) Speed differences between various driver subgroups The questionnaire data enabled the average speeds of drivers to be estimated in terms of the following characteristics of the trip they were making when observed: (i) whether they were driving a private or a company car, and (ii) whether they were carrying a passenger or not per cent of the male driver respondents and 86.7 per cent of the female driver respondents in the survey were driving private cars. The average (uncorrected) speeds of male and female drivers driving private cars was about the same (50.7 and 49.7 mph respectively), but the average speed of those driving company cars was about 7-8 mph faster than that of private car drivers (57.8 and 57.5 mph for male and female drivers respectively). With regard to the effect of passengers, just over a third of both men and women were carrying a passenger (35.5 per cent of men and 33.2 per cent of women). Drivers carrying passengers drive somewhat slower than those who are not carrying passengers. The average speeds for male and female drivers who were carrying passengers was 50.9 and 49.5 mph respectively, compared with 54.5 and 51.4 mph for male and female drivers driving without passengers. Table 8 shows the proportion of drivers in the sample Table 8 Occupational group and mean speeds Percentage of drivers Mean speed (mph) Occupational group Males Females Males Females 1. Senior managerial, administrative or professional 2. Junior managerial, administrative or professional 3. Skilled manual Semi skilled or unskilled manual 5. Student, housewife/ husband, unemployed 6. Retired Total numbers

13 classified into six occupational groups. About 60 per cent of the male drivers claim to be in the senior or junior managerial, administrative or professional group, with less than 3 per cent classing themselves as students, househusbands or unemployed. Just over 56 per cent of women drivers assessed themselves as being in the two managerial, administrative or professional group categories, with a further 25 per cent in the fifth category which includes housewife. There are systematic speed differences between these occupational groups with the senior managerial drivers driving the fastest, and retired drivers the slowest. 2.5 Multiple regression analysis of speed Introduction The aim of the analysis reported in this section is to relate the observed speeds of individual drivers to road type, demographic and psychological variables using multiple regression which can take into account the inter-correlations between the potential explanatory variables. Figure 1 shows a box plot of the speed distributions observed at the 43 sites at which measurements were made - ordered in terms of ascending mean speeds; the central band of the box plot is the percentile range. It will be seen that the spread of speeds increase as the mean speed increases. In order therefore to stabilise the variance of these observations and to improve the Normality of the speed distributions, the following analysis has been based on the natural logarithm of observed speed (Ln(speed)) as the independent variable. The use of the log transformation has the added advantage that the resulting regression equation expresses the explanatory effects as speed ratios rather than speed differences. This is particularly useful in relation to the site to site variation which will now be considered Site to site variation in speed Figure 1 shows that there are relatively large site to site variations in mean speed. The first step therefore in the speed analysis was to fit site as a category variable with Speed (mph) Figure 1 Speed distributions for individual sites ordered by mean speed 7

14 43 levels. The resulting model is: ln(v ij ) = [Mean ln(v i )] j=1 to 43 + ln(s ij ) (1) Where: V ij is the speed of vehicle i (i=1...n j ) at site j, and the term in the square brackets is the mean value of lnv averaged over all n j vehicles observed at site j. Ln(S ij ) - the residual of the regression equation at this stage - is the difference on a log scale of the speed of the individual vehicle i to the mean (logged) speed at the site j. S therefore (dropping the subscripts and the brackets for clarity) represents the within site speed effect relative to the geometric mean speed of vehicles at that site. LnS will be used as the dependent variable in the subsequent analysis. The site to site variation in speed represents a large proportion of the total variation in the speed measured in the survey. Based on lnv, the total sums of squares about the mean for the dataset as a whole is The addition of the site term (equation (1)) results in a residual sum of squares of Thus 77 per cent of the variation in observed speed is due to between-site variations in road type, road geometry and other site specific effects such as weather and road surface condition. It is also possible, even though the observed vehicles were intended to be free flowing, that traffic flow conditions may have influenced site mean speeds. As will be seen from the subsequent analyses the use of the speed ratio (lns) largely eliminates site to site effects. This means that the ratio of an individual driver s speed on a given section of road to the mean speed of all drivers on that road is largely independent of the type of road involved and its absolute mean speed Age, sex and annual mileage effects The present section aims to determine the age, mileage, trip frequency, road type and sex effects on relative speed using lns as the dependent variable. A linear regression model was fitted in three stages. The first stage was to fit a model to a sub-set of the data containing 6435 cases having no missing data for age, annual mileage, trip frequency, sex or road type. Trip frequency is the driver s estimate of the average number of days he or she drives in a year; it has been found in other studies to be a useful exposure measure in addition to annual mileage (Forsyth et al, 1995). The resulting model may be written as follows: lns= k + b 1 AGE + b 2 MILEAGE + b 3 FREQUENCY + b 4 SEX + b 5 SIND (2) Where: SIND is a 2-level factor distinguishing single carriageway roads from dual carriageways (including motorways), and k and the b s are coefficients to be determined. The model coefficients for equation (2) are shown in Table 9. The right hand column of the Table shows the size of the effect in terms of the percentage change in relative speed resulting from the predictor variable changing from the 5th percentile point of its distribution (below the mean) to the 95th percentile point (above the mean) - other variables in the model being held at their mean values. In the case of variables such as sex or road type with only two categories, the single figure indicates the percentage difference in speed between the two categories. Table 9 LnS (speed ratio) as a function of age, annual mileage, trip frequency, sex and road type Coefficient Variable (t-value) Effect size for S Constant (Male driver on k a single carriageway road) Age b (13.9) +4% to -4.2% Mileage b (6.5) -1.2% to +2.5% Trip frequency b (3.6) -2.2% to +0.2% If female, add... b (3.0) -1.2% If dual carriageway/ motorway, add... b (3.1) -1.1% Age has by far the largest effect on relative speed. The effect size for age in Table 9 shows that the speed of a driver aged 23 (the 5th percentile age) is 4 per cent above the site mean speed, whilst a driver aged 71 (the 95th percentile age) is just over 4 per cent below the site mean speed - an overall change of 8 per cent in speed across the age range. The mileage effect has the opposite sign. A driver driving only 3,000 miles a year (the 5th percentile mileage) has a speed which is just over 1 per cent below the site mean speed, whilst a driver driving 45,000 miles a year (the 95th percentile mileage) has a speed which is 2.5 per cent above the site mean - an overall 3.7 per cent effect. The frequency of making trips by car is also a significant factor in determining speeds. Most drivers drive nearly every day (365 times a year), however, drivers who drive less frequently (the 5th percentile point has been taken to be 100 times a year - about twice a week) drive at a speed which is just over 2 per cent slower in relation to the site mean than drivers who drive more frequently. The overall difference between the sexes corresponds to an effect on speeds of about 1.2 per cent relative to the site mean speed, with women driving slower than men. There were no sex interactions with any of the other variables in this model. The road type factor (SIND) is also of interest. In the preliminary analysis, this factor was a three level factor identifying three types of road - single carriageways, dual all purpose roads and motorways. It turns out that only a 2- category distinction between single and dual carriageways (including motorways) can be justified statistically. The negative coefficient of this term shown in Table 9 implies that in relation to the site mean speeds, drivers on dual carriageways and motorways drive just over 1 per cent slower than drivers on single carriageways - it must be stressed that this is relative to the site mean speeds across the range of sites, not that drivers on dual carriageways and motorways are driving more slowly in absolute terms. Although this effect is statistically significant, it is not very large, and probably reflects a subtle difference in the shape of the speed distribution on the two types of road. It seems 8

15 unlikely to be of any great practical significance. In terms of residual variance, the addition of age and mileage terms reduces the initial residual sum of squares of (that is after the site to site variation has been removed) to a reduction of 5.3 per cent Speed differences between various driver subgroups In order to evaluate the effect of the various category variables on speed adjusting for age, annual mileage, driving frequency and sex, each of the variables of interest were added to the above model singly, with the resulting coefficient quantifying the magnitude of the effect and its statistical significance. The results are shown in Table 10. Table 10 The associations between speed, and other category variables corrected for age, mileage, trip frequency and sex effects Category Level N Coefficient (b n ) Effect (t-value) on speed Vehicle Private 4530 R ownership Company (6.22) 2.7% Carrying Yes 2219 R passengers? No (5.8) 2.2% Journey Other 3743 R purpose (including work) Leisure, (4.8) -1.9% shopping etc. Occupational Senior 3758 R group and junior managerial, administrative or professional Manual (7.3) -2.8% workers, students, housewives, unemployed, retired The coding scheme illustrated in Table 10 is such that the first level of each category variable is always the reference category (denoted by R in the table); the values of the coefficients estimated for the other levels of each category then reflect the differences in lns between the reference level (the first) and each of the other levels. Moreover since S is the ratio of speed to the site mean speed, it follows that 100(exp(b n ) - 1) - the effect on speed given in the final column of the Table - is the percentage amount by which the speed associated with a particular level is greater (positive) or less (negative) than the speed associated with the first (reference) category - all effects being corrected for differences in age and mileage between the various levels. The results are given for men and women drivers combined. Table 10 shows that company car drivers drive 2.7 per cent faster than those driving private cars. Drivers carrying passengers drive just over 2 per cent slower than those not carrying passengers and drivers engaging in driving for leisure and shopping activities, drive just under 2 per cent slower than those travelling for other purposes - including travelling to and from work and driving during the course of work. In the case of occupational group, the original five questionnaire categories (senior managerial, junior managerial, skilled manual, semi-skilled manual, housewives, and retired people) have been combined to give the two categories shown in the Table. Senior and junior managers drive about 2.8 per cent faster than the other occupational groups. The third stage in the analysis of speed was to explore the effect of the four psychological variables - driving style, hazard involvement, violation score, and decision making style. These variables were added to the model of Table 9 above using a restricted sample of drivers for whom data was available for all four scales. This procedure showed that as in the univariate analysis (Table 7) violation score was a strong predictor of speed - not surprisingly since the violation scale includes a measure of the driver s tendency to exceed speed limits. The coefficient of this term was (t=8.6), highly significant statistically, with the positive sign indicating that higher violation scores are associated with higher speeds. Unlike the univariate results, the driving style scale was not a significant predictor of speed once the other effects had been taken into account, but decision making style was just significant with a coefficient of (t=2.8). 2.6 Analysis of drivers by speed group Segmentation analysis (CHAID) As described earlier, the questionnaire respondents were sampled in five speed bands - the lowest and highest 15th percentile bands and three intermediate bands. The kind of data available by classifying driver responses in the five speed bands is illustrated in Table 6. This section describes a segmentation analysis based on this banded data using the CHAID (Chi-square automatic interaction detector) algorithm. CHAID is a statistical algorithm designed to divide a set of cases (such as drivers in the present study) into mutually exclusive groups each of which differs significantly from the others with respect to a specific parameter or variable. The CHAID algorithm builds hierarchical tree structures which are useful for visualising interactions in complex data sets. In the present study the CHAID process was used to develop a simple stratification of drivers based on the variables which have been shown to be statistically sound predictors of the relative speed of drivers. The process requires a dependent variable - in this case, the numbers of drivers in each of the five speed bands - and a list of predictor (splitter) variables which are used for building the tree structure. Because the programme uses categorical (grouped) data only, the continuous variables of interest - age and mileage - were grouped into simple binary categories split at 44/45 for age (<44, 45+) and 10,000 for mileage (<10,000, 10,000+). Figure 2 9

16 10 ALL DRIVERS 1: 18.1% 5: 22.5% N=5979 Spd: 52.5 mph Age: 46.4 yrs Miles: 17,170 AcF: 0.289/3y MALE 1: 19.2% 5: 21.8% N=4340 Spd: 53.2 mph Age: 47.6 yrs Miles: 19,680 AcF: 0.300/3y FEMALE 1: 15.4% 5: 24.2% N=1639 Spd: 50.7 mph Age: 43.0 yrs Miles: 10,480 AcF: 0.258/3y <10,000m 1: 12.1% 5: 28.8% N=1322 Spd: 48.2 mph Age: 54.5 yrs Miles: 6,970 AcF: 0.181/3y >10,000m 1: 22.3% 5: 18.8% N=3018 Spd: 55.4 mph Age: 44.6 yrs Miles: 25,160 AcF: 0.351/3y <10,000m 1: 12.6% 5: 25.8% N=1050 Spd: 48.8 mph Age: 45.1 yrs Miles: 5,890 AcF: 0.223/3y >10,000m 1: 20.4% 5: 21.2% N=589 Spd: 54.0 mph Age: 39.4 yrs Miles: 18,460 AcF: 0.319/3y YOUNGER OLDER YOUNGER OLDER YOUNGER OLDER YOUNGER OLDER 1: 18.8% 5: 19.6% N=373 Spd: 50.2 mph Age: 33.0 yrs Miles: 7,230 AcF: 0.244/3y 1: 9.5% 5: 32.5% N=949 Spd: 47.4 mph Age: 63.0 yrs Miles: 6,870 AcF: 0.156/3y 1: 28.3% 5: 15.7% N=1476 Spd: 56.5 mph Age: 34.1yrs Miles: 26,360 AcF: 0.448/3y 1: 16.5% 5: 21.7% N=1542 Spd: 54.3 mph Age: 54.9 yrs Miles: 23,980 AcF: 0.257/3y 1: 15.5% 5: 23.3% N=515 Spd: 49.5 mph Age: 32.9 yrs Miles: 6,110 AcF: 0.277/3y 1: 9.7% 5: 28.2% N=535 Spd: 48.1 mph Age: 57.4 yrs Miles: 5,670 AcF: 0.169/3y 1: 24.2% 5: 18.4% N=385 Spd: 54.5 mph Age: 31.8 yrs Miles: 18,890 AcF: 0.387/3y 1: 13.2% 5: 26.5% N=204 Spd: 53.2 mph Age: 53.2 yrs Miles: 17,670 AcF: 0.197/3y SINGLE 1: 22.2% 5: 17.7% N=221 Spd: 42.1 mph Age: 33.5 yrs Miles: 7,260 AcF: 0.235/3y DUAL & MWAY 1: 13.8% 5: 22.4% N=152 Spd: 60.1 mph Age: 32.4 yrs Miles: 7,200 AcF: 0.254/3y SINGLE 1: 34.9% 5: 13.5% N=631 Spd: 46.1 mph Age: 33.9 yrs Miles: 24,840 AcF: 0.464/3y DUAL & MWAY 1: 23.4% 5: 17.4% N=845 Spd: 63.1 mph Age: 34.2 yrs Miles: 27,340 AcF: 0.437/3y SINGLE 1: 17.5% 5: 19.4% N=320 Spd: 42.7 mph Age: 33.5 yrs Miles: 5,900 AcF: 0.291/3y DUAL & MWAY 1: 12.3% 5: 29.7% N=195 Spd: 58.6 mph Age: 32.3 yrs Miles: 6,400 AcF: 0.258/3y Figure 2 CHAID tree constructed using sex, annual mileage, age and road type as predictor variables

17 shows a CHAID tree for the datafile consisting of 5979 driver responses with no missing data for any of the splitter variables used. Shown in each leaf of the tree are the percentages of drivers in speed bands 1 and 5 (as before band 1 includes the faster drivers and band 5 the slower drivers), and immediately below these figures, the number of drivers contributing to these percentages. In the lower section of each leaf are the average speeds, average ages, average annual mileages and the mean accident frequencies for all drivers in the sub-groups defined by the tree segments. Figure 2 shows that the first split has been made on the basis of sex. As expected from the earlier analyses, the sex effect though significant is not large. Compared with an overall 18.1 per cent of drivers in the fast speed band, 19.2 per cent of men are in this band and only 15.4 per cent of women. Average speeds for men and women are 53.2 and 50.7 mph respectively. The next level has been split for both men and women on the basis of annual mileage - drivers driving less than or equal to 10,000 miles annually and those driving more than 10,000 miles annually. In terms of percentages in the speed bands and average speeds, Figure 2 shows that low mileage drivers drive more slowly than high mileage drivers; for male drivers for example, moving one level down in the hierarchy, the 19.2 per cent of drivers in the highest speed band becomes 12.1 per cent for the low mileage drivers and 22.2 per cent for the high mileage drivers. The average speed for these two groups of drivers are 48.2 and 55.4 mph respectively. The percentage occupancy of the speed bands and the mean speeds are similar for high and low mileage female drivers as for the male drivers. This result would seem to imply that the main reason for the difference in mean speed between male and female drivers is not due fundamentally to different driver behaviour between the sexes, but arises because there are far more high mileage male drivers than there are high mileage female drivers in the sample. In fact, the characteristics of all the sub-groups of the tree below the first level are very similar for men and women drivers such that a tree consisting of mileage and age alone would be a fair representation of the structure of the data. The next split (level 3) has been made on the basis of age, and a binary split at about the mean age for the sample has been used - drivers aged less than or equal to 44, and those aged 45 or over. As expected younger drivers drive faster than older drivers and a higher proportion of young drivers feature in the fast speed band. Thus for example, 12.1 per cent of male, low mileage drivers are in band 1 (fast) and this splits to 18.8 per cent for the younger drivers (mean speed 50.2 mph) compared with only 9.5 per cent of older drivers (mean speed 47.4 mph). Only one further split was possible. For younger drivers only (and in the case of women drivers, younger low mileage drivers only), the 2-level road type factor SIND, distinguishing single carriageways from dual carriageways proved a significant splitting factor. As with the earlier analyses, the effect is counter-intuitive. In relation to the mean speeds on these two groups of road, driving speeds are relatively higher on single carriageways than on dual carriageways - an effect which is consistent across the three younger driver groups. For male, high mileage, younger drivers for example, the combined percentage of drivers in the highest speed band is 28.3 per cent (mean speed 56.5 mph averaged across all road types); on single carriageways this increases to 34.9 per cent (mean speed 46.1 mph) and on dual carriageways and motorways it falls 23.4 per cent (mean speed 63.1 mph). This road type effect confirms the hypothesis expressed earlier that the effect arises from differences in the shape of the speed distributions on these roads for some drivers; it is not immediately obvious that the effect is of any practical significance. In summary, it would appear that once the mileage and age categorisations have been made, distinguishing drivers by sex is no longer necessary. Accordingly, a basic stratification using only annual mileage (<10,000 and 10,000+ miles annually) and age (<44 and 45+ years of age) is proposed as a simple system of stratification which could be of some practical value to road designers. The other factors - notably, occupational group, vehicle ownership, and violation score - together with the relatively small and counter intuitive road factor SIND, can best be regarded as modifiers to the effects of the basic discriminators, age and annual mileage The proposed stratification system The hierarchical segmentation procedure (CHAID) has been used as described in the previous section to develop a simple driver stratification scheme. Dropping the distinction between the sexes and the term distinguishing between single and dual carriageways (SIND), the stratification tree of Figure 2 reduces to that shown in Figure 3 based only on binary splits of drivers according to annual mileage and age. Figure 3 shows for each leaf in the hierarchical tree the proportion of drivers in the fast and slow speed bands (band 1 and 5 respectively) and the number of drivers contributing to these proportions. The lower section of the leaves of the tree show average speeds, average ages, average annual mileages and mean accident frequencies for the drivers within each group defined by the stratification system. 2.7 Accident analysis Introduction In the questionnaire survey, drivers were asked to report the number of accidents of all kinds occurring on public roads in which they have been involved in the previous 3 years or for as long as they had been driving. Most of these accidents were damage only and only about 10 per cent involved injury. This section reports the relationship between these reported accidents and the demographic and psychological variables collected in the study, and determines how these accidents are related to drivers speed choice. In previous studies it has been shown that compared with the most recent year, substantial numbers of accidents in the previous year are forgotten (Maycock et al, 1991). Because of this progressive memory loss effect, and to avoid distortion in carrying out accident analyses, the accident frequencies presented in the following sections are based on only those drivers who have been driving for at least three years. 11