Traffic accident reduction by monitoring driver behaviour with in-car data recorders

Transcription

1 Accident Analysis and Prevention 32 (2000) Traffic accident reduction by monitoring driver behaviour with in-car data recorders Peter I.J. Wouters *, John M.J. Bos SWOV Institute for Road Safety Research, P.O. Box 1090, 2260 BB Leidschendam, Netherlands Received 1 September 1998; accepted 11 June 1999 Abstract People who are aware of being observed tend to modify their behaviour. This phenomenon could potentially be used to encourage individuals to behave more safely when driving if there were means of providing feedback to the drivers about their behaviour on the road. So-called vehicle data recorders offer such a means of providing behavioural feedback by confronting drivers with their recorded driving actions. A field trial of a matched experimental/control group design with intervention was planned with the objective of investigating whether this feedback mechanism would reduce the number of road traffic accidents in everyday driving conditions. In the study, seven experimental vehicle fleets were involved, which varied widely in terms of the kind of transport sector concerned, the type of vehicles used, and the traffic circumstances in which the vehicles are operated. Each fleet was matched with at least one control vehicle fleet, taking into account those relevant traffic safety characteristics. In total, accident and exposure data were collected for 840 vehicles of which 270 equipped with a recorder. During an observation period representing a total of about 3100 vehicle years, these vehicles were involved in 1836 road accidents. Analysis of the effects of the use of data recorders in these fleets resulted in an average estimated accident reduction of some 20%. The analysis shows that the actual savings vary depending on the transport sector involved and on the prior level of the fleet s safety record. Further studies are needed to identify the more promising application of such a use of traffic data recorders as a means of reducing road accidents Elsevier Science Ltd. All rights reserved. Keywords: Driver behaviour; Intervention; Feedback; Telematics; Field trial; Traffic safety 1. Introduction Human behaviour is a determining factor in traffic safety. It is of crucial importance therefore, that drivers are aware of the safety implications of their driving behaviour. It is known that people who are aware of being observed tend to modify their behaviour as a result, especially if important others can give feedback on their perceived behaviour (Ajzen, 1988). This phenomenon could potentially be used to encourage individuals to behave more safely in traffic by providing feedback on their driving. So-called vehicle data recorders using modern digital recording techniques, offer a means of providing behavioural feedback to drivers in terms of their * Corresponding author. Tel.: ; fax: address: peter.wouters@swov.nl (P.I.J. Wouters) recorded driving actions. The potential effect is twofold. Firstly, drivers knowing that they are being monitored may well adjust their behaviour as a result, and secondly, data recording will enable drivers to be confronted after the event with their recorded behaviour in specific driving situations. If this proves effective in changing driving behaviour, fewer (or less severe) road traffic accidents should ultimately result. In order to assess whether such feedback mechanisms would reduce the number (and/or severity) of road traffic accidents in real driving condition, a real-world practical experiment was carried out in fleets of commercial vehicles. The feedback process was not the subject of the study; it is its effect on accident occurrence, which was the prime subject. The first phase of this experiment was undertaken within the framework of SAMOVAR, a project which formed a part of the European Union Commission s DRIVE 2 research programme, and an additional /00/$ - see front matter 2000 Elsevier Science Ltd. All rights reserved. PII: S (99)

2 644 P.I.J. Wouters, J.M.J. Bos / Accident Analysis and Pre ention 32 (2000) phase was implemented to better meet the requirements of the trial design. The latter phase was made possible as a result of the co-operation of the Association of Dutch Insurers. The study is reported in greater detail elsewhere (Wouters and Bos, 1997). In the present paper, the design and execution of the experiment will be briefly described, the data gathered on accidents and exposure will be summarised, and the analysis of the data will be presented and discussed. 2. Method 2.1. General approach The experimental trial was aimed at assessing the effect of behavioural feedback using data recorders. In principle, feedback particularly to members of the general driving public could be given in a variety of ways, for example, by training organisations, the police, or insurance companies. In the present study, it was decided to focus on vehicles which were a part of a business fleet. Management of such fleets provides a natural means of feeding back information available from in-vehicle data recorders to the drivers. Collaboration of the fleet-owners in this project was on a voluntary basis and was implemented at their own expense. Fleet owners co-operating in the project agreed to give feedback to their drivers. The methods of providing feedback were not being specifically investigated in the study. For recording driving behaviour data, two types of commercial recorders, which were available from several different manufacturers, were used. One of these is the accident data recorder (ADR). As the name suggests, this recorder is designed to collect data just before and during an accident or incident for the purpose of facilitating the reconstruction of the event. The other type of recorder is the so called on-board computer or journey data recorder (JDR) and is primarily used for the purpose of fleet management. Data on time schedules, mean speed, rapid accelerations and decelerations, fuel consumption, etc., were recorded, as well as more detailed data on driving speeds over the last 90 seconds or so, before the vehicle stopped. Both types of recorded information are relevant to driving behaviour and safety, albeit of some different nature, and offering different opportunities for feedback as well. Since neither the exact information recorded, nor the specific way it was fed back to drivers, was a focus of study, recorders of both types could be used in the experiment Design of the field trial In assessing feedback effects on accident occurrence in a real-world setting, the experimental design needs to be relatively uncomplicated, avoiding practical constraints as far as possible. In view of this, a so-called matched experimental/control group design with intervention was preferred in which the numbers of traffic accidents among an experimental group and a control group of vehicles are compared for periods before and after the start of the intervention. The experimental group was a group of vehicles in which the data recorders were to be fitted and whose accident histories were collected during the periods before and after the intervention. The intervention itself consisted in the installation of recorders and briefing the drivers about the purpose of the monitoring and the arrangements for giving feedback. To enable changes in the general accident trends to be allowed for, accident information was collected for a control group of vehicles which was matched with the experimental group and which did not receive the intervention. In the context of this study an accident was defined according to the following international definition: a traffic accident is an event on public roads, which is related to traffic participation, in which at least one moving vehicle is involved and in which, as a result, one or more road users have been fatally injured, and/or injured, and/or in which material damage has been caused. The following aspects of the experimental design were regarded as being of particular importance. Drivers were explicitly informed as to whether a vehicle was equipped with a recorder and its purpose: the prime stimulus for adapting their behaviour. They were arbitrarily allotted to these vehicles by the fleet managers (in order to avoid selection bias). In practice, it resulted in fixed driver/vehicle combinations for some fleets, and in combinations per shift for others. Since the numbers of accidents among the experimental and control groups were to be compared, the two groups had to be comparable with respect to relevant traffic safety characteristics. The characteristics of particular relevance are: the type of vehicle, the nature of the transport business in which the vehicles are used, and the traffic conditions in which they operate. These characteristics are also closely related to the vehicle s exposure to traffic hazards. Ensuring comparability between the experimental group and the relevant control group (or control groups, since more than one control group was used in some cases) might be achieved by choosing the participants in the two groups either on a purely random basis or by deliberate attempts to match the groups. Because the vehicle fleets which were willing to co-operate in the experiment were quite diverse, the random process did

3 P.I.J. Wouters, J.M.J. Bos / Accident Analysis and Pre ention 32 (2000) not seem appropriate, and a matching procedure was used. The criteria used for matching were the safety characteristics mentioned above. It was not always possible to match on all relevant characteristics. For example, changes in the composition of a fleet over time or a fleet manager s attitude towards safety were difficult to match, so that the trial has the character of a so-called quasi-experiment (e.g. Cook and Campbell, 1979). Another important aspect of the experimental design relates to exposure to risk a measure which reflects the extent of the participants involvement in traffic and therefore the risk of an accident. In traffic safety research, mileage is often considered as the most appropriate exposure measure. In the present study however, vehicle mileage in the pre-test period could only be reliably determined for one (big) fleet. For this reason it was decided to use the number of months of vehicleuse as a measure of exposure. In the case of the single fleet for which both measures were available, the correlation coefficient between mileage and vehicle-months of use was This strongly suggests that these two exposure measures are interchangeable, provided, of course, that the fleet s vehicles are in similar use and that the comparisons are made over the same time periods. In order to ensure that accident data was recorded in a consistent way during both the pre- and the post-implementation periods, the relevant insurance companies were asked to record and supply the accident data. Insurance companies could also provide insights into the safety history of a particular fleet and were sometimes willing to anonymously supply data on other comparable fleets as well. The before and after periods were chosen as a rule to be 12 month periods in order to avoid seasonal variations in both accident occurrence and exposure. In order to estimate in advance the required sample size (the number of vehicles required for detecting the expected effect of installing data recorders), the annual traffic accident involvement rate of vehicles likely to take part in the study were required. During a feasibility study, data on the accident involvement rates for a number of fleets were collected. This exercise showed that large differences existed between the different fleets, and that involvement rates between 0.2 and 0.5 accidents per vehicle per year might reasonably be expected. By assuming that the fleets in the experimental and the control groups are about equal in size, and that vehicles in both fleets will cover the same mileage under similar safety conditions, a simple model for estimating the required sample size was devised. Two kinds of errors had to be considered: the type I or -error that of finding a positive effect where there is none in reality, and the type II or -error that of not finding Table 1 Minimum number of vehicles required in the trial to give =5% (or type I error) and =20% (or type II error) Annual accident involvement Effect of installing the recorders in the experimental vehicles 10% 15% 20% 25% a positive effect when such an effect actually exists. Clearly, an error of the first kind is particularly serious, and a one-sided test using the conventional value of 0.05 for seemed appropriate. A higher probability of making a type II error was acceptable, so that a value of 0.2 was considered suitable for. Assuming that vehicles involved in a trial of the kind envisaged in this study would be involved in an average of one accident in 2 years, it was calculated that if an accident reduction of some 15% is to be detected, then about 570 vehicles fitted with a recorder and an equal number of vehicles without a recorder would have to be included in a trial that lasted for a period of one year before and one year after installing the recorders (conform Schneiderman, see: Sachs, 1968). Indications suggested that an effect of this magnitude was to be expected, and the experiment had therefore to be designed for detecting such an effect. Table 1 shows the estimated sample sizes required for three levels of accident involvement and four potential accident reductions arising from the trial. A smaller experimental group than that shown in Table 1 will be sufficient if (a) the control group is larger, (b) the observation periods are extended (beyond a year), (c) the actual accident involvement rate is higher than estimated, (d) the influence of the recorder is bigger than mentioned, or if (e) lower levels of confidence are accepted. In practice, vehicles for the trial had to be drawn from different fleets of various sizes. This may well have produced some inhomogeneity in the sample, and variation in the intervention effect. It therefore required careful matching of experimental and control groups, and weighting of the results from different groups within the sample when estimating an overall effect, as described in Ch Data collection Data for the present study was collected from 11 different fleets. In total, more than 840 different vehicles were involved, of which 270 vehicles were equipped with a recorder. The vehicles formed seven sets. Within each set, the experimental group (vehicles fitted with a recorder in the post-intervention period) was matched

4 646 P.I.J. Wouters, J.M.J. Bos / Accident Analysis and Pre ention 32 (2000) Table 2 Accident and exposure data for the matched experimental and control groups for the pre-test and post-test periods studied a Fleet set Experimental or Sample Pre-test Post-test control? size Period Accidents Vehicle-months Period Accidents Vehicle-months (A) Heavy Experimental /92 1/ /94 12/ trucks Internal control 50 As above As above External control 105 As above /94 10/ (B) Medium Experimental 25 1/89 7/ /94 7/ and heavy trucks Internal control 28 As above As above Internal quasi-control 160 1/91 7/ As above External control 53 As above As above External quasi-control 23 As above As above (C) Coaches Experimental 25 7/93 2/ /95 2/ External control I 25 As above As above External control II 25 As above As above (D) Taxis and Experimental 54 8/93 3/ /95 3/ vans Internal quasi-control 66 As above As above (E) Company Experimental 23 1/94 12/ /94 1/ cars Internal control 21 As above As above (F) Coaches Experimental 23 1/93 1/ /94 7/ Internal control 9 As above As above (G) Taxis Experimental 10 11/93 5/ /94 4/ Internal control 5 As above 0 25 As above 4 60 a The number of vehicles cannot always be given exactly, because the sizes of fleets changed over time. This is particularly true for those fleets participating in the project over longer periods. The numbers of vehicle months however, are aggregated only over the period the vehicle was actually in use. with one or more control groups. Four of these control groups were external i.e. they consisted of vehicles outside the company operating the fleet; seven are internal i.e. they consisted of vehicles owned by the same company; two were internal quasi-control groups. A quasi-control group is one in which not all matching criteria could be applied. One of the external control groups was also a quasi-control group. These seven sets differed widely from each other in the type of transport sector in which they were involved, in the type of vehicles being used, and in the composition and management of the fleet to which the vehicles belonged. An overview of the experimental groups and the controls, the sample sizes, and the period over which accident and exposure data were available, is given in Table 2. The seven sets may be briefly characterised as follows: 1. Set A (Heavy trucks) consisted of vehicles which were owned by an international transport company located in Belgium, and both the experimental group and the internal control group consisted of heavy trucks and trucks with full- or semi-trailers. The vehicles of the experimental group were equipped with ADR s. The external control group consisted of vehicles of basically the same type and use, owned by a Dutch international transport company. 2. Set B (Medium and heavy trucks) was formed from two fleets whose business was the distribution of cattle food in different regions of The Netherlands. The experimental fleet in this set, was equipped with ADR s, and both the internal and external control groups were composed of heavy trucks. The internal quasi-control group consisted of vehicles of different types and sizes, mostly medium and heavy trucks, used mainly for general goods deliveries. The external quasi-control group consisted of medium weight lorries, also mainly used for general goods transport.

5 P.I.J. Wouters, J.M.J. Bos / Accident Analysis and Pre ention 32 (2000) Set C (Coaches) consisted of coaches belonging to three different fleets, owned by a Dutch touring operator. They were mostly employed for international travel and touring purposes. The experimental group was equipped with JDR s. Two external control groups consisting of coaches from another company were available for this set. Drivers in the second control group received special instructions with regard to safety regulations, working hours, speed, etc. 4. Set D (Taxis and vans) was drawn from a fleet which specialises in the private transport of individuals or small groups. The experimental group were cars drawn from taxis operating inside a medium sized Dutch town. These vehicles were equipped with ADR s. Another section of this fleet consists of vans used for the transport of small groups of people, and in particular for the transport of handicapped and elderly people, and for local ambulance services. These vans formed the internal quasi-control group for set D vehicles. 5. Set E (Company cars) were cars belonging to a Dutch insurance company participating in the study. The experimental group was equipped with ADR s. The vehicles in set E are mainly used for business purposes during working hours, but are used for private travel as well. 6. Set F (Coaches) consisted of vehicles drawn from a fleet owned by an international travel and touring company in Belgium. Its coaches are employed for the national and international passenger transport. The experimental group was equipped with ADR s. 7. Set G (Taxis) consisted of taxis operated in a large Belgian city. The cars in the experimental group were equipped with ADR s. 3. Data analysis and results 3.1. Analysis procedure To assess the effect of the use of data recorders on accident occurrence, the following procedure was adopted. First, the accident rates (i.e. accidents per vehicle month) for the pre- and post-test periods of each group were calculated, in order to take into account their exposure to traffic hazards. The ratio of the post- and pre-test rates, the so called risk ratio, gives the change in accident occurrence over time. The risk ratio for the control group reflects the autonomous trend in accident occurrence. In the case of the experimental group, this trend is combined with the intervention effect. In order to determine the intervention contribution to the compound effect, an adjusted risk ratio was calculated next, by dividing the risk ratio of the experimental group by that of its control group. If a set contains more than one control group, weighting is required, as will be described in Ch If the adjusted risk ratio is smaller than one, then a reduction in accident rates has occurred for vehicles equipped with a recorder; conversely, if the adjusted risk ratio is greater than one, then accident rates have increased. As can be seen in Table 2, the pre- and post-test periods of set A and set B do not completely coincide in time. This lack of matching of the observational periods in particular relevant to seasonal influences, could affect the estimated safety effect of the intervention. Therefore, in making estimates of the safety effects, we distinguish the following three approaches which make use of different sub-sets of experimental and control data: 1. Total. All data from both the experimental and the control groups were used. From the numerical point of view, the larger accident numbers yield estimates of the effect size with smaller confidence intervals. 2. Internal matching. For those analyses labelled internal matching in the following tables, all available data were used in so far as they coincided in time. Thus, from set A, only the internal control group data were applied. Data from the external control group were omitted. From set B, the data concerning the very beginning of the pre-test period of the experimental and the internal control groups were omitted. This approach places some emphasis on data from the experimental group, adjusted for the accident trends internal to the fleet operation. 3. External matching. For those analyses labelled external matching, all data of the external control groups were used. In order to achieve full time overlap, data of the experimental and internal control groups from set A were omitted at the very end of the post-test period. The selection of set B is the same as in the case of internal matching. This analysis places some emphasis on the adjustment of accident rates according to trends in the external control group. Comparison periods, other than those listed above, were not thought to be useful; nor were they practical due to the limited amount of accident data available in the study. Obviously, each of the outcomes generated by the three approaches outlined above offers an estimate of the effect of installing the data recorder (and the associated feedback), but each of them does so from a different point of view. Given the quasi-experimental character of the trial, it is not possible to be dogmatic about which of the three is to be preferred.

6 648 P.I.J. Wouters, J.M.J. Bos / Accident Analysis and Pre ention 32 (2000) Table 3 Estimated effects for each set of vehicle fleets, 90% confidence intervals and proportional contribution made by each set to an overall weighted average a Fleet set Adjusted risk ratio Confidence interval Contribution to overall result Set A 1.04 [0.79, 1.28] 57% Internal matching 0.69 [0.45, 0.92] {53%} External matching 1.13 [0.85, 1.40] {56%} Set B 0.54 [0.18, 0.90] 14% Internal=External matching 0.83 [0.26, 1.40] {11%} Set C 0.58 [0.04, 1.13] 6% Set D 1.54 [0.91, 2.16] 13% Set E 0.97 [0.08, 1.87] 4% Set F 0.28 [0, 0.68] 6% Set G 0% a The contribution percentages for the cases of matching of data periods are placed within brackets {}. Set A with internal matching matches time periods for experimental data with internal control data but makes no use of external control group data. Set A with external matching uses experimental and internal control data which matches external control data but makes no use of 14 months post-test data for both the experimental and the internal control groups. Set B with internal=external matching uses only those data periods during which experimental data, internal and external control data fully coincide, omitting two early years pre-test data of the experimental and of the internal control group Results Table 3 shows the adjusted risk ratios (the safety effect of feedback) and their confidence limits for the seven sets of fleets. It will be seen that the estimates vary e.g. from a non-significant increase in accident risk of 13% (Set A (heavy trucks) adjusted risk ratio 1.13) to a reduction of 72% (Set F (coaches) adjusted risk ratio 0.28). The estimates of effect size presented in Table 3, indicate the potential effectiveness of applying the feedback mechanism. Although relatively small sample sizes were involved, a statistically significant accident reduction is apparent for some of the fleet sets. This is indicated in the table by values of the adjusted risk ratio less than one and upper confidence limits also less than one. At the same time, the results given in Table 3 suggest a wide divergence in feedback effects among the fleets. Although sets A (heavy trucks) and D (taxis and vans) have risk ratios greater than one (implying a reduction in safety), neither of these adverse safety effects are statistically significant. In order to determine an overall result of the feedback effect on accident rates (all seven sets combined) a weighted average has to be computed because the fleet sets have different samples sizes; in addition, the various pairings of experimental and control groups need to be taken into account. In order to estimate a combined effect therefore, each set of fleets included in Table 3 is assumed to produce an independent estimate of a common safety effect. An overall weighted average has been computed using as weights the inverse of the variances of the estimates for the individual sets. The variances were calculated by assuming that the numbers of accidents included in each of the samples have a Poisson distribution (e.g. Hedges and Olkin, 1985). In this way, each estimate of the adjusted risk ratio shown in Table 3 contributes to the overall effect in proportion to its reliability, thus producing a minimum variance estimate of the common effect. The contribution of each estimate is shown in Table 3 as a percentage under the heading: Contribution to overall result. As can be seen in Table 3, set A contributes more than 50% to the overall effect. In order to present a more balanced picture the combined result for sets B to F has been estimated, and is given in Table 4; in this table the risk effect is expressed as a percentage change. The results for Sets B to F combined (Table 4) show an accident reduction of 39%. In this case, all available data were used. The reduction becomes 31% in the case of internal ( =external) matching. Both results are statistically significant at the 90% confidence level. The combined result for sets B to F is very similar to that obtained for set A when using the internal matching to adjust for general accident trends. The set A result with the external control group included, appears Table 4 Effect for set A (heavy trucks), for the combination of sets B to F, and for all sets of fleets, with 90% confidence intervals a Aggregated results Mean effect on risk Confidence interval Set A + 4% [ 21%, +28%] Internal matching 31% [ 55%, 8%] External matching +13% [ 15%, +40%] Sets B to F combined 39% [ 60%, 17%] Internal=external 31% [ 56%, 7%] matching All sets combined 20% [ 36%, 4%] Internal matching 31% [ 48%, 14%] External matching 12% [ 30%, + 7%] a Set G does not contribute to the overall effect. For other notes see Table 3.

7 P.I.J. Wouters, J.M.J. Bos / Accident Analysis and Pre ention 32 (2000) Fig. 1. The estimated overall safety effects in the fleets of the trial, and their 90% confidence intervals. (Not to scale). different from most of the other estimates. As the data in Table 2 shows, the positive safety effect found for set A (heavy trucks) in the case of internal matching, is largely due to an increase in risk arising from the internal control group, rather than a decrease in the risk of the experimental group. Thus, there would appear to be some inconsistency between the two results, which may be more to do with changes in the number of accidents in the control group than in the experimental group. Seen the calculated confidence intervals, such results may stem from statistical variations. The overall results in Table 4 are illustrated in Fig. 1. The figure shows that: 1. The overall accident rate in the experimental groups (those in which data recorders have been installed in the post-implementation period), when adjusted for changes in safety occurring in the control groups (which have not been subject to the intervention), is reduced by 20% after the intervention. This reduction exceeds the 95% significance level. 2. The overall result based on the internal matching of pre-test and post-test periods for the experimental and control groups, was that the intervention produced a reduction in accident risk of 31%-exceeding the 99% significance level. 3. When the intervention effect was estimated in the case of the external matching, the accident reduction was 12%. This result was not statistically significant; the probability of no real reduction being about 15%. 4. Discussion and conclusions The motivation for the research described in this paper was to improve traffic safety by changing the safety related behaviour of the drivers. The principle explored in this project is that monitoring a driver s behaviour offers a means for influencing behaviour, providing the driver knows that he/she is being monitored and receives feedback as a result of the monitoring. In this context, the objective of the field trial described in this paper was to determine whether drivers accident rate would be reduced if drivers were monitored using in-car data recording devices and were given feedback as a result. It should be emphasised that demonstrating such an effect on the accident rates of the vehicle fleets available to this study does not provide a comprehensive assessment of the potential accident reduction. The actual accident rates of the various types of vehicle fleets, as well as the exposure to risk in different transport sectors, are too varied for an overall measure of effectiveness. Moreover, neither the feedback actually provided by the various fleet managers, nor the subjective impact of the feedback on the driver, were investigated as a part of this study. Such factors will also contribute to variations in the effect in an unknown way. The trial was set up as a matched experimental/control group design with intervention. The sample size and other requirements of the experimental design were determined in advance as a part of a feasibility study of the preferred methodology. The voluntary co-operation of a number of companies had to be sought, and the greater part of the trial s funding raised. In order to meet scientific standards of validity and reliability for the trial not only was strict control of the process of data gathering necessary, but continuing contacts with the participating companies had to be maintained. The field trial was a major logistic operation, in which the wide geographical spread of the various participants was a considerable handicap. Eventually, the trial involved seven sets of fleets, which varied widely in type and characteristics. Accident and exposure data were collected for 840 vehicles, of which 270 were equipped with a recorder. During an observation period of about 3100 vehicle-years, these vehicles were involved in 1836 road accidents. In fact, the length of time over which data were collected, as well as the number of vehicles in the control groups, were rather greater than those envisaged when designing the trial. The actual average annual accident rate of the vehicles in the trial was also greater than expected at the outset of the study. An analysis of the traffic accident and exposure data obtained in this study, suggested that the use of data recorders and the feedback the drivers was given, resulted in a reduction in accident risk of some 20% on average. The safety effect of the intervention varied considerably between the different vehicle fleets included in the study. As well as a significant overall effect, statistically significant accident reductions were found for some of

8 650 P.I.J. Wouters, J.M.J. Bos / Accident Analysis and Pre ention 32 (2000) the fleets, though the estimated risk ratios generally show relatively wide confidence intervals. This lack of precision stems partly from the small sample sizes available for some fleets, but the influence of the fleet s prior accident record and the way feedback was applied might have played a role too. This being the case, it is desirable to extend the present study to investigate the effect of the use of in-vehicle recorders on a larger scale, and to examine their use in other than the commercial fleet settings used here. In this context, greater insight into the origins of the variations of the effect size among fleets of different types would be worthwhile, since such studies would offer opportunities for optimising the effect of driver monitoring. It would also be of particular interest to examine the effect of the different ways feedback to the drivers was achieved. Feedback could have been given, for example, by means of incentive schemes offered by the fleet management, or from regular or occasional feedback depending on the type of the recorder installed. It is also important to evaluate the feedback actually provided by the management or others, and to determine how this feedback is received by the drivers. The way feedback is given to drivers is also of importance for the long-term effectiveness of the measure. One method for reducing the possibility that the effect will not last, and to make safety an habitual aspect of driving, might be to provide feedback regularly. In this respect the journey data recorder seems to be preferable to the accident data recorder because, in the context of a commercial operation at least, it provides an operational motive for regular monitoring. References Ajzen, I., Attitudes, Personality and Behavior. Dorsey Press, USA, p. 112 et seq. Cook, T.D., Campbell, D.T., Quasi-Experimentation; Design & Analysis for Field Settings. Houghton Mifflin Company, Boston, USA, p. 103 et seq. Hedges, L.V., Olkin, I., Statistical Methods for Meta-Analysis. Academic press, San Diego, USA, p. 109 et seq. Sachs, L., Statistische Auswertungsmethoden, et seq. Springer Verlag, Berlin, Germany, p Wouters, P.I.J., Bos, J.M.J, The Impact of Driver Monitoring with Vehicle Data Recorders on Accident Occurrence, Report R-97-8, SWOV Institute for Road Safety Research, Leidschendam, The Netherlands.