FATAL TRAIN ACCIDENTS ON BRITAIN S MAIN LINE RAILWAYS: END OF 2007 ANALYSIS

Transcription

1 FATAL TRAIN ACCIDENTS ON BRITAIN S MAIN LINE RAILWAYS: END OF 2007 ANALYSIS Andrew W Evans Lloyd s Register Educational Trust Professor of Transport Risk Management Imperial College London March 2008 Summary This paper updates the author s previous statistical analyses of fatal train accidents on running lines of the national railway system of Great Britain to the end of 2007, based on fatal accident data over the 41-year period 1967 to The year 2007 saw the first fatal train derailment since 2002, other than at level crossings, and there were also three fatal collisions between trains and road vehicles. That is a poorer performance than in 2006, but it is broadly in line with the estimated mean frequencies. In consequence, the trends in the estimated frequencies are changed only slightly. The paper estimates that the mean frequency of fatal collisions, derailments and overruns per year in 2007 was 0.53, and the mean number of fatalities per year in these accidents was 2.1. The paper also estimates that the mean frequency of fatal collisions between trains and road motor vehicles in 2007 was 2.7 per year, with 3.8 fatalities per year. The estimated means for collisions, derailment and overruns are slightly higher than the corresponding estimates made last year for 2006, and those for collisions between train and road vehicles are slightly lower. The paper examines the evolution of the estimates since 2001, and makes comparisons with the results of the Safety Risk Model of the Rail Safety and Standards Board. Keywords Railways, safety, accidents, fatalities, train protection, road vehicles. Centre for Transport Studies a.evans@imperial.ac.uk Department of Civil and Environmental Engineering Tel: Imperial College London Fax: London SW7 2AZ

2 CONTENTS Summary page cover 1 Introduction 3 2 Basic approach 3 3 Fatal train collisions, derailments and overruns Data Observed accidents and accident rates The Train protection and Warning System (TPWS) Estimates of trends in accidents per train-km Estimates of mean numbers of accidents per year in Accident consequences Estimates of mean numbers of fatalities per year in A simpler model of the trend 11 4 Collisions between trains and road motor vehicles Data Accident rates and trends Accident consequences Estimated mean numbers of accidents and fatalities per year in Summary of principal results 16 6 Evolution in the estimates over time and comparisons with safety risk model 16 Sources and acknowledgements 18 References 19 Appendix 1: detailed data tables 20 Appendix 2: statistical results for estimates of trends in accident rates 24 2

3 1 INTRODUCTION This paper updates the author s previous statistical analyses of fatal train accidents on running lines of the national railway system of Great Britain. The paper presents estimates of frequencies of fatal train accidents and fatalities based on accident data over the 41-year period from 1967 to Two kinds of fatal accidents are analysed: (1) train collisions, derailments and overruns on running lines; and (2) collisions between trains and road motor vehicles, both at level crossings and elsewhere. The basic methodology was presented in the Journal of the Royal Statistical Society (Evans 2000). The previous version of this paper included data to the end of 2006, and was circulated in January 2007 (Evans, 2007). This paper incorporates: (1) data on fatal train collisions, derailments and overruns for ; (2) data on fatal collisions between trains and road motor vehicles for ; (3) data on train-kilometres to the end of 2006, with a projection to The main statistical results presented are: (1) estimates of the mean frequencies of fatal train collisions, derailments and overruns in 2007, and of the trends in fatal accidents per train-kilometre; (2) estimates of the mean frequencies of fatal collisions between trains and road motor vehicles in 2007 and of their trends; and (3) estimates of mean fatalities per year in The year 2003 saw the completion of installation of the Train Protection and Warning System (TPWS), which was expected to reduce substantially the frequency of collisions, derailments and overruns. As last year, we make no independent estimate of the effectiveness of TPWS in this paper. We estimate pre-installation risks in this paper, and we combine these with industry estimates of the effectiveness of TPWS to estimate post-installation risks. The paper continues as follows. Section 2 summarises the basic approach to risk estimation adopted in this paper. Section 3 presents the data and analysis of train collisions, derailments and overruns. Section 4 presents the data and analysis of collisions between trains and road motor vehicles. Section 5 presents a brief summary of the main results. Section 6 compares the results of this paper with those of the Railway Safety and Standards Board s (RSSB) Safety Risk Model. Appendix 1 contains the detailed accident data; Appendix 2 gives statistical information about the fit of the accident frequency models to the data. 2 BASIC APPROACH The basic model of accident occurrence is that accidents, either all together or of a specified type, are presumed to occur randomly at a rate of λ per billion train-kilometres. Once an accident has occurred, the number of fatalities is also random, and has a probability distribution with mean μ. The mean number of fatalities per billion train-kilometres is then the product λμ. This quantity is the primary measure of the risk of accidental death. Either or both of λ and μ may change over time, and there may be sets of different λ s and μ s for different types of accident. 3

4 The parameters λ and μ are not directly observable, but they can be estimated. In this paper, the approach is to estimate them directly from data on past accidents. Estimating fatality risks λμ requires three steps: (1) Estimating mean accident rates λ, and trends in λ, from data on accident frequencies; (2) Estimating mean accident consequence μ, from data on accident consequences; and (3) Multiplying the estimates of λ and μ to give fatality risk λμ. These steps are followed separately for train collisions, derailments and overruns, and for collisions between trains and road vehicles. 3 FATAL TRAIN COLLISIONS, DERAILMENTS AND OVERRUNS 3.1 Data The year 2007 saw the fatal derailment of a London-Glasgow passenger train at Grayrigg in Cumbria on 23 February, with one fatality. That was the first fatal collision or derailment in Great Britain since the derailment at Potters Bar in There were also three fatal collisions between trains and road vehicles at level crossings, which are discussed further in section 4. Table A1 presents the updated list of all fatal main line collisions, derailments or overruns in , including the derailment at Grayrigg. There are 81 accidents in the dataset, with 321 fatalities. It may be noted that one of the accidents (at Milford in 1978) resulted in a collision between a train and a car; it is included in this dataset rather than the dataset of collisions between trains and road motor vehicles because it was due to a signal passed at danger (SPAD). However, the collision at Great Heck in 2001 between a train and an errant car from the M62 motorway, which then led to a collision between two trains, is counted with the collisions between trains and road vehicles. The derailment of a passenger train at Ufton Nervet level crossing in 2004 due to a collision with a car is likewise counted with the collisions between trains and road vehicles. Each accident in Table A1 is classified by whether it was preventable by Automatic Train Protection (ATP) or not. The ATP-preventable accidents are subdivided into those due to signals passed at danger (SPADs), excess speed, or buffer overruns. The SPAD accidents are subdivided into those involving a train passing a signal protecting a conflicting movement, and those involving trains proceeding in the same direction on the same track, a plain line SPAD. Of the 81 accidents, 15 were due to conflicting movement SPADs, 8 were due to plain line SPADs, 9 were due to ATP-preventable excess speeds or buffer overruns, and 49 were non-atppreventable. Each accident is also classified by whether a loaded or empty passenger train was involved or only non-passenger trains. The passenger trains are then further classified by the type of train and rolling stock: Mark 1 or post-mark 1, and multiple unit or locomotive hauled. Collisions involving both a multiple unit and locomotive hauled passenger stock are classified as multiple unit. The penultimate column of Table A1 gives data on the speed of the trains involved in accidents: in the case of collisions, this is the closing speed ; in the case of buffer overruns, it is the speed at which the train hit the buffers; in the case of derailments, it is the speed at which the train derailed. The speed data were assembled as part of an investigation for HMRI which was written up separately (Evans, 2001). 4

5 The consequences of fatal accidents are measured by the number of fatalities. It should be noted that, except where otherwise indicated, the data and discussion in this paper refer to fatalities rather than fatalities and weighted injuries (FWIs), which is used by the Railway Group to measure a combination of fatalities and injuries. The Railway Group s current weights are 1 for a fatality, 0.1 for a major injury, and for a minor injury. During 2000 the author estimated the ratio of FWI to fatalities in collisions and derailments, using HMRI cumulative figures of the numbers of fatalities, major injuries and minor injuries to passengers and staff in train accidents over the 18¼ years 1978 to 1995/96, which were 159, 550 and 5,774 respectively. When these figures are combined with the weights above, the ratio of FWI to fatalities is found to be Observed accidents and accident rates In this and previous analyses, fatal train accidents are subdivided into the following four categories: (a) ATP-preventable accidents due to conflicting movement SPADs; (b) ATP-preventable accidents due to plain line SPADs; (c) Other ATP-preventable accidents; and (d) Non-ATP-preventable accidents. Table 1 gives the numbers of fatal accidents in each of these four categories in seven five-year periods covering 1967 to 2001, and one six period The table also gives trainkilometres, and the calculated numbers of accidents per billion train kilometres for each category of accident. Table 1: train-kilometres; fatal train collisions, derailments and overruns; and accident rates: national railway system: Period Years Train- Number of accidents Accidents per billion train-km kilo- metres ATP-preventable Non- All ATP-preventable Non- All (billion) ATP- ATP- Conf. Plain Other prev. Conf. Plain Other prev. SPAD SPAD SPAD SPAD There are two main reasons for subdividing the accidents into the four groups. First, it has been desirable to separate ATP-preventable accidents from the rest, in order to inform the longrunning debate about ATP and TPWS as safety measures. Secondly, it was found empirically that the trend in the frequency of conflicting movement SPAD accidents was statistically 5

6 significantly different from the trends in the other types of ATP-preventable accident, and non- ATP-preventable accidents. These different trends can be seen in the data in Table 1. Plain line SPADs and other ATPpreventable accidents declined from appreciable rates per train-km in the late 1960s and early 1970s to none in later years, whereas conflicting movement SPAD accidents persisted until the period at a roughly constant rate of 1 fatal accident per billion train-km. At the historical level of railway activity of about 0.4 billion train-km per year, this was equivalent to an average of 2 accidents in 5 years. It may be seen from Table 1 that almost every five-year period in did indeed happen to have two fatal conflicting movement SPAD accidents. However, there were no fatal ATP-preventable accidents in At the time of writing, the most recent such accident was at Ladbroke Grove on 5 October The year 2003 saw the attainment of the longest continuous period without a fatal ATP-preventable accident since at least 1946, and this continued to the end of The Train Protection and Warning System It was largely in response to the persistent problem of conflicting movement SPADs that the Train Protection and Warning System (TPWS) was developed as an alternative full ATP to reduce the frequency of ATP-preventable accidents. TPWS was installed quickly and had been completed by the end of 2003 (Network Rail, 2003). It became effective largely during 2002 and TPWS appears to have made a step reduction in ATP-preventable risk. Before installation, it was estimated to be capable of reducing ATP-preventable accidents and casualties by about 70% (Davies, 2000). In early 2003, Railway Safety (2003) made another estimate of its effectiveness using the Safety Risk Model, and came to much the same conclusion. In this paper we use an estimate of 70%. Because TPWS became effective during 2002 and 2003, ATP-preventable accident data up to 2001 can be presumed to reflect the pre-tpws situation. The data for 2002 and 2003 reflect the transitional period while TPWS was becoming effective, and the absence of ATP-preventable accidents in those years may be in part due to TPWS. The years 2004 to 2007 had full TPWS. Therefore in order to estimate the pre-tpws ATP-preventable risk, we use the ATP-preventable accident data for , disregarding the data for , because these will have been affected by TPWS. We have at present no independent empirical estimate the post-tpws ATPpreventable risk, so we adopt the estimate referred to above that it reduces ATP-preventable accidents and casualties by 70%. 3.4 Estimates of trends in train accidents per billion train-kilometres Table 2 presents estimates of the trends in the mean numbers of fatal accidents per billion trainkilometres for each of four categories of accident separately, and also for all accidents taken together, assuming that the actual accidents in each category occur as a Poisson process with a mean which changes exponentially over time. As noted above, the pre-tpws trends in the three ATP-preventable categories are based on the data. The trend in non-atp-preventable accidents is based on the data. Table 2 shows that conflicting movement SPADs had 6

7 an almost constant rate per train-kilometre over the long-term, whereas the other three categories were all decreasing at statistically significant rates. The bottom line in Table 2 gives the trend in all accidents taken together over the period We discuss this in section 3.8. Table 2 Estimated trends in fatal train accidents per billion train-kilometres: Accident period over Estimated rate type which trend of change in estimated accidents per train-km (with standard error) Conflicting movement SPADs % (se 2.5%) p.a. Plain line SPADs % (se 4.7%) p.a. Other ATP-preventable % (se 5.4%) p.a. Non-ATP-preventable % (se 1.3%) p.a. All accident types together (se 1.1%) p.a. 14 Figure 1: Fatal train accidents per billion train-kilometres: Accidents per billion train-km Year 7

8 Figure 1 shows the fitted trend from 1967 to 2007 in all fatal accidents per billion trainkilometres, obtained by summing over the four different categories above, together with the eight observed data points from the last column of Table 1. The scatter of the data points about the trend is well within what would be expected of Poisson-distributed variables. Appendix 2 and Table A4 present the key statistical results and discussion of the goodness of fit of the adopted model to the data. The curve in Figure 1 bifurcates at the right-hand end. The upper branch shows the continuation of the pre-tpws trends estimated in Table 2. The lower branch shows the effect of TPWS. TPWS is assumed to have started to be effective in 2002 and to have been fully effective from the beginning of Thus the steep downward segment of the lower branch represents the introduction of TPWS during 2002 and 2003, and the flatter segment from 2004 represents the resumption of the previous trends in fatal ATP-preventable accidents, albeit at a lower level. It can be seen that the effect of TPWS is substantial relative to current fatal accident rates, though smaller relative to the long-term reduction in rates that has already been achieved. 3.5 Estimates of mean numbers of accidents per year in 2007 The upper half of Table 3 presents estimates of the mean numbers of accidents per year in The estimates are obtained by multiplying the mean numbers of accidents per train-km in 2007 from the trends above by the estimated number of train-km in 2007 (estimated at billion). Two sets of estimates are given. The first corresponds to the upper the upper branch of Figure 1, and hypothetically assumes that there was no TPWS in 2007; the second corresponds to the lower branch, and assumes that TPWS was fully operational in TPWS is assumed to have reduced the frequency of ATP-preventable accidents by 70%. Table 3 shows that the estimated the mean number of accidents in 2007 with TPWS was 0.53 per year, of which 0.36 per year were in non-atp-preventable accidents. This is slightly higher than last year s corresponding estimate: this is partly because the Grayrigg accident has slightly flattened the long-term downward trend in accidents per train-km, and partly because the number of trainkm is increasing. Table 3 Estimates of mean train accidents and fatalities per year in 2007 Without With Reduction TPWS TPWS due to TPWS Accidents Conflicting movement SPADs Plain line SPADs Other ATP-preventable Non-ATP-preventable Sum of all four types Fatalities Conflicting movement SPADs Plain line SPADs Other ATP-preventable Non-ATP-preventable Sum of all four types

9 3.6 Accident consequences As noted in section 3.1, the consequences of accidents are measured by the number of fatalities. Table 4 gives the distributions and averages of the numbers of fatalities in accidents by type of train, and, for accidents involving passenger stock, for multiple units, locomotive-hauled stock, Mark 1 stock, and post-mark 1 stock. Although the type of train, rolling stock and impact speed may be expected to affect the numbers of fatalities in accidents, and the writer has undertaken various analyses in the past using these factors, his present estimate of the mean fatalities in train collisions, derailments and overruns is simply the observed overall average number of fatalities per accident in the bottom right-hand corner of Table 4, which is currently There are two reasons for using this simple estimate. First, the observed mean numbers of fatalities per accident in the various sub-categories of passenger train accident in Table 4 are not statistically significantly different from each other. Secondly, there is no significant trend over time in the mean number of fatalities per accident. Figure 2 plots the fatalities in the individual accidents over time, and shows the trend line fitted by least squares. It can be seen that the slope is positive, but it is not statistically significantly different from zero, so we treat the mean as constant. Table 4 Fatalities in main line collisions, derailments and overruns: Number of accidents with given number of fatalities Total Fata- Type of - fata- lities/ accident All lities acc. Passenger stock Multiple unit Loco hauled Pre- & Mark Post-Mark All passenger Non-passenger All

10 Figure 2: Fatalities in train collisions, derailments and overruns: Hither Green Fatalities Clapham Junction Ladbroke Grove Year 3.7 Estimates of mean numbers of fatalities per year in 2007 The lower half of Table 3 (above) gives estimates of the mean numbers of fatalities per year in 2007, with and without TPWS. These estimates are simply the estimated frequencies in the top half of the table multiplied by 3.96, which is the mean number of fatalities per accident. The estimated mean number of fatalities in 2007 with TPWS is Again this is slightly higher than the estimate for The corresponding paper for the end of 2003 included the derivation of an estimate of the cost per fatality or equivalent fatality prevented by TPWS of 10.8 million. We do not repeat the calculation this year, because any new estimate by the same method would differ only slightly from the previous estimate. 10

11 3.8 A simpler model of the trend As an alternative to subdividing the accidents into the four categories discussed in Section 3.2, it is possible to fit a simple exponential trend to all accidents taken together over the whole period This also fits the data well. As was shown in Table 2, the single overall trend in accidents per billion train-km is a decline of 6.2% per year (with a standard error of 1.1% per year). Figure 3 shows this single exponential trend as the solid curve, together with the data points and the trend with the downward bifurcation due to TPWS from Figure 1 as the dotted curve for comparison. The two curves are not very different. The mean number of accidents per year in 2007 with the single exponential curve is 0.55, compared with 0.53 with the after-tpws branch from Figure 1. The disadvantage of the single curve is that it provides no detail about the different types of accident. On the other hand, it provides a simple description of the overall trend, and it makes no assumption about the effectiveness of TPWS. Figure 3: Fatal train accidents per billion train-kilometres: Accidents per billion train-km Year 11

12 4 COLLISIONS BETWEEN TRAINS AND ROAD MOTOR VEHICLES 4.1 Data Three fatal collisions between trains and road vehicles occurred in 2007, all at level crossings. These were at Delny near Invergordon on 2 February with two car occupant fatalities, at Gailes Crossing near Irvine on 23 February with one van occupant fatality, and at Swainsthorpe south of Norwich on 1 March with one car occupant fatality. That is a poorer performance than in 2006, but in line with the previous recent frequency of such collisions. As well as these additions to the data, two retrospective changes to the data have been made. These are because the analysis in this paper omits suicides, and inquests have determined that two previously-included level crossing fatalities were suicides. These were the car occupant in the collision at Ufton Nervet on 6 November 2004, and the car occupant at Swainsthorpe on 13 November 2005 (the same crossing but a different collision from that in 2007). In consequence, the Ufton Nervet collision remains in the data, but with 6 fatalities all train occupants rather than 7, and the 2005 Swainsthorpe collision is deleted from the data. This reduces the number of fatal collisions between trains and road vehicles in 2005 from 4 to 3. Including the changes above, there were 179 fatal main line accidental collisions between trains and road motor vehicles between 1967 and Of these, 165 occurred at level crossings and 14 elsewhere. Seven of the collisions at level crossings and 2 elsewhere resulted in fatalities to train occupants; in all others the fatalities were not train occupants. Appendix Table A2 lists the accidents in which there were train occupant fatalities and those not at level crossings, but only summarises the 158 level crossing accidents in which only road vehicle occupants were killed, because these are too numerous to list individually, and for many few details are recorded. Appendix Table A3 gives the number of accidents in each year, and the distribution of fatalities in these. The 179 accidents caused a total of 252 fatalities, of which 42 were train occupants and 210 were road vehicle occupants. The four most serious accidents all involved train occupant fatalities: they were at Hixon level crossing in 1968 with 11 fatalities, Lockington level crossing in 1986 with 9 fatalities, Great Heck in 2001 with 10 fatalities, and Ufton Nervet level crossing in 2004 with 6 fatalities. Table A2 shows that no collisions between trains and road vehicles not at level crossings were recorded between 1967 and Since such accidents have subsequently occurred fairly regularly, albeit infrequently, there must be doubt about whether the pre-1976 record is complete. It is possible, for example, that such accidents were recorded at that time as trespassers. For that reason, the number of such accidents in the period is treated in the statistical analysis as unknown. The consequences of fatal train/road vehicles collisions are again measured by the number of fatalities. However, the ratio of fatalities and weighted injuries (FWIs) to actual fatalities for occupants of road vehicles is lower than in train collisions and derailments, because the proportion of injuries that are fatal is higher. The most relevant RI data are casualties to third parties in train accidents, because these are mostly road vehicle occupants. The numbers of fatalities, major injuries and minor injuries over the 18¼ years 1978 to 1995/96 were 101, 88 and 301 respectively. When these figures are weighted with the Railway Group s usual weights, the ratio of FWIs to fatalities for road vehicle occupants is found to be Given that 83% of fatalities in train/road vehicle collisions are road vehicle occupants and 17% are train occupants, the overall ratio of equivalent fatalities to actual fatalities is

13 4.2 Accident rates and trends Table 5 presents the numbers of fatal collisions between trains and road vehicles at level crossings and not at level crossings, and the collision rates per billion train-kilometres for the eight periods. As noted in Section 4.1, the number of non-level crossing accidents in is treated as unknown in the analysis. Figure 4 plots the data and also shows the fitted trend, which falls at a common rate of 3.1% per year for accidents both at level crossings and not at level crossings. It should be noted that the vertical scale in Figure 4 is different from that in Figure 1. The estimated mean numbers of fatal collisions between trains and road vehicles in 2007 are 2.45 per year at level crossings and 0.26 per year at other locations. Appendix 2 and Table A5 present the key statistical results and discussion of the goodness of fit of the modelled trend to the data. In contrast to the train collisions and derailments above, the scatter of the data about the trend is somewhat greater than would be expected of Poisson variables, which suggests that the data do not have a smooth trend as assumed by the model. It can be seen from Table 5 and Figure 4 that there was an increase in the accident rate between the late 1970s and the early 1980s; this may have been associated with the replacement of manned level crossings by certain types of automatic ones. Table 5: Train-kilometres, fatal train/road vehicle collisions and accident rates: national rail system Period Years Train- Number of accidents Accidents per billion train-km kilo- metres At Not at All At Not at All (billion) LCs* LCs LCs LCs *LC = level crossing No fatal non-level crossing accidents are referred to in the RI reports for However, there must be some doubt about whether there really were none, or whether this type of accident was not recorded at the time. Therefore this figure is treated as unknown in the statistical analysis. 13

14 Figure 4: Fatal collisions between trains and road vehicles per billion train-km: Accidents per billion train-km Year Figure 5: Fatalities in collisions between trains and road vehicles: Hixon level crossing 10 Lockington level crossing Great Heck 8 Fatalities 6 Ufton Nervet level crossing Year 14

15 4.3 Accident consequences Table 6 gives the distributions and averages of the numbers of fatalities in fatal collisions between trains and road vehicles by type of location. Although the observed average number of fatalities per accidents differ between the types of location, the difference is not statistically significant. Table 6 Distributions of fatalities in main line collisions between trains and road vehicles: Number of accidents with given number of fatalities Total Fata- fata- lities/ All lities accident At level crossings Not at level crossings All Figure 5 plots the fatalities in individual accidents over time; it should be noted that most of the points for one-fatality accidents represent several accidents. The least squares trend line is shown; its slope is almost flat, and not statistically significantly different from zero. In the light of these results, the mean number of fatalities per accident in collisions between trains and road vehicles is taken to be constant at 1.41 at both types of location. This is materially lower than the mean for train collisions, derailments and overruns because of the different nature of the accidents. 4.4 Estimated mean numbers of accidents and fatalities per year in 2007 Table 7 summarises the trends and estimated mean numbers per year of fatal collisions between trains and road motor vehicles in 2007, and the mean numbers of fatalities in these. The overall frequency of accidents is estimated to be 2.71 per year causing 3.81 fatalities per year. Table 7 Fatal collisions between trains and road motor vehicles: Accident Estimated rate Accidents Fatalities Fatalities Location of change in per year per per year accidents per in 2007 accident in 2007 train-km (with standard error) At level crossings Not at level crossings All 3.1% (se 0.6%) p.a

16 5 SUMMARY OF PRINCIPAL RESULTS Table 8 presents a summary of the estimated numbers of fatal accidents and fatalities per year in 2007 from earlier sections of this paper. Fatalities and weighted injuries are added, using the ratios of FWIs to fatalities discussed previously. With TPWS installed, collisions between trains and road vehicles collectively present higher risks than train collisions, derailments and overruns. Table 8 Estimated mean numbers of fatal accidents and fatalities per year in 2007 Mean Mean Mean FWI* Mean accidents fatalities fatalities per FWI per year per per year fatality per year in 2007 accident in 2007 in 2007 Collisions, derailments & overruns Train/road motor vehicle collisions *FWI = fatalities and weighted injuries 6 EVOLUTION IN THE ESTIMATES OVER TIME AND COMPARISONS WITH THE SAFETY RISK MODEL The Rail Safety and Standards Board has developed a detailed risk model, entitled the Safety Risk Model (SRM), for the purpose of estimating all risks of accidental fatalities and injuries on the main line railway. It does this by modelling the precursors and the consequences of about 120 accidental events. It uses a wide range of data on accidents and precursors, as well as expert judgement. The first two versions of the SRM were published in 2001; Version 3 was published in February 2003 (Railway Safety 2003); Version 4 was published in February 2005 (RSSB, 2005). Version 5 was published in August 2006 (RSSB, 2006). Each version of the SRM uses data closing some months before its publication in order to allow time for its production. The SRM was reviewed by Bedford et al (2004); the review includes comment on the results of the SRM in relation to the type of analyses presented in this paper, which Bedford et al label the AEM (for Andrew Evans model); we adopt that label in this section. Among many other outputs, the SRM estimates the same risks as those in this paper, but it does so in a different way, by the detailed modelling indicated above. It is possible to compare the estimates of the mean numbers of fatalities per year in train collisions, derailments and overruns and in collisions between trains and road vehicles from the two sources. However, it is not possible to compare the mean numbers of fatal accidents per year or fatalities per fatal accident, because the SRM works in terms of all accidents, not just the fatal ones, and the AEM considers just fatal accidents. Table 9 presents the estimated mean numbers of fatalities per year in fatal train collisions, derailments and overruns for from successive versions of the present paper for the AEM and from successive issues of the SRM. The table shows two AEM estimates for 2003, one with and one without TPWS. Estimates before 2003 are without TPWS; estimates after

17 are with TPWS. Table 9 shows that both the AEM and the SRM have estimated sharply reducing mean fatalities since 2001, to which TPWS made a major though not the only contribution. However, the SRM has consistently estimated higher mean fatalities per year than the AEM, by factors of since 2002, albeit on a reducing base. Table 9: Estimated mean fatalities per year in train collisions, derailments and overruns: AEM and SRM: 2001 to 2007 Year With TPWS? AEM source AEM estimate SRM source SRM estimate Ratio SRM/AEM 2001 No End Issue No End Issue No End Yes End Yes End Issue Yes End Issue Yes End Yes End Table 10 shows the corresponding estimated mean fatalities per year in in collisions between trains and road vehicles from the AEM and the SRM. The table shows that both models agree that there has been only a slow reduction in these means in recent years, which is a reflection of recent safety performance. For these accidents the two models are in close agreement on the level of mean fatalities per year. Table 10: Estimated mean fatalities per year in collisions between trains and road vehicles: AEM and SRM: 2001 to 2007 Year AEM source AEM estimate SRM source SRM estimate Ratio SRM/AEM 2001 End Issue End Issue End End Issue End Issue End End Technical note. The sources of the SRM results above are Table A1 of the Risk Profile Bulletin Issues 2 and 3 (Railway Safety 2001, 2003) and Table A2 of the Risk Profile Bulletin Issues 4 and 5 (RSSB 2005, 2006 ). The SRM s hazardous events currently contributing to collisions, derailments and overruns are HET 01, HET 02, HET 03, HET 06, HET 26 (collisions), HET 09 (buffer stop overruns), HET 12 and HET 13 (derailments). The hazardous events contributing to collisions between trains and road vehicles are HET 04 (train colliding with object not derailing), HET 10 and HET 11 (level crossings). The reason for counting HET 04 here is that all the fatality risk from this type of accident is to road vehicle occupants that may be struck by trains not at level crossings. It may also be noted that the two models classify 17

18 Great Heck type accidents differently. These are the rare accidents in which trains are derailed by striking road vehicles not at level crossings. The present writer has classified these with train/road vehicle collisions; the SRM classifies them with derailments. The risk from such events is small (of the order of 0.1 fatalities per year, notwithstanding Great Heck). SOURCES AND ACKNOWLEDGEMENTS As noted in Section 3.1, the principal sources of information for this paper are the Railway Inspectorate (RI) annual reports and accident reports. The Inspectorate is now part of the Office of Rail Regulation. The published RI reports are supplemented by RI computer records from 1984 and by data from the RSSB and the RAIB. The author is grateful to all for their help in providing information, but is alone responsible for the contents of this paper. LIST OF ABBREVIATIONS AEM Andrew Evans model ATP Automatic train protection ECS Empty coaching stock FWI Fatalities and weighted injuries GLIM Generalised linear interactive model HMRI HM Railway Inspectorate HSE Health and Safety Executive LC Level crossing LH Locomotive-hauled MU Multiple unit NLC Not at level crossing ORR Office of Rail Regulation RAIB Rail Accident Investigation Branch RI Railway Inspectorate RSSB Rail Safety and Standards Board RTA Road traffic accident SPAD Signal passed at danger SRM Safety risk model TPWS Train Protection and Warning System REFERENCES Bedford, T, S French and J Quigley. A statistical review of the Safety Risk Model: WP1 Report. RSSB, June At: 0safety%20risk%20model%20(wp1).pdf Davies, D E N (2000). Automatic Train Protection for the railway network in Britain A study. Royal Academy of Engineering, London. Evans, A W (2000). Fatal train accidents on Britain s main line railways. Journal of the Royal Statistical Society Series A: Statistics in Society, 163(1)

19 Evans, A W (2001). Speed and rolling stock of trains in fatal accidents on Britain s mainline railways. HSE Contract Research Report 334/2001. HSE Books, Sudbury, Suffolk. Evans, A W (2007). Fatal train accidents on Britain s main line railways: end of 2006 analysis. Centre for Transport Studies, Imperial College London, January. Network Rail (2003). Step change in safety delivered on time and under budget. Press Release PR 03078, 29 December Rail Safety and Standards Board (2005). Profile of safety risk on the UK mainline railway. Issue 4, RSSB, London. Rail Safety and Standards Board (2006). Profile of safety risk on the GB mainline railway. Issue 5, RSSB, London. Railway Safety (2003). Risk Profile Bulletin: Profile of safety risk on the UK mainline railway. Issue 3, Railway Safety, London. 19

20 APPENDIX 1: DETAILED DATA TABLES Table A1 Fatal collisions, derailments and overruns: national railway system: Date Location Nature of accident ATP*- Rolling Speed Fatalpreventable? stock km/h ities Grayrigg Derailment No Post-Mk1 MU* Potters Bar Derailment No Post-Mk 1 MU Hatfield Derailment No Post-Mk1 LH* Ladbroke Grove Train collision, fire Yes: C-SPAD* Post-Mk 1 MU Southall Train collision Yes: C-SPAD Post-Mk 1 LH Watford Junction Train collision Yes: C-SPAD Post-Mk 1 MU Rickerscote Derailment, then collision No Non-passenger Ais Gill Derailment, then collision No Post-Mk 1 MU Cowden Train collision Yes: C-SPAD Mark 1 MU Branchton Derailment No Mark 1 MU Morpeth Train collision No Non-passenger Newton Train collision Yes: C-SPAD Mark 1 MU Cannon Street Buffer stop collision Yes: Overrun Mark 1 MU Stafford Train collision No Post-Mk 1 MU Holton Heath Train collision No Non-passenger Bellgrove Junction Train collision Yes: C-SPAD Mark 1 MU Purley Train collision Yes: C-SPAD Mark 1 MU Warrington Train collision No Non-passenger Clapham Junction Train collision No Mark 1 MU St Helens Derailment No Post-Mk 1 MU Glanrhyd Bridge Collapsed bridge; train fell No Mark 1 MU Colwich Train collision Yes: C-SPAD Post-Mk 1 LH Chinley Train collision No Mark 1 LH Eccles Train collision, fire Yes: P-SPAD Post-Mk 1 LH Longsight Train collision No Mark 1 MU Wembley Central Train collision Yes: C-SPAD Post-Mk 1 MU Polmont Derailment No Post-Mk 1 MU Wigan Train collision No Non-passenger Wrawby Junction Train collision No Mark 1 MU Elgin Derailment No Post-Mk 1 LH Linslade Derailment No Mark 1 LH Alvechurch Train collision No Mark 1 MU Seer Green Train collision No Mark 1 MU Ulleskelf Derailment No Post-Mk 1 LH Crewe Train collision No Non-passenger Appledore Derailment Yes: Excess speed Mark 1 MU Invergowrie Train collision Yes: P-SPAD Mark 1 LH Paisley Gilmour St Train collision Yes: C-SPAD Mark 1 MU Fratton Train collision No Mark 1 LH Milford LC* Train/car collision Yes: C-SPAD Mark 1 MU Hassocks-Brighton Train collision Yes: P-SPAD Mark 1 MU Farnley Junction Train collision No Mark 1 MU Newton-on-Ayr Train collision Yes: C-SPAD Non-passenger Unknown Worcester Tunnel Jc Train collision No Non-passenger Lunan Bay Train collision No Post-Mk 1 LH Corby Train collision No Non-passenger Unknown Carstairs Train collision No Non-passenger Nuneaton Derailment Yes: Excess speed Mark 1 LH Watford Junction Derailment, then collision No Post-Mk 1 LH *ATP = Automatic train protection; C-SPAD = Signal passed at danger protecting a conflicting movement; P-SPAD = Signal passed at danger protecting preceding train on same line; MU = multiple unit; LH = locomotive-hauled; LC = Level crossing. Continued 20

21 Table A1 (continued) Fatal collisions, derailments and overruns: national railway system: Date Location Nature of accident ATP- Rolling Speed Fatalpreventable? stock km/h ities Bridgwater Train collision Yes: P-SPAD Non-passenger Pollokshields E Jc Train collision Yes: C-SPAD Mark 1 MU West Ealing Derailment No Mark 1 LH Shields Junction Train collision, fire Yes: P-SPAD Mark 1 MU Kidsgrove Train collision Yes: P-SPAD Non-passenger Leicester Train collision No Non-passenger Eltham Well Hall Derailment Yes: Excess speed Mark 1 LH Nottingham Train collision Yes: C-SPAD Non-passenger Beattock Train collision No Non-passenger Tattenhall Jc Derailment No Mark 1 LH Cheadle Derailment, then collision No Non-passenger Finsbury Park Train collision No Mark 1 MU Sheerness Buffer stop collision Yes: Overrun Mark 1 MU Middlesbrough Train collision No Non-passenger Kirkstall Train collision No Mark 1 MU Guide bridge Derailment No Mark 1 MU Roade Junction Derailment, then collision No Post-Mk 1 MU Streatham Hill Train collision No Mark 1 MU Beattock Train collision No Mark 1 LH Morpeth Derailment Yes: Excess speed Mark 1 LH Monmore Green Train collision, fire Yes: C-SPAD Post-Mk 1 MU Ashchurch Derailment, then collision Yes: Excess speed Mark 1 LH Paddock Wd-Marden Train collision Yes: P-SPAD Mark 1 MU Castlecary Train collision No Mark 1 MU Hatfield Train collision Yes: Excess speed Non-passenger Peterborough North Train collision No Non-passenger Hither Green Derailment No Mark 1 MU Didcot Derailment Yes: Excess speed Mark 1 LH Copy Pit Train collision Yes: P-SPAD Non-passenger Thirsk Derailment, then collision No Mark 1 LH Connington South Derailment No Mark 1 LH Stechford Train collision No Mark 1 MU

22 Table A2 Fatal collisions between trains and road motor vehicles: national railway system: Date Location Nature of accident Fatalities Train Road All Occs Veh Occs Collisions at level crossings with one or more train occupant fatalities Ufton Nervet Passenger train/car collision Lockington Passenger train/van collision Naas Passenger train/lorry collision Chivers No 1 Passenger train/lorry collision in fog Shalmsford St ECS/lorry collision 1 1 2?.?.70 Chivers Decoy Passenger train/lorry collision in fog Hixon Passenger train/road transporter collision Total 7 accidents Collisions not at level crossings with one or more train occupant fatalities Great Heck Car from M62/pass/freight train collision Annan Lorry fell from bridge; hit by pass train Total 2 accidents Collisions at level crossings without train occupant fatalities (summary) Total 158 accidents Collisions not at level crossings without train occupant fatalities Copmanthorpe Car crashed thro fence at old LC; hit by train Nocton, Lincs Car crashed thro old bridge wall; hit by train Weeton Car crashed thro bridge approach; hit by ECS Burbage Wharf Car fell from bridge; hit by freight train Balcombe Tun Jc Car fell from M23; hit by pass train Stourbridge Car fell down bank; hit by train Four Ashes Car fell after RTA; hit by freight train, fire Dunhampstead Car fell from bridge; hit pass train North Wembley Car driven through fence; hit by ECS 0 2 2?.?.81 Not known Car crashed through fence; hit by train Walkeringham Car fell from bridge; hit by freight train Cleland Van driven thro fence; hit by freight train Total 12 accidents All fatal collisions between trains and road vehicles Total 179 accidents Source: Extracted from Railway Safety (HSE, annual) or HMRI fatal train accident database Abbreviations: RTA = road traffic accident; ECS = empty coaching stock; LC = level crossing. 22

23 Table A3 Numbers of fatal collisions between trains and road motor vehicles: Number of collisions with given number of fatalities Total Fata- fata- lities/ All lities accident

24 APPENDIX 2: STATISTICAL RESULTS FOR ESTIMATES OF TRENDS IN ACCIDENT RATES Table A4 gives the main statistical results and test statistics for variants of the models for estimating the trends in the accident rates for collisions, derailments and overruns; Table A5 gives corresponding results for collisions between trains and road vehicles. The scaled deviance in Tables A4 and A5 is the measure of goodness of fit of the model variants to the data; a large scaled deviance indicates a poor fit. If the data are indeed generated in the way presumed in the model (that is Poisson-distributed with a timedependent mean), the scaled deviance can be assumed to be approximately χ²-distributed with mean equal to the number of degrees of freedom. As noted in the section 3.3, the data used for the analysis of train collisions, derailments and overruns are for for the three classes of ATP-preventable accidents, and for non-atp-preventable accidents. The currently adopted variant of the model is (e). The conclusion that variant (e) fits statistically significantly better than variant (d) is based on the fact that the reduction of 4.7 in the scaled deviance in moving from (d) to (e) is larger than would be expected when tested against the χ² distribution with 1 degree of freedom. That is the justification for concluding that the two types of fatal SPAD accidents had significantly different trends. The fact that the scaled deviance for variant (e) is less than the number of degrees of freedom indicates a good fit between the model and the data. As noted in section 4.1, the data used for the analysis of collisions between trains and road vehicles are for for level crossing accidents and for non-level crossing accidents. The adopted variant of the model is (b), that is a common trend for level crossing and non-level crossing accidents. The reduction in the scaled deviance of 24.1 between variants (a) and (b) is strongly statistically significant when tested against the χ² distribution with 1 degree of freedom, but the reduction of 1.8 between variants (b) and (c) is not. It follows that the trend is certainly non-zero, but the difference between the trends for accidents at level crossings and that for accidents elsewhere could be due to chance. Thus we assume a common trend for both types of location. The residual scaled deviance of 15.5 in variant (b) with 12 degrees of freedom indicates that the fit of the model to the data is not very good. It appears that the long-term trend in the accident rates at level crossings has not been smoothly downward, as assumed by the model. 24