Statistically is the UK railway network now less safe?

[TransportHub]

I'm sure there are cleverer statisticians on this forum than me. Can the increase in fatal train crashes in recent years be considered statically significant? Or tragically unfortunate. Purely from a statistics point of view is the railway less safe in the 2020s than the 2010s or 2000s?




A table showing the number of fatal train crashes from 1990 to 2026
A graph showing the same data

 

[Dieseldriver]

I'm sure there are cleverer statisticians on this forum than me. Can the increase in fatal train crashes in recent years be considered statically significant? Or tragically unfortunate. Purely from a statistics point of view is the railway less safe in the 2020s than the 2010s or 2000s?

View attachment 206819View attachment 206820


A table showing the number of fatal train crashes from 1990 to 2026
A graph showing the same data

Great Heck (2001) and Ufton Nervet (2004) are missing from that list

 

[Swedenorer]

There is not enough data here to make a judgement, as it limits the sample to fatal incidents since the terms safety and fatal are not synonymous.

You'd need to look at say all the incidents looked into by RAIB for each year and then somehow grade them for seriousness. For example, Norwood Junction will not qualify for your chart but has a number of factors that would put it to the top of a risk index as would the unbelievable confusion at Bookham.

Take another set of data - most would proabaly consider Heritage Lines as pretty safe, but I did an analysis some years ago and between 2004 and 2021 there was only one year when the RAIB did not issue a relevant report, bulletin or digest. This period included a recreation of Abermule on a narrow gauge line. Would I miss my annual trip to the Isle of Wight Steam Railway? No way.

I've been most concerned recently by the spate of overspeeding incidents but the only time I ever really worried about things was from 1995-2002 when the risks being run with infrastructure management were obvious.

Your chart which shows the three years between Sarum & Talerdig might indicate that lessons are not always learnt. But statistically significant? Not on this basis for me.

 

[mr_moo]

So the first question of course is to define "safe".

Using only fatal crashes paints a very limited picture. In the industry we use something called the "Fatality Weighted Index", which equates a certain number of smaller incidents to one larger, a certain number of larger to one major etc,.
There's some good info here: https://www.rssb.co.uk/about-rssb/k...nsights/risk-management-rssb-graphic-insights

 

[TransportHub]

Great Heck (2001) and Ufton Nervet (2004) are missing from that list

Ah good spot

== Doublepost prevention - post automatically merged: 28 Jun 2026 ==

So the first question of course is to define "safe".

Using only fatal crashes paints a very limited picture. In the industry we use something called the "Fatality Weighted Index", which equates a certain number of smaller incidents to one larger, a certain number of larger to one major etc,.
There's some good info here: https://www.rssb.co.uk/about-rssb/k...nsights/risk-management-rssb-graphic-insights

Very interesting thanks

 

[Russel]

I think, to make an accurate claim on weather it's less safe, you'll also have to add the total number of services ran, which will also have increased, for perspective.

 

[JG89]

Surely you would also need to include some data about the number of train movements in each given year too.

Edit - I see @Russel had the same idea as I was posting.

 

[bluenoxid]

Trying to work out if passenger miles would be a measure to consider

 

[zwk500]

What you also need to be asking is how many accidents were *prevented*, which is of course impossible to quantify properly.

Measuring crashes per passenger-km or per vehicle-km would be good guides though.

 

[Russel]

What you also need to be asking is how many accidents were *prevented*, which is of course impossible to quantify properly.

Measuring crashes per passenger-km or per vehicle-km would be good guides though.

Indeed.

It's clear that number of fatal accidents alone doesn't tell you much at all.

 

[GB]

Sailsbury wasn't fatal.

 

[mikeb42]

The first problem you've got is that the answer depends a lot on how you define "safety" and that isn't clear cut.

The thing that matters from my point of view, as a passenger, is:

I get on a train and go somewhere. What's the probability of stepping off at the other end having been neither killed nor seriously injured (KSI). Mileage is sort of neither here nor there.

There's been a large increase in journeys over the last 30 years. I'm not sure if these figures are really correct but a quick attempt dredges up

1995: 0.8bn
2005: 1.1bn
2015: 1.7bn
2025: 1.8bn

So you'd need to find out the KSI figures and see how those correlate. The K in particular and even the SI might be so low as to be statistically insignificant, though taking whole 10 year periods might do something useful.

On the basis of lots more journeys but no sign of a corresponding increase in dead and injured people it's probably a lot safer than 1995. The interesting one is 2015 to 2025 where it's either going to be too hard to tell (the recent Bedford crash is statistically meaningless in isolation) or a bit less safe.

On a different measure, one thing that seems to be getting widely overlooked is that as the number of train movements and passengers goes up we'd expect more crashes and injuries while "safety", depending on how we define it, is remaining constant. If the probability of each way bad stuff can happen is remaining constant but we're creating more opportunities per unit time for it to happen we can expect to see more of it happening.

One other avenue that might give insight would be to look at all-cause mortality and morbidity of train drivers over the long term. What that's very likely going to show is that by far the most dangerous part of their job is the health effects of shift work and sedentary employment. After that, way out in front as the most dangerous part of most of their working lives will be a road journey before and after their shift. Actual train driving will be way down the list as it's probably not much less safe when done conscientiously than sitting in front of the TV at home.

As someone who fiddles with probabilities for a living quite a lot of the time (in an entirely unrelated field) the risk of train travel in the UK remains at the "not worth worrying about" level. I'm vastly more likely to get run over or drop dead of a heart attack while walking to the station.

 

[zwk500]

Indeed.

It's clear that number of fatal accidents alone doesn't tell you much at all.

Quite.

For the OP, the most useful set of statistics are probably ORR tables 5260: https://dataportal.orr.gov.uk/stati...es-and-number-of-incidents-split-by-severity/
1253: https://dataportal.orr.gov.uk/stati...253-passenger-vehicle-kilometres-by-operator/
And table 1333: https://dataportal.orr.gov.uk/stati...le-1333-freight-train-kilometres-by-operator/

Note 1253 is vehicle-KM not train-km so would need to be divided by operator's formation sizes.

 

[Djgr]

So the first question of course is to define "safe".

Using only fatal crashes paints a very limited picture. In the industry we use something called the "Fatality Weighted Index", which equates a certain number of smaller incidents to one larger, a certain number of larger to one major etc,.
There's some good info here: https://www.rssb.co.uk/about-rssb/k...nsights/risk-management-rssb-graphic-insights

Indeed, even if unsafe can be defined as (=) fatal crashes, we can't infer any trends from this data because the number of relevant incidents is so small.

 

[The exile]

Ultimately you can prove almost everything you want to with statistics, especially when the actual figures are very low (as they thankfully are in the case of UK railway accidents). Statistically, the railway was considerably less safe in 2026 five minutes after Bedford than it was 5 minutes before. In reality, there was no change (or if anything an infinitesimal improvement).

 

[swt_passenger]

I'm sure there are cleverer statisticians on this forum than me. Can the increase in fatal train crashes in recent years be considered statically significant? Or tragically unfortunate. Purely from a statistics point of view is the railway less safe in the 2020s than the 2010s or 2000s?

Did you mean to write “statistically” in the title and in the text of the first post? Attacked by autocorrect yet again?

 

[Oxfordblues]

I remember reading somewhere (it might have been Alan Williams in "Modern Railways") that the main risk involved in rail travel is in getting to the station in the first place.

 

[mikeb42]

Ultimately you can prove almost everything you want to with statistics, especially when the actual figures are very low (as they thankfully are in the case of UK railway accidents). Statistically, the railway was considerably less safe in 2026 five minutes after Bedford than it was 5 minutes before. In reality, there was no change (or if anything an infinitesimal improvement).

You can only "prove" anything you want with statistics to people who don't understand statistics.

E.g. the railway was only statistically less safe after the Bedford incident at more-or-less the 0% confidence interval, and the first question any statistically competent analyst would be asking about the data is "What's the confidence interval?"

For general consumption I do agree though.

Someone, somewhere, will be working out the answer to the question posed by the OP. The raw data is (thankfully) so limited that the statistically competent answer may actually be "It is impossible to say with a reasonable degree of confidence".

 

[norbitonflyer]

Statistically, Bedford made no difference to the fatality rate of railway passengers.
And, notwithstanding the fatality at Bedford, drivers are statistically safer than track workers.
And more passengers have accidents on platforms than on trains - including me.

 

[TransportHub]

Quite.

For the OP, the most useful set of statistics are probably ORR tables 5260: https://dataportal.orr.gov.uk/stati...es-and-number-of-incidents-split-by-severity/
1253: https://dataportal.orr.gov.uk/stati...253-passenger-vehicle-kilometres-by-operator/
And table 1333: https://dataportal.orr.gov.uk/stati...le-1333-freight-train-kilometres-by-operator/

Note 1253 is vehicle-KM not train-km so would need to be divided by operator's formation sizes.

Very interesting reading, thanks for posting

== Doublepost prevention - post automatically merged: 28 Jun 2026 ==

Someone, somewhere, will be working out the answer to the question posed by the OP. The raw data is (thankfully) so limited that the statistically competent answer may actually be "It is impossible to say with a reasonable degree of confidence".

This is exactly what prompted my original question, I remember we did 'statistical significance' at school but that was a while ago and the methodology escapes me now

== Doublepost prevention - post automatically merged: 28 Jun 2026 ==

Did you mean to write “statistically” in the title and in the text of the first post? Attacked by autocorrect yet again?

Auto correct and a mix of dyslexia!

 

[mikeb42]

Very interesting reading, thanks for posting

== Doublepost prevention - post automatically merged: 28 Jun 2026 ==


This is exactly what prompted my original question, I remember we did 'statistical significance' at school but that was a while ago and the methodology escapes me now

== Doublepost prevention - post automatically merged: 28 Jun 2026 ==


Auto correct and a mix of dyslexia!

Depending on how advanced they went in your school you may have been taught about things like the standard error in the mean and t-tests and the like. These are only meaningful subject to a whole lot of assumptions (the underlying probability distribution of the data, stationarity etc) that likely don't apply or can't be established here.

Intuition isn't necessarily a bad guide in the first instance and you only have to look at your graph to intuit that it's all a bit of a "don't know". The falling trend on the left is probably real. What is happening on the right looks suspicious but could also be a fluke. The underlying position has likely changed from left to right, but there's not enough data to ascertain what a new equilibrium might be nor if it's being deviated from latterly.

Fatal accidents is the wrong end of the severity spectrum to look at. Get the data for near-misses/concerning events and you might be in a very different position as there are a lot more of them. Anecdotally these have been going up significantly. If this is actually substantiated by the data,  and there's an apparent upward trend in fatal accidents (which you can bet are positively correlated with near misses in general) that adds up to concerning evidence.

 

[Lucan]

The thing that matters from my point of view, as a passenger, is:

I get on a train and go somewhere. What's the probability of stepping off at the other end having been neither killed nor seriously injured (KSI). Mileage is sort of neither here nor there.

That is an interesting point of view. However, with car, bike, foot, and train journeys the probability of an accident is more-or-less proportional to the length of the journey, all other things being equal. On the other hand, in the cases of planes and space travel the danger is almost entirely in the take-off and landing phases and the fact that those journeys tend to be long is the reason why aircraft do so well in deaths/km comparisons. In fact even space travel does rather well in deaths/km, not much worse than a car, because most trips cover huge distances, despite the fact that the chance of not surviving any one space trip is around a sobering one in a few hundred. For plane and space travel the risk per journey can be a more appropriate figure to consider than risk per km.

But shorter plane journeys, even professionally operated ones like private jets, have a much worse death/km figure, and owner-piloted light aircraft, which tend to make even shorter journeys, are even worse for numerous reasons. The following chart is somewhat out of date, and is world rather than UK data, and I don't believe that "vans" can be that safe, but it illustrates how figures can be misunderstood or manipulated. The three tables are the same data but arranged in order (safest at the bottom) by journeys, hours and km. Buses always do well for safety, but I believe that is because it is difficult to get killed in a London bus for example doing under 20 mph most of the time.

better than

 

[bramling]

I'm sure there are cleverer statisticians on this forum than me. Can the increase in fatal train crashes in recent years be considered statically significant? Or tragically unfortunate. Purely from a statistics point of view is the railway less safe in the 2020s than the 2010s or 2000s?

View attachment 206819View attachment 206820


A table showing the number of fatal train crashes from 1990 to 2026
A graph showing the same data

It’s so difficult to quantify, as (thankfully) the number of incidents in the entire time is so small that one incident will skew the results massively. And whilst there was a long period devoid of no serious incidents, there were certainly a few near misses during that time. There’s also the matter of whether one includes the Croydon tram derailment, which depending on how one counts could skew figures heavily due to the fairly high number of fatalities.

I think the message one can take is that British railways are extremely safe, clearly the industry was doing something very right to go through such a long period with no serious consequence accidents, but there’s definitely no room for complacency.

 

[Magrar]

Well on the one hand, the number of fatal crashes is tiny compared to overall number of journeys, so it's tempting to make a backround noise type argument.

Then again, if you stripped the data of all context and merely asked a statistician, is this data just random noise, given the time axis is statistically significant, they'd clearly have to say no.

 

[Deepgreen]

For fatal incidents, given the extremely low numbers involved, there is no significant trend that can be identified. The range is from 0 to 2 per year, which is too low to present any useful satistics. However, the gap from 2008 to 2020 in your table suggests that, if anything, that period is 'safer'. The fact remains, though, that the numbers are so tiny, that despite variations in journey numbers occurring from year to year throughout your chosen time period, the chances of dying in a rail accident are vanishingly low.

 

[Russel]

For fatal incidents, given the extremely low numbers involved, there is no significant trend that can be identified. The range is from 0 to 2 per year, which is too low to present any useful satistics. However, the gap from 2008 to 2020 in your table suggests that, if anything, that period is 'safer'. The fact remains, though, that the numbers are so tiny, that despite variations in journey numbers occurring from year to year throughout your chosen time period, the chances of dying in a rail accident are vanishingly low.

Regarding the last sentence, I'd say the chances of being in any kind of rail accident are vanishingly low.

 

[corsaVXR]

Some considerations:

• Regression to the mean can mean that if the actual average rate of fatal incidents is 0.2 a year, and you go a ~13 year period without incident, then statistically, you may see an increase without the actual mean being affected (can work the other way too) - this would assume there's no underlying changes to the mechanic at work. There was clearly a big shift around 2002, around the end of Railtrack.
• Your incidents are ranked as just fatal / non-fatal, whereas the reality is that we've had 4 fatal incidents since 2005, of which 1 had more than 1 fatality (Stonehaven)
• when doing health and safety training at work, we were shown an analogy to earthquake predictions, which is that for a given higher severity of incidents, a larger number of smaller severity incidents would have happened, for which each would have a corresponding larger number of even smaller severity incidents - to really understand trends, you'd want to have a full picture of all incidents with their severities, and understand the trends of each severity.

 

[bramling]

Regarding the last sentence, I'd say the chances of being in any kind of rail accident are vanishingly low.

It feels a little close-to-home for me knowing (albeit fairly loosely) not one but two separate people who were on the 360 last week.

 

[TheoBald]

I'll start by echoing the above comments on the reliability of statistics drawn from such small samples, and the need to ask whether we're comparing comparable data (passenger kilometres etc). But you prompted me to answer a version of your question for my own interest: "Can the increase in fatal train crashes in recent years be considered statically significant" - compared to the noughties. I'll share my answer, for your interest only, but with all the above provisos (and my fingers crossed!).

Firstly "significance": a standard value is to answer "what is the likelihood of this data being a purely random occurrence, happening by chance?" by requiring it to be 5% or less. In medicine 1%, 0.1% or even smaller is commonly required.

We have had a 6.5 year period in the 20s (to the end of this month).
In the noughties there were 3 fatal crashes in 10 years - averaging 0.3 per year = 1.95 per 6.5 year period.

We can model the number of fatal accidents per 6.5 years (from the noughties) as a Poisson distribution; Po(1.95). We then need to determine the number of accidents that we might expect with a confidence of 95%:

P(0) = 0.1423 .......... means probability of no accidents = 14.23%
P(1) = 0.2774
P(2) = 0.2705
P(3) = 0.1758
......... 0.8660 ........... probability of 3 or fewer = 86.6%
P(4) = 0.0857
......... 0.9517 .......... likelihood of 4 or fewer just over 95%

This shows there was an 86.6% chance of there being 3 accidents or fewer. The data would only be significant if there were 5 or more.

Ducks for cover!

 

[Russel]

It feels a little close-to-home for me knowing (albeit fairly loosely) not one but two separate people who were on the 360 last week.

Not a nice situation, but prior to this, how many people do you know that have been involved in a rail accident?