We've been ranking dangerous intersections wrong

Executive Summary
I analysed ten years of crash and volume data of signalised intersections in Melbourne, priced crashes using willingness-to-pay values and used an Empirical Bayes model to flag dangerous intersections.

Victoria’s road network has historically been, and continues to be, built for maximising capacity. We have widened urban arterials that divide neighbourhoods and an expansive network of roads stretching into regional areas; all at a cost of $714 per Australian resident per year.

There is a need to find and upgrade dangerous intersections. Since adopting the Safe System approach (pictured below), attention has been turned to improving the safety of our existing roads. However, we still have a long way to go to reach Vision Zero by 2050. There were 1314 fatalities on Australian roads in 2025, and the total social cost of crashes in Australia is $27.6 billion per year.

Previous reports such as RACV’s annual survey and AAMI’s recently published top 10 intersections look at data to find dangerous intersections, but I wanted to look at the relationship between crash volume and traffic volume. Additionally, RACV chooses not to use crash data to influence its ranking, while AAMI uses its own motor insurance claims database which makes the analysis irreproducible. Meanwhile, Transport Victoria does not publish any intersection rankings.

What makes an intersection dangerous?

To calculate the true impact of each signalised Melbourne intersection over the past ten years, I started with a simpler approach: ranking intersections by crash frequency. However, this ranking ignored crash severity and traffic volume. For example, these top 3 intersections may have the most (tied) recorded crashes but they also all recorded zero casualties and #2 and #3 experience extremely high traffic.

Similarly, ranking by number of serious injury crashes retains this bias. We can find the most deadly intersections, but these aren’t necessarily the most dangerous. Fourteen sites have two fatalities however the number of crashes range from just 5 to 38 at these sites.

Using a crash index metric does emphasize severity. For example, the Bureau of Infrastructure and Transport Research Economics (BITRE) report titled ‘Evaluation of the Black Spot Program’ weights fatalities and serious injuries at 9.5, minor injuries at 3.5, and property damage only at 1. (Page 59) A key limitation is that it only counts the most severe casualty in the crash, and uses subjective dimensionless weightings. Each of the intersections below have 50+ crashes with no fatalities:

I chose a cost-based approach, because expressing crash severity in a monetary value is consistent and comparable. Additionally, it allows for benefit-cost ratios (BCR) to be calculated, which is an important tools in advocacy, as BCR hurdles often implement a baseline filter of 1.0 to not be rejected, and greater than 1.0 when funds are relatively scarce. For example, the Australian Black Spot Program requires a BCR of at least 2 as well as 2-3 casualty crashes and an average of 0.13-0.2 casualty crashes per km over a 5-year span.

Putting a price on a crash

It’s uncomfortable to ask what a crash truly costs. There will never be standards for evaluating the value of a human life. However, it’s also infeasible to spend an infinite amount of money on one life.

Through surveys and economic research, we can estimate how much Australians collectively value avoiding a crash, which gives us a broad measure taking into account all of the individual costs from legal to medical related costs.

I looked at two cost-based approaches that differ in how they estimate the cost of a crash:

1. Hybrid Human Capital (HHC) using the Bureau of Infrastructure and Transport Research Economics (BITRE) 2022 report titled ‘Social Cost of Road Crashes’, which estimates the social cost of a fatality at $2.9 million, hospitalised injury at $241k, and non-hospitalised injury at $26k. (p. 4, 2022 dollars)
2. Willingness to Pay (WTP) using the Australian Transport Assessment and Planning (ATAP) 2024 report titled ‘Willingness-to-pay…Research report’, which estimates the cost of a fatality at $6.7 million, hospitalised injury at $650k, and non-hospitalised injury at $54k. (Table 6.13, 2024 dollars)

I favoured using WTP estimates because it captures the massive intangible cost of pain and suffering.

Which intersections have cost us the most?

I adapted the methodology from the approach Ng (2022) used for Adelaide. While Ng uses a 3-year period and a selected amount of intersections, this post used a 10-year period to increase the volume of crash data and vehicle volume data, and ranks every signalised intersection in a larger study area.

The intersections below have recorded the highest total social cost using the WTP estimate. Note that each one has at least one fatality, and #2 (Springvale Junction) and #3 have extremely high traffic volume.

Intersections with higher traffic volume are generally associated with an increased frequency of crashes (R²=0.32), so I normalised each intersection’s metric by the number of entering vehicles using the Victorian SCATS data which contains traffic volumes at all signalised intersections. I aggregated crashes labelled as intersection crashes within 50m of a SCATS site, which lets me normalise any metric per million entering vehicles (MEV).

Without filtering out the quietest intersections, the top three only contains intersections in the first percentile of volume. Therefore, I removed the quietest 20% of intersections and ranked the remainder by Cost per MEV:

#2 has high volumes and a high social cost from a large number of recorded crashes. In comparison, #1 is relatively quiet in terms of volume, while the total cost exceeds $12 million dollars, most of which is derived from a single fatality. This raises the question: are intersection like this truly dangerous or rather just unlucky?

What if an intersection just had a bad run of luck?

There’s value in estimating the crash frequency expected at similar intersections and comparing it to how many crashes were actually recorded. I adapted Ezra Hauer’s 2002 tutorial for Melbourne, but be warned: the methodology gets technical fast!

To do this, I fitted a negative binomial regression to the data to model a Safety Performance Function (SPF). The independent variables chosen were logarithmic MEV and geometry because they were both statistically significant.

SPF Equation: log(E[crashes]) = −0.2040 + 0.4240 × log(MEV) + 0.2849 × geometry

I then wanted to weight the recorded data and the expected crash frequency using the SPF equation, which the Empirical Bayes (EB) method allows me to do. I estimated each severity separately so that each site’s EB estimate reflects its own severity mix.

This method reduces ‘regression to the mean’ bias, which arises where society is often being too interested in the safety of select intersections because they recently had unusually many crashes. The EB method also increases precision as it allows older data to be used with less risk of bias.

I was then able to calculate the difference between the SPF and EB estimates to find a value called Potential for Safety Improvement (PSI), which is the number of crashes above expected were recorded. These intersections have the highest PSI value, with each recording significantly more crashes than expected from similar intersections.

On the other end, these sites recorded less crashes than expected. These intersections should also be studied to check if any installed safety features are working.

However, PSI is a count of crashes, and as we saw earlier, ignores crash severity. Additionally, both rankings are dominated by high-volume intersections because high traffic amplifies PSI values. So let’s convert them to WTP dollar values and normalise for traffic volume.

Which intersections do we need to fix?

We can now find the intersections with a higher social cost than expected. The following have the highest excess social cost (PSI WTP) and should be flagged for potential investment in safety:

Again, to see if these intersections are a symptom of high traffic volume, the ranking can be normalised per million entering vehicles (PSI WTP MEV). As the ranking is once again dominated by sites in the first percentile of volume, the quietest 20% of intersections were filtered out to achieve more meaningful results.

Because every signalised intersection in Melbourne was included in the analysis, I could map all 2000+ intersections. I’ve included a close-up map of inner Melbourne colour-coded by PSI WTP and sized by volume.

But will changes be cost-effective?

Using SIDRA software and the SCATS diagram, I could model changes at the intersection with the highest PSI WTP per MEV and complete a full benefit-cost analysis. I’ve left this outside this post’s scope due to time constraints, however this is the next step to take the flagged intersections and model their changes to advocate for recommendations to the Transport Accident Commission (TAC).

A limitation of this study is that SCATS only includes vehicle volumes. For example, the CBD intersection pictured above has huge pedestrian volumes too which can be partially attributed to the high number of crashes.

Improving cycling and pedestrian infrastructure should be a priority because cyclists and pedestrians are the most vulnerable road users and will remain over-represented in crash statistics. However, they remain under-funded. Australia records 38 fatalities and more than $2 billion in social costs for cyclists alone each year. Yet just 90 cents per person is spent on walking and cycling infrastructure.

Conclusion

Ranking intersections is harder than counting crashes. Raw crashes flag busy intersections, fatality counts flag unlucky ones, and subjective weightings can’t be actioned easily. By using willingness-to-pay, empirical bayes, and per-vehicle normalisation approaches, we get a ranking that identifies the truly underperforming intersections. However, the ‘most dangerous intersection’ label cannot be applied without first understanding what we value as a city.

Interested in working with data like this? Reach out on LinkedIn or by email.