At a glance
A Stellenbosch University engineering thesis found that not one completed accident report form in its sample was free of errors, and proposed a data quality framework to fix the leak at source.
Every completed accident report form Oluwatimilehin Okeowo examined contained at least one error. Some had empty boxes. Others contradicted themselves: a sketch showing a pedestrian struck down, while the accident-type box recorded a collision with a fixed object.
The forms came from the Stellenbosch Traffic Department and the local South African Police Service, the two bodies authorised to record crashes in the Western Cape town. Okeowo, then a master's student in industrial engineering at Stellenbosch University, worked through them field by field, checking each tick and handwritten note against the accident sketch and the officer's written description. His 2018 thesis describes a data pipeline that leaks at the exact point where a person first puts pen to paper.
That leak has a price. South Africa recorded 11,883 road deaths in 2023, and the Road Traffic Management Corporation (RTMC) puts the economic cost of crashes that year at R205 billion, or 2.74% of GDP. Of those deaths, 5,360 were pedestrians. Municipalities choose where to calm traffic, and provinces decide which corridors get rebuilt, from the patterns this data produces. Skewed input produces skewed spending.
Okeowo's study is small in geography and blunt in conclusion: the form itself needs restructuring, and the people filling it in need far better training. It is also one of the few pieces of research that opens up the actual paperwork rather than the published statistics.
Why South Africa's road safety budget rests on a handwritten form
The Accident Report form, usually shortened to ARF, is a multi-page document completed at the scene by a traffic officer or a police officer. It captures the date, the junction type, the light and weather, the drivers' ages, the vehicle manoeuvres, the injuries and a hand-drawn sketch. Carbon copies travel from the local department to provincial capture systems such as TRAFMAN and the Integrated Provincial Accident System, and from there into national reporting.
Okeowo built his investigation around what he calls a four-stage approach model. First, he mapped the process inside the Stellenbosch Traffic Department and interviewed the people running it. Second, he identified the vulnerable areas of the form and the factors behind the errors. Third, he assessed the raw records and circulated a questionnaire to officers who had completed ARFs, with approval from his supervisor, the university's ethics committee, the traffic department and the provincial police in Cape Town. Fourth, he analysed the results and used them to justify a quality framework.
The raw material was the set of accident reports filed in Stellenbosch in 2012, assessed against three classic data quality dimensions: accuracy, completeness and the timeliness of validation. The full thesis is held in SUNScholar, Stellenbosch University's institutional repository.
A town's crash file, read line by line
What Okeowo found in the paperwork was not random noise. The mistakes clustered. He sorts them into two families. Response errors are answers that contradict another answer on the same form: the driver recorded as performing a reverse or a parking manoeuvre in the box reserved for traffic control type, a sideswipe recorded but only one vehicle counted, a speed limit of 6 km/h written where a built-up road normally carries 60. Item non-response errors are simply blanks, and he found them in every section of the form.
Some entries were internally impossible. Daylight ticked alongside overcast visibility. A pedestrian collision recorded under accident type while the sketch and the written description showed two vehicles. A date and a day of the week that didn't belong to each other. Street names misspelled badly enough to make the location hard to recover later.
Okeowo traces the pattern to human inaccuracy along the processing line rather than to machinery, and attributes it to thin and inconsistent training, instructions that officers bypass, and slow turnaround on corrections. A data control officer he interviewed put part of it down to inexperience and lapses in concentration at the scene, where a shaken driver, a hit-and-run, or an incident reported hours after the fact all work against a clean record.
Where the form fails its own users
Roughly 17.6% of the officers surveyed said specific fields were hard to complete: accident location, driver particulars, road details and the information on passengers and pedestrians. The bigger complaint was the sketch. Many drawings in the file were, in Okeowo's assessment, rendered very poorly, and he argues this costs traffic engineers the most diagnostic part of the record. A good sketch shows how the vehicles met. A bad one shows nothing.
He also faults the form's layout. The current arrangement, he writes, was driven by how much information could be squeezed into the available space rather than by any relational logic that helps the officer move from one field to the next. Fields such as road type and junction type get skipped as a result.
None of this is exotic. Anyone who has filled in a badly designed form under time pressure knows the failure mode. The difference is that these particular boxes decide whether a dangerous intersection gets a traffic light.
Photo by Nqobile Vundla on Unsplash
Three questions you can ask of any number
Okeowo's framework starts from a simple idea: never treat a crash statistic as a single fact. Ask whether the value is right, whether everything that should be there is there, and whether it arrived while someone could still act on it. Accuracy, completeness and timeliness are separate tests, and a record can pass one while failing the others. A form with a spotless speed limit and a blank injury section is accurate on one field and useless on another.
His proposal places the check at the local level, before the carbon copy ever leaves for provincial capture. Reporting officers and their supervisors would work to a zero-tolerance standard on errors, with validation built into the handover between the traffic department and the police rather than bolted on afterwards. The framework is explicitly aimed at the gap between those two locally authorised departments, whose training differs and whose communication, in Okeowo's account, has been poor enough to undermine earlier quality efforts.
He frames the goal as cultural as much as procedural. A supervisor who signs off a form with a blank junction type is, in effect, approving a gap in the national record.
The patterns that survived the errors
Even a noisy dataset produced a recognisable picture of Stellenbosch's crashes in 2012. Collisions peaked in the dawn hour between 07:00 and 07:59 and again at dusk between 16:00 and 16:59, which lines up with the school and work runs. They were most frequent in winter, in 60 km/h zones, in built-up areas, on dual and single carriageways, and most often on Thursdays and Fridays.
The people involved skewed young and male. Drivers and cyclists under 40 dominated, and South Africans outnumbered foreign nationals. Light conditions were recorded as daylight for about 76% of crashes, night with street lighting for 16%, and unlit night for 6%. On injuries, roughly 1% of people involved were killed, 6% seriously injured, 16% slightly injured, and 77% unhurt.
Okeowo is careful with that last set of figures. Missing injury data, he writes, blunted the estimates for the slightly injured and uninjured groups, and the improvement he could make by folding in passenger and pedestrian details was limited. The distributions are indicative rather than exact, which is the whole point of the thesis. Bad forms don't produce no information. They produce information you can't quite trust.
What the study can't tell us
The obvious limitation is scale. This is one municipality, one year of records, and one researcher doing a visual review by hand. The questionnaire captures what officers say about their own work, which is not the same as observing them at a crash scene. And the thesis is a master's dissertation, not a peer-reviewed journal article, so the findings carry the weight of a careful local audit rather than a national survey.
A sceptical reader might also point out that the study can't fully separate the form from the person holding the pen. If training is inconsistent, a redesigned form may not help much; if the form is genuinely confusing, more training may only paper over it. Okeowo's own evidence cuts both ways, which is why his recommendations pair the two rather than picking one.
The national literature does rhyme with his local findings. A 2011 assessment of South Africa's national road collision database by Marian Sinclair documented missing values, duplication and underreporting across the system, and later reviews of provincial data in the Western Cape have found missing ages in a large share of records and gaps in geolocation precise enough to hamper planning. That supports Okeowo's argument that the problem starts at collection. It also means a Stellenbosch fix can't be assumed to travel to another municipality with different staff, systems and politics.
Six days, thirty days, and a blind spot
One structural issue sits above the form entirely. South Africa counts a road death as a road death only if the person dies within six days of the crash. The United States and France use a 30-day window, with follow-up to confirm the link. Okeowo argues the six-day rule should be extended, noting that families and insurers would benefit and that a longer window creates more opportunity to capture details about the victim.
His point is that policy choices, not just clerical ones, quietly remove cases from the record. A cyclist who dies on the tenth day never appears as a road fatality, so the intersection where it happened never accumulates the count that would justify rebuilding it. That is how a data problem becomes an infrastructure problem.
Photo by Boitshoko Morobeng on Unsplash
The forms are already changing
New blank Accident Report forms were being distributed across Western Cape municipalities around the time of the study, with carbon copy duplicates attached to reduce lost records and speed up processing. Okeowo welcomes the change and then says plainly that it doesn't resolve the errors committed while completing the form. Better paper is not better data.
The Stellenbosch Traffic Department engaged with the work directly. In a letter reproduced in the thesis, J. Waldis of Traffic Services confirmed in June 2016 that the document had been proofread with minor changes to structures and reporting lines, that the analysis was done well, and that the proposals had been discussed with the authorities responsible for road safety, who asked for the findings to be evaluated further.
That request is the thread to pull. The next decade of South African road safety will run on whether crash data gets captured digitally at the scene, with drop-downs that make a 6 km/h speed limit impossible to enter, or whether officers keep sketching collisions by hand on a form designed around available space. Okeowo's framework is a low-tech answer to a high-stakes question: who checks the box before it becomes a budget line.
- road-safety
- south-africa
- stellenbosch-university
- industrial-engineering
- municipal-governance

