Sub-Theme Road Safety (MIROS)

Road policing and safety engineering are resource-constrained. Pahang covers 36,000 km² with 11 districts and a road network dominated by long federal and state routes through low-density terrain. Enforcement patrols, speed cameras, and remedial engineering all have to be placed somewhere, and the decision of where and when determines how many crashes are prevented per ringgit spent.

MIROS has provided a crash dataset covering Pahang for 2017: 3,000 vehicle-level records drawn from 1,673 crash reports, with 42 fields spanning vehicle, driver, road, environment, and location attributes.

Before a dataset can support a decision, someone has to establish what it can and cannot be asked. That step is routinely skipped. An analyst trains a model, reports an accuracy figure, and the result looks authoritative whether or not the data ever contained the relationship being modelled. Crash data is particularly exposed to this, because records are often synthesised or de-identified to protect victims' privacy, a necessary safeguard that can break the links between fields while leaving each field individually realistic.

This dataset shows that pattern. The driver fatality rate is 43.1%, roughly twenty times the rate in real Malaysian crash statistics, and fatality is statistically unrelated to every established risk factor: helmet use, road surface, lighting, and speed limit all show no effect on whether a crash is fatal. A severity model trained on this data would return confident predictions resting on nothing.

What survives is the structure that does not depend on links between fields: where crashes happen, when they happen, and what kinds of vehicles are involved. The problem this project addresses is therefore how to build a road safety decision tool from the parts of a dataset that remain trustworthy, and how to make that tool state plainly what it does not know, so that no one mistakes an artefact of data processing for a finding about road safety.

1. To build a reusable checking routine for crash data that tests whether records agree with one another, whether the values they contain are physically possible, whether locations fall where they should, and whether outcomes are related to known risk factors at all.
2. To establish which fields in the dataset still carry usable information and which do not, and to publish that as a short profile other analysts can rely on.
3. To turn that model into a patrol schedule, naming the districts and times of day where enforcement would encounter the most crashes.
4. To show, as a deliberate control, that a model trained to predict whether a crash is fatal does no better than simply guessing the most common outcome every time, and to explain why that is the correct result rather than a failed attempt.

• A reusable checking tool that can be run on any crash dataset of similar structure.
• A short field-by-field profile stating which parts of this dataset can be relied on and which cannot.
• A tested model of how many crashes occur in each district at each time of day and day of the week.
• A ranked patrol schedule: the districts and time bands where a patrol hour would meet the most crashes.
• A documented result showing why predicting crash severity from this data cannot work.

• The dataset covers a single year (2017) and a single state, so no trend analysis is possible and seasonal effects cannot be separated from year-specific events.
• Crash counts reflect reported crashes, so district variation partly measures reporting and policing intensity rather than underlying risk.
• Without traffic volume data, exposure is approximated by population and labour force, which is imperfect for transit corridors such as Rompin that carry through-traffic disproportionate to their residents.