The Road Death Gap

Executive Summary. Bangladesh's official road-safety record and the World Health Organization's modeled burden must not be merged into one series. The WHO country profile lists 5,084 reported fatalities in 2021. Separately, WHO estimates 31,578 road traffic fatalities, with a modeled confidence interval from 27,441 to 35,716, and an estimated mortality rate of 18.6 per 100,000 people. The reported total measures deaths captured by national systems. The WHO estimate uses a mortality model to correct for incomplete capture and comparability. The gap cannot tell us which individual deaths were missed, but it is too large to treat as a footnote. Bangladesh needs a linked road-injury surveillance system spanning police, hospitals, emergency care, and civil registration, with one case definition and privacy-preserving record linkage. Better enforcement matters, but the state cannot target roads, vehicles, users, or post-crash care confidently while the denominator remains uncertain.

The reported record is one dataset

The WHO profile presents 5,084 reported road fatalities for 2021. This is the nationally reported count cleared through the country data process. It belongs on its own chart because it describes what the national reporting chain captured.

Source: World Health Organization, Bangladesh road-safety country profile, reported fatalities for 2021.

A reported death count is built through institutions. A crash occurs. Police may attend or receive a report. A victim may die at the scene, in transport, in a facility, or later. A death certificate may identify the external cause. Each transition creates a chance for loss, duplication, or inconsistent classification. A person who dies after leaving the crash scene can disappear from a police total even when the death is registered elsewhere.

The count should therefore be read as a property of the reporting system as well as a measure of harm. A lower reported total can mean safer roads, weaker capture, a narrower time definition, or some combination. Without linkage across police, health, and civil-registration records, policymakers cannot separate those explanations.

This is not an argument to distrust national officers. It is an argument to give them a system capable of following a case. Front-line police and clinicians collect information for immediate purposes, not for a later national merge. Common identifiers, dates, locations, and case definitions have to be designed into the workflow.

The time window after a crash is especially important. A scene record can classify a seriously injured person as surviving because the immediate task is transport. A hospital may later record a death without the crash location or police case number. Civil registration may capture the death but code the underlying cause differently. A linked registry preserves each institution's operational record while creating a verified outcome. It should never force a police officer to predict a later medical result.

WHO's estimate is a different dataset

WHO separately estimated 31,578 fatalities for 2021. Its modeled confidence interval runs from 27,441 to 35,716. The central estimate corresponds to 18.6 deaths per 100,000 people. These are model outputs, not additional reported cases.

Source: World Health Organization, Bangladesh road-safety country profile, modeled fatality estimate and confidence interval for 2021.

Keeping the estimate separate preserves honesty. It would be wrong to splice the estimated total into a police time series and describe the jump as a sudden deterioration. It would also be wrong to subtract the reported total from the estimate and label every difference an identified unreported death. The model expresses a likely burden under WHO's method and uncertainty, not a case list.

The estimate is still policy-relevant. Its scale says the national system may be observing only a fraction of fatal harm. The confidence interval says the exact level remains uncertain even after modeling. Both points strengthen the case for direct measurement.

Uncertainty should shape language and budgeting. A modeled central estimate can support the conclusion that road injury is a large public-health burden. It cannot support a precise claim about one district, road class, or user group unless the underlying local data do. Prevention funds should therefore combine broad safe-system investments with targeted projects drawn from case-level evidence. The model sets urgency. The registry sets location.

World Bank WDI provides a separate modeled rate series, with 10.8 deaths per 100,000 people in 2000, 17.6 in 2005, 16.9 in 2010, 15.8 in 2015, and 15.3 in 2019. That series should also remain labelled as modeled and should not be used to claim that the nationally reported count followed the same path. A trend can inform broad comparison while still failing to locate crashes or identify the people missing from administrative records.

The fleet shows where prevention must look

The WHO profile reports 5,013,908 registered vehicles in 2021. Powered two- and three-wheelers accounted for 3,822,446, compared with 767,797 four-wheel vehicles, 195,094 heavy trucks, and 76,903 buses.

Source: World Health Organization, Bangladesh road-safety country profile, national vehicle registry categories for 2021.

Fleet composition does not equal fatality composition. Exposure depends on kilometres travelled, road type, speed, passenger load, helmet and restraint use, vehicle condition, and conflict with pedestrians. The registry also describes registered vehicles, not every vehicle in use. Yet the dominance of powered two- and three-wheelers makes their safety environment impossible to treat as a side issue.

Safe roads must manage kinetic energy, not merely driver morality. Median barriers, protected walking space, safe crossings, forgiving roadsides, speed management, vehicle standards, and emergency response reduce harm even when a person makes a mistake. Enforcement is part of the system, but a strategy that relies on perfect behaviour will fail predictably.

Data design should support this systems view. A crash record needs a precise location, road characteristics, vehicle types, road-user roles, injury severity, and outcome after follow-up. Hospital records need an external-cause field that can link back without exposing the patient publicly. Road authorities need a routine method for turning clusters into engineering changes and publishing whether the change worked.

Exposure data complete that method. A corridor with many deaths may also carry far more people and vehicles than another. Risk can be expressed per trip, distance, or road user, but Bangladesh first needs consistent traffic counts and travel surveys. Until then, agencies should publish both absolute harm and available exposure proxies, with the limitation stated. Ranking places by counts alone can overlook a lightly travelled road with an extreme rate, while ranking by uncertain exposure can create false precision.

The strongest counterargument is that the model may be wrong

The strongest objection is that WHO's estimate is far above the reported total and may be too dependent on cross-country assumptions. Bangladesh should not reorganize policy around a model whose central figure cannot be matched to named cases. Local reporting practices, health access, and crash definitions may not fit the model perfectly.

That objection is exactly why the estimate must remain labelled and separate. A modeled value should not be granted the authority of a complete registry. The confidence interval also makes uncertainty explicit.

But rejecting the model does not validate the reported count. The appropriate response is to measure the missing overlap among existing systems. If police, hospital, and civil-registration records can be linked, analysts can estimate capture directly, reconcile duplicates, and test how many deaths occur after the initial report. The national total can then improve without borrowing its level from an external model.

Record linkage should be conservative and auditable. Exact matches will be rare when names, times, and locations are entered differently. Probabilistic matching can help, but every rule creates false matches and missed matches. Agencies should validate a sample manually, publish error rates, retain source records, and allow the annual total to be revised when a later outcome is confirmed. A revised count is evidence of learning, not institutional embarrassment.

There is a governance benefit as well. When agencies own separate totals, disagreement can become institutional competition. A linked system changes the question from whose number is correct to which cases each system captured and why. That is answerable.

What would change this conclusion

A specific test would change the argument: if record linkage shows that the national reporting chain captures nearly all deaths under a published common definition, with minimal unmatched hospital and civil-registration cases, then WHO's estimate would be a poor guide to Bangladesh's level. The result should be published with matching rules, audit samples, and uncertainty.

First, build a linked fatal-crash registry. Connect police crash records, emergency care, hospital outcomes, and civil registration through privacy-preserving identifiers and a common fatality definition. Owner: road-safety lead agency with police, health authorities, and the civil registrar. Success signal: an annual reconciliation table shows records matched, unique cases, duplicates removed, late deaths added, and unmatched cases by source.

Second, target corridors from linked severity data. Rank locations by fatal and serious injury burden, inspect road design and speed, then implement engineering and enforcement packages. Owner: Roads and Highways Department and local road authorities. Success signal: treated corridors show a sustained decline in linked fatal and serious injuries relative to comparable corridors, with exposure changes reported.

Third, close the post-crash information loop. Track a seriously injured person from dispatch through facility outcome so deaths after the scene are not lost and response delays can be addressed. Owner: emergency and hospital services. Success signal: the share of serious crash records with a verified outcome rises, while median response intervals and preventable transfer delays fall.

The road death gap is not one number minus another. It is the distance between a case-based system and a modeled estimate. Bangladesh can close that distance only by seeing the same person across the institutions that already encountered the crash.

Sources

1. World Health Organization. Global status report on road safety 2023, Bangladesh country profile. Retrieved 2026-08-23: https://cdn.who.int/media/docs/default-source/country-profiles/road-safety/road-safety-2023-bgd.pdf?download=true&sfvrsn=7e673626_3
2. World Bank WDI. Mortality caused by road traffic injury, SH.STA.TRAF.P5. Retrieved 2026-08-23: https://data.worldbank.org/indicator/SH.STA.TRAF.P5?locations=BD