The figure the New Zealand work health and safety regulator publishes for deaths from work-related ill health, 750 to 900 a year, is a 2019 estimate modelled from international attributable fractions rather than a count, and all four system-level measures in its 2024/25 annual report cover acute injury or notified events

What is happening?

New Zealand has a published national figure for deaths caused by work-related ill health. It is 750 to 900 a year, set alongside 5,000 to 6,000 hospitalisations a year, and both come from a single piece of work: a report published by WorkSafe New Zealand in August 2019. That report states how the figures were produced. They apply international work-related attributable fractions to New Zealand mortality and hospital event data; they exclude deaths and hospitalisations caused by acute injury; and the ranges are described as scenario-based rather than confidence intervals, rounded to the nearest 50 and the nearest 1,000.

The regulator has repeatedly described this category as the larger part of the harm it exists to reduce. Its October 2021 factsheet on the 2020 injury target refers to long-term ill health as the largest source of work-related harm. Its December 2019 outcomes dashboard records that, compared with work-related acute injuries, it knows less about work-related ill health, and that it does not have detailed data on exposures to work-related carcinogens or on how well those risks are controlled in workplaces. The same dashboard puts the modelled burden of all work-related harm at about 50,000 disability-adjusted life years a year, of which ill health is 89 percent, about 45,000.

The numeric targets have been attached to the other part. In 2013 the Government set a target to cut fatal and serious non-fatal work-related injuries by 25 percent by 2020, with an interim 10 percent by 2016, measured by three indicators: fatal injuries, serious non-fatal injuries, and injuries involving more than a week away from work. The October 2021 final report recorded that the fatal rate had met its target early but that progress had stalled since about the middle of the previous decade, at a three-year average of 2.2 per 100,000 full-time equivalents for 2018 to 2020 against a baseline of 3.3 and a target of 2.5. Serious non-fatal injuries reached 16.5 in 2020 against a target of 14.3, and injuries with more than a week away from work reached 12.6 per 1,000 full-time equivalents against a target of 8.4. A December 2022 factsheet recorded that the period for the system targets had ended and that injury trend monitoring would continue.

The measures now in force keep that shape. The 2024/25 annual report, whose statement of responsibility was signed on 2025-10-31, carries four system-level performance measures on printed pages 22 and 23: the fatal acute injury rate, age-standardised, with a target of under 2.0 per 100,000 full-time equivalents by 2027/28; the serious non-fatal injury rate, with a target of under 14.0 by 2027/28; the same serious non-fatal rate broken down by ethnicity; and the share of notified events at high-risk workplaces that were high-potential events, reported as 26 of 296 for 2024/25. Two of the four carry a number and a date. All four cover acute injury or notified events. The same pages record that the Ministry of Business, Innovation and Employment led the development of the system-level measures and that WorkSafe contributes to reporting on progress.

Whose problem is this?

The body that publishes the estimate of how large this harm is and the body that sets the national measures sit inside the same system. The quantity that system calls the largest source of work-related harm is the one that arrives as a model output rather than as a count.

Where does this problem end?

The boundary matters because the dispute here is not about whether the harm is large. The regulator has already put its own estimate of the size on the record, and has already recorded that it knows less about this part of the harm than about the other part. What is at issue is the kind of number that exists, and what a number of that kind can be asked to carry.

What is the state now, and what should it be?

Now

Needs a new measurementthe target state: no source opened here sets a number or a date for work-related ill health. Not a reduction target for deaths, not a target for the burden, not a deadline for a counted measure to exist. The closest thing any opened source contains is a statement the regulator made about itself in the December 2019 dashboard: that there is a need to develop a measurement framework to track progress in reducing work-related harm, together with a commitment to produce the first dashboard report by the end of 2019 and to update it annually. That is a statement of need rather than a target, and it carries no figure.

How big is it?

On the 2019 basis the regulator puts work-related ill health at 750 to 900 deaths and 5,000 to 6,000 hospitalisations a year. Cancer and respiratory disease account for most of both — about 650 of the 750 deaths and about 3,900 of the 5,000 hospitalisations on the lower-bound scenario. The whole modelled burden of work-related harm is about 50,000 disability-adjusted life years a year, with a social cost the regulator states as at least 2 billion dollars, and ill health is 89 percent of that burden, about 45,000.

The quantity that carries the national numeric targets is counted, and it is smaller. The regulator recorded 63 fatal acute injuries in the 2018 calendar year, 53 workers and 10 bystanders, drawn from notifications and compensation claims. The most recent provisional figure in the 2024/25 annual report is 139 fatal acute injuries across 2021 to 2023, a rate of 1.6 per 100,000 full-time equivalents.

The two quantities are not the same kind of thing, and this document does not divide one by the other. One is a model output for disease categories with latency measured in decades. The other is a count of events that are notified when they happen. The regulator states on one of its own topic pages that work-related ill health kills about 15 times as many people as acute injury at work; that page gives no year and no underlying counts for the comparison, so the multiple is reported here as a statement of the regulator and not as a computed value.

The one series that counts individual cases rather than modelling them is administrative. Gradual process claims to the Accident Compensation Corporation averaged 5,281 a year across 2015 to 2017 on an extract dated 2019-06-06, and the dashboard records that only a limited subset of work-related ill health is eligible for compensation. That series therefore measures compensable cases rather than cases. On the exposure side, the 2024 overview of harm and risk reports that more than 1 in 2 workers are exposed to at least one carcinogen at work, on data as at 2023-06.

Under what conditions does it arise?

1. Individual attribution does not work for this category. Latency runs for decades and most of the conditions have several possible causes, so a single case cannot generally be assigned to work. The regulator says as much about itself: its work-related health position document, last updated in 2017, describes population-based modelling as the route it takes instead of case-by-case attribution.

2. The administrative count that does exist is shaped by compensation. Gradual process claims are the one series that counts individual work-related illness, and the dashboard records that only a limited subset of work-related ill health is eligible for compensation. A series built on eligibility answers a question about entitlement, not a question about how much disease work causes.

3. Notification duties reach some health events, and no opened source connects them to the estimate. A person conducting a business or undertaking must notify the regulator of notifiable events, and the published list includes serious infection and medical treatment within 48 hours of exposure to a substance. A Medical Officer of Health notifies the regulator of a notifiable disease believed to have arisen from work. No source opened here states that these notifications feed the national ill-health estimate, or gives the annual number of such notifications.

4. A numeric target has to attach to a series. Both target rounds attached to injury series — the three indicators used for the 2013 target, and the two rates carrying 2027/28 targets in the 2024/25 annual report. Those series have baselines, a reporting cycle and stable definitions, which is what makes a number and a date possible.

5. The exposure side is not measured in detail either. The 2019 dashboard records that detailed data on carcinogen exposures, and on how well those risks are controlled in workplaces, is not held. Without exposure data a leading measure cannot be built to stand in for the lagging one.

None of these five conditions is a statement about intent. They describe why the quantity arrives as a model output, and why the series that carry targets are the injury series.

What has been tried?

Two directions have been run side by side for a decade. The injury side has been given baselines, indicators, numbers and dates twice over. The ill-health side has been given one estimate, a statement that a measurement framework is needed, and no number to be measured against.

What was found?

Why is it still unsolved?

Measurement absence — the part of work-related harm the regulator itself calls the largest arrives as a model output, and the measures that carry numeric targets are attached to the series that are counted.

A numeric target is a promise about a series. It needs a baseline, a cycle on which new values arrive, and a definition stable enough that a later value can be compared with an earlier one. The injury indicators have all three, which is why they could be given a 25 percent cut by 2020 in 2013 and a rate of under 2.0 by 2027/28 in the current annual report. Whether those targets were met is a separate matter; what matters here is that they could be written at all.

The ill-health quantity does not have those three things in the same sense. It was produced once, in 2019, by applying international attributable fractions to national mortality and hospital data. A model output can be recomputed, but recomputing it moves the method as well as the value, and the sources opened here show no second edition against which the first could be compared. A number that arrives once does not behave like a series, and a target written against it would be a promise about a model rather than about workplaces.

Underneath that sits an obstacle the regulator names itself. For disease with long latency and several possible causes, individual cases cannot generally be attributed to work, so the count a target would need cannot simply be taken. The only series that does count individual cases counts compensable ones, and the dashboard records that only a limited subset of work-related ill health is eligible for compensation. The exposure side, which could support a leading measure, is recorded in the same dashboard as lacking detailed data.

What the opened record does not contain is a reason. The annual report says that the Ministry of Business, Innovation and Employment led the development of the system-level measures and that WorkSafe contributes to reporting on progress. It does not say why the scope of those measures is what it is, and no other source opened here says either. This document does not supply one.

The last part is what keeps the pattern stable. An injury target produces events — a baseline, a midpoint, a final report, a factsheet saying the period has ended. A measurement gap produces none of those. The 2019 dashboard recorded a need to develop a measurement framework and a commitment to annual updates; no later edition turned up in the sources opened here, and nothing in the process marks the difference between a framework still being developed and a framework that is not. The estimate simply stays on the page with its 2019 date, and each new annual report reports faithfully against the measures that exist.

What observation would mean it is solved?

Candidates — first, the regulator publishes a counted national series for work-related ill health, with a stated definition and a stated cycle; second, a system-level performance measure covering work-related ill health appears in an annual report with a baseline, a number and a date; third, the 2019 estimate is superseded by a later edition that can be compared with it.

The first is the hardest and the most load-bearing. It is also the one the regulator has already said is obstructed, because individual attribution does not work for long-latency disease. A series built instead on notifications would count what gets notified, and a series built on compensation claims would count what is compensable. Either could be published and described as a measure of work-related ill health while measuring something narrower, and the gap between the label and the content would not be visible in the number.

The second can be satisfied without the first. A measure can be declared against a modelled quantity. If that happens, progress against the target becomes a statement about successive runs of a model, and a change in the attributable fractions used would move the reported value without anything changing in any workplace.

The third is the weakest of the three. A second edition would show that the estimate is maintained, not that the underlying harm is counted. It would also be hard to read against the first, because a revision that changes the method changes what the two numbers mean, and the 2019 report already describes its own range as scenario-based rather than statistical.

The three have to be read together, and the second has to be read against whichever of the first and third it rests on.

What is it connected to?

Fills with researchoccupational disease measurement in other jurisdictions, the design of compensation eligibility for gradual process conditions, carcinogen exposure surveillance, and the relationship between statutory notification duties and national estimates. Relation type and evidence grade were not confirmed in this round.

What these sources do not say

• Why no system-level measure covers work-related ill health. The 2024/25 annual report records that the Ministry of Business, Innovation and Employment led the development of the measures and that WorkSafe contributes to reporting on progress. No source opened here gives a reason for the scope of those measures, and none carries an on-the-record explanation from either body.
• Whether the 2019 estimate has been revised. No source opened here carries a revision. Three regulator pages that give the figure all show a 2019 update date. The regulator data centre, which might hold a newer edition, returned only a script shell and its contents are unknown.
• Whether a later outcomes dashboard exists. The December 2019 edition states a commitment to annual updates. No later edition appears among the sources opened here.
• Whether notification duties feed the national estimate. The published notifiable event list includes serious infection and medical treatment within 48 hours of exposure to a substance, and a Medical Officer of Health notifies notifiable disease believed to have arisen from work. No opened source states that these notifications are an input to the ill-health estimate.
• How many work-related disease notifications are received a year. No opened source gives that count. The figure of 296 notified events in the annual report covers high-risk workplaces and is not a disease count.
• Why the two acute-injury burden figures do not reconcile. The December 2019 dashboard gives the acute injury share of the modelled burden as 11 percent and separately states 4,550 disability-adjusted life years, against a stated total of about 50,000. No opened source explains the difference, and this document does not resolve it by division.
• What the 15 times comparison rests on. The regulator page carrying it gives no year and no underlying counts, and no opened source states the basis on which a modelled figure and a counted figure are being compared.
• Worker and full-time-equivalent denominators. No source opened here supplies the workforce totals that would turn the ill-health estimates into a rate, or turn the exposure share into a count of exposed workers.
• What Healthy Work set as a number. The opened page describes outcomes to be reached by 2026 in terms such as improved understanding by businesses. No opened page gives a numeric reduction target for work-related ill health under that plan.
• Whether the 2026 reform bill has been enacted. The opened summary records introduction to Parliament on 2026-02-12 and goes no further. No opened source confirms passage, assent or commencement.

See the evidence

Five documents were read directly as published files — the 2024/25 annual report, the 2019 work-related health estimates report, the December 2019 outcomes dashboard, the October 2021 final progress report and the December 2022 indicators factsheet. Every figure this document grades as high comes from those five. The remaining regulator pages, the national statistics office explanatory item and the two secondary summaries were opened through a reading tool that returns a summary, so their wording may differ by a word from the originals, and no phrase from them is reproduced here as a quotation.

Three limits on scope are worth stating plainly. Only printed pages 22 to 25 of the annual report were read, so every statement here about that report is limited to the system-level performance measures on printed pages 22 and 23 and is not a statement about the whole report. The 2019 report contains an appendix on suicide which was not opened and is not used. Where the opened sources disagree the disagreement is left visible rather than resolved — the December 2019 dashboard gives the acute injury share of the modelled burden as 11 percent in one place and as 4,550 disability-adjusted life years in another against a stated total of about 50,000, and one regulator topic page restates the 750 to 900 figure as deaths from work-related causes without the acute injury exclusion that the 2019 report states. This is a Path A output, a research-based definition, so observation_refs is empty and provenance_mode: press-derived.

People affected

Estimated range 750–900 As of 2019 estimate, report published 2019-08

Sensitivity This band is modelled rather than counted, which is the subject of the document it belongs to. The regulator produced it by applying international work-related attributable fractions to New Zealand mortality and hospital event data, and it states that the range is scenario-based and not a set of confidence intervals. It excludes deaths caused by acute injury, so it does not overlap the injury series that carry the national numeric targets. It counts deaths only. Work-related ill health is separately estimated in the same 2019 work at 5,000 to 6,000 hospitalisations a year, and the whole modelled burden of work-related harm at about 50,000 disability-adjusted life years a year with ill health at 89 percent of it, so the number of people affected in any year is larger than this band by an amount no source opened here quantifies. The workforce denominators that would turn this into a rate, or turn the exposure finding that more than 1 in 2 workers are exposed to at least one carcinogen into a count of exposed workers, were not obtained. The one administrative series that counts individual cases, gradual process claims to the Accident Compensation Corporation at 5,281 a year averaged over 2015 to 2017, cannot substitute, because the regulator records that only a limited subset of work-related ill health is eligible for compensation. No source opened here carries a revision of the 2019 estimate, so the band may be stale in either direction and the 2019 as-of should be read as the age of the model run rather than as the year the deaths occurred.

Regional breakdown No source opened here breaks the ill-health estimate down by region. The estimate is national and is derived from national mortality and hospital event data through attributable fractions taken from international literature, so no regional layer exists inside it to be recovered. Splitting it by regional population would be proportional allocation, and it would be wrong in both directions at once, because exposure to the hazards behind these disease categories is concentrated in particular industries rather than distributed with population.