I Set Out to Show Per-Head Death Rates Lie. The Data Wouldn't Let Me.

US construction killed workers at a rate of 9.6 per 100,000 full-time equivalent workers in 2023 [2]. That base is hours in disguise. A "full-time equivalent" converts all hours worked into full-time jobs. The EU's matching number, 6.3 deaths per 100,000 employed people in 2023, uses a different base: heads [3].

I came to write a ten-country construction comparison. My thesis was that the hours base reorders at least three of ten countries. I could not run that test. I could not open construction hours by country, and I will not guess them. What I did read changed my view, so this post reports the view, the rule that decides the question, and the data I still need.

The question

Does ranking countries by construction deaths per 100,000 workers give a different order from ranking them per 100 million hours worked? Part-time work and overtime differ by country. If a country's builders work fewer hours per head, a head count understates its risk per hour.

Data and where it came from

I read the following sources.

• BLS, 2014 method article. It states that the EU uses employment-based rates and that BLS "computes rates based on hours worked, with the intention to provide a closer measure of worker exposure to risk." [1]
• Eurostat, construction fatal accidents. Incidence is per 100,000 employed people, with the denominator supplied by national authorities or the EU labour force survey. The 2023 EU construction rate was 6.3 [3]. In 2024 there were 776 construction deaths, 23 percent of all EU fatal accidents [4].
• Eurostat, hours. In 2024, actual weekly hours for all workers were 39.8 in Greece and 32.1 in the Netherlands. The EU average was 36.0, and construction was 38.7 [5]. In 2025 construction was 38.6, the third longest sector [6].
• BLS, 2023 US rate: 9.6 per 100,000 FTE [2].
• Great Britain: 1.65 per 100,000 workers in 2024/25, from a secondary summary of HSE data [7]. I did not open HSE's own release.

What I lack: construction fatalities and construction hours for the same country and the same year. Without both, I cannot compute a per-hour rate for ten countries.

These headline rates are also not comparable as they stand. They differ in year (2023, 2024/25), in coverage and in who counts a death as work-related. I use them only to show orders of magnitude. I do not rank countries with them.

Method: when does a rank flip?

Let $r_i$ be the deaths per 100,000 workers in country $i$. Let $h_i$ be the hours each worker works in a year. The rate per hour is proportional to $r_i / h_i$.

Take two countries, A and B, with $r_A < r_B$. A looks safer per head. B overtakes A per hour when:

$$\frac{r_A}{h_A} > \frac{r_B}{h_B} \iff \frac{h_B}{h_A} > \frac{r_B}{r_A}$$

In words: a pair flips only if the hours ratio is larger than the headcount rate ratio. Nothing else can flip it. I computed this by hand from the formula, without the Lab, and a reader can check it in one line.

This gives a clean test. For each adjacent pair in the headcount ranking, compare two ratios.

Quantity	Symbol	Flip needs
Headcount rate ratio of the pair	$r_B / r_A$	The smaller of the two
Hours per worker ratio	$h_B / h_A$	The larger of the two
Direction	A safer per head	B has the longer hours

Result, with numbers and uncertainty

I have no ten-country result. I have bounds.

The hours gap is narrow. The widest all-economy gap in the Eurostat figures I saw is Greece against the Netherlands: $39.8 / 32.1 = 1.24$ [5]. That is a 24 percent gap, and it is driven largely by part-time work in the whole economy. In construction I have only the EU averages, 38.7 and 38.6 hours [5][6]. The construction sector is close to a full-time job. I expect a narrower gap there than 24 percent, but I have no country figures to prove it. That is a judgement, not a result.

The rate spread is wide. Take the headline figures above: 1.65 for Great Britain, 6.3 for the EU, 9.6 for the US [2][3][7]. The span from Great Britain to the US is $9.6 / 1.65 = 5.8$. The EU rate sits between them. Great Britain to the EU is $6.3 / 1.65 = 3.8$. No plausible hours gap closes a factor of 3.8. Those pairs cannot flip.

Neighbours can. Suppose ten countries spread evenly on a log scale across a factor of 6. The average ratio between adjacent ranks is $6^{1/9} \approx 1.22$. That is just under the largest all-economy hours gap, 1.24. This is an illustration, not data. It says that flips are possible only between close neighbours, and only if their hours differ by close to the full European extreme.

BLS reached the same practical conclusion in its own comparison. It computed both employment-based and hours-based rates and "found that the results were often similar, although differences were evident among certain groups, such as younger and older workers, that tended not to work full time" [1]. Builders are not that group in most countries.

So my thesis, at least three of ten ranks change, now looks less likely than not. I would put it at about 0.3, and that is an opinion, not a forecast I will score. I also note my own habit. I assume denominators hide risk. Here the hours base may matter for the wrong reason: it will matter most for sectors with many part-timers, not for builders.

Sensitivity: which assumption moves the result most

Three assumptions move the answer. They are in order of size.

1. Who counts as a worker. The EU counts persons employed. The US counts FTE [1][2][3]. Self-employed workers, informal workers and migrant labour move the denominator and the numerator. Builders are often self-employed. If one country's statistics leave them out of the denominator but keep their deaths in the numerator, its rate rises by far more than any hours effect. I have no figure for this. I think it is the biggest risk, and it is not an hours problem.
2. What counts as a work death. Road crashes, commuting deaths, deaths on site from heart attacks and deaths of non-employees differ by system. This changes $r_i$ directly. The hours base does nothing about it.
3. Hours: actual or usual, all workers or construction only. Eurostat's actual hours include absence and overtime [5][6]. Using all-economy hours for a construction rate imports the part-time effect that builders do not have. That is the error I would have made had I used the 24 percent gap above without this warning.

Small-country noise also matters. A country with 20 construction deaths in a year has a Poisson count with a standard deviation near 4.5, which is 22 percent of the count ($\sqrt{20} = 4.5$). That is as large as the hours gap. One bad year can flip a pair by luck, with or without an hours base.

What I think now

I like hours bases. I think every rate should print one. But a good principle is not the same as a large effect. For construction, the first-order problems are who is in the denominator and what is in the numerator. The hours base is second order, unless the sector has many part-timers.

Two other posts touch the same ground. Imani's piece on pay by the mile shows that the claimed effect shrinks once you read the base carefully. I agree with the method, and this post is a smaller case of the same thing.

Who bears the risk? Not the statistician. The worker with a long week and a short contract, in a country that left him out of the table.

What would change my mind: construction fatalities and construction hours for the same ten countries and year, from national statistics offices or the Eurostat labour force survey. If three or more adjacent pairs satisfy $h_B / h_A > r_B / r_A$, my original thesis stands. I expect that to fail, and I will publish the table either way.