A French cohort study1The French study: https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2842305 followed almost 28 million adults (under 65) for four years. They divided the population into vaccinated and unvaccinated and calculated the probability of dying. The main result made all the newspapers: vaccinated people had a hazard ratio of 0.26 for death from COVID‑19. According to the model, this meant 74% protection compared to unvaccinated people. A large number.
But the same table also shows something else: those vaccinated did better on almost all causes of death. Now we have often seen the vaccine being promoted as an elixir of life and then the results were considered unreliable.

Critics who raise this issue are being reprimanded this time: the authors have actually calculated and weighed very carefully, they write. Fewer heart attacks: explained by lifestyle, nutrition, of course fewer accidents (think of properly wearing a bicycle helmet, heavier and safer cars), fewer drownings because there are many migrants in the lowest quintile, even 22% fewer2HR 0.88 = 22% less suicides (logical because less debt problems). All those advantages are exactly as they are: you see the socio-economic differences here. Those who are vaccinated would on average be healthier, wiser and better educated because they have a higher socio-economic status. And suicide is certainly less common among 'better situated people'.
That sounds plausible – it just doesn't match their own data. The study does record SES3SES: Socio-Economic Status-differences, but biological and behavioral markers that are inextricably linked to these are missing. That cannot be true at the same time.
1. Missing SDI indicators
A difference between vaccinated and unvaccinated people is visible in the 'Social Deprivation Index': 27% of unvaccinated people are in the most disadvantaged quintile4SDI quintiles: This study works with 5 SES classes (quint=5), against 19% of those vaccinated: relatively 42% more. At first glance, this seems to support the SES explanation. But that makes the rest of the data even more mysterious. It is then epidemiologically impossible that we do not see a difference in smoking, obesity, diabetes and alcohol addiction. These should be 2-3x higher in unvaccinated people.
This basic table shows that important SES-related differences are actually missing from the group distribution:

| Feature | Vaccinated | Unvaccinated |
|---|---|---|
| Alcohol addiction | 1,4 % | 1,5 % |
| Tobacco use | 5,0 % | 4,5 % |
| Obesity | 0,9 % | 0,7 % |
| Diabetes | 2,0 % | 2,0 % |
The pattern is not only minimal; it is even partly going in the wrong direction: there are more smokers and more cases of obesity among those vaccinated!
It goes against everything known epidemiologically:
- Alcohol addiction is 2 to 3 times less common among the better off. Source: Lancet5Probst et al., Lancet Public Health 2020: https://doi.org/10.1016/S2468-2667(20)30052-9
- In France (and the Netherlands) the lowest income group smokes 2–2.5x more often then the highest. (Sources: WHO6WHO Europe Tobacco Report 2023: https://www.who.int/europe/publications/i/item/9789289059283, Eurostat7Eurostat tobacco statistics: https://ec.europa.eu/eurostat)
- Obesity is coming about 2x more often for low SES. (Sources: OECD8OECD Obesity Update 2024: https://www.oecd.org/health/obesity-update.htm, ObÉpi9ObÉpi-Roche (Inserm / Kantar Health): https://www.obepi.org/)
- Type‑2 diabetes is in the lowest SES quintiles 2–3x as frequent (in the Netherlands for example 9.8% versus 4.1%10Diabetes by income: https://www.vzinfo.nl/diabetes-mellitus/inkomen and even a factor of 411Maastricht 2024: https://repository.ubn.ru.nl/bitstream/handle/2066/313262/313262.pdf).
So there are three options:
- the SDI doesn't measure what we think, either
- the risk factors are seriously under-registered in exactly one group, or
- the model has corrected the groups in such a way that reality has disappeared.
None of these options are reassuring for the reliability of the study.
The absence of those SES-related differences means that the supposed “SES artifacts” are created by the model. Meaningful weighting corrects differences to make groups comparable. The opposite seems to be happening here.
2. The doubling of suicide
Even more remarkable: according to the model, vaccinated people had a 12% lower risk of suicide (HR = 0.88). But research has shown for decades that suicide occurs about 60% less in higher social classes than in lower ones. That's an HR of around 0.4 or 0.5, not 0.88.12Lorant et al. 2005 – European comparative study: https://doi.org/10.1192/bjp.187.1.49 13Lorant et al. 2021 – Socioeconomic disparities in suicide: https://doi.org/10.1371/journal.pone.0243895
In this study, vaccinated people, who are more likely to come from those higher strata, commit suicide twice as often as expected.
That's a strong indication that something is wrong. The model corrects incorrectly or the groups are actually composed very differently than expected.
Unless, of course, we are prepared to believe that vaccination actually leads to more suicides.
And if we refuse to assume that for this rule, we must also ask how reliable the other rules in the table are. You can't cherry-pick from the results: I believe that one, I don't believe that one.
What does such a Cox analysis actually do?
The study uses a Cox proportional hazards‑model. That model compares the chance of dying at any time between two groups, with the assumption that that ratio remains constant.
In reality, this ratio often changes: after vaccination the risk may be temporarily higher (also due to stress, inflammation, or coincidental diseases) and then return to normal. A Cox analysis smooths out such temporary peaks into one friendly-looking average. Even if the difference in socio-economic status is only limited, the structurally lower background mortality of the higher quintiles causes the hazard ratio to fall sharply over the entire period. and can even become lower than 1, while there is more mortality per 100,000 people.
From hazard ratio so answers not the question: who dies more often within four years? But: How do the instantaneous mortality probabilities compare on average over time, under fixed assumptions? If those assumptions are incorrect, the outcome will also be incorrect. Regardless of the fact that it primarily concerns the first question: the mortality outcome.
3. Raw numbers vs. model figures
For suicide, perhaps the hardest and best recorded and studied cause of death, the actual numbers in the study are: 229 per million vaccinated versus 222 per million unvaccinated – a ratio of 1.03. In the model this becomes “12% less risk”.
The same pattern occurs for other causes: the crude incidence hardly differs, but the weighed HR’s jump to 0.7–0.8. The more it is “corrected”, the better it looks. This gives you a table in which it appears that vaccination protects against everything from falls to suicide. That is not a biological effect, but a sum of model assumptions, SES adjustment and time averaging.
Prioritizing actual mortality numbers over hazard ratios is not a simplification but a conscious and correct choice of an effect measure that fits the research question.
It is methodologically completely legitimate to give more weight to the incidence rate and absolute survival than hazard ratios in mortality research. When the research question revolves around who makes it to the finish line, the fact that someone has died is decisive, not the exact time within the life course and average risks. Hazard ratios describe time-dependent risks and are, of course, model-dependent.
So?
If the assumed socio-economic differences are not visible because the raw figures hardly differ and if the model then calculates a benefit for each cause of death, then the problem does not lie with people's behavior or vaccinations, but with the calculation method itself.
A model that structurally corrects in one direction produces an illusion of protection. All that really matters are the absolute outcomes: who is alive, who is not.
The French mega-study therefore shows less about the population than about the effect of the applied statistics. Anyone who accepts the calculated COVID-19 protection as true must also recognize that, according to the same model, suicides among better-off vaccinated people have roughly doubled.
In short: “These findings suggest that caution is required when interpreting the Hazard Ratios”, an observation that could also be formulated very differently.
References
- 1The French study: https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2842305
- 2HR 0.88 = 22% less
- 3SES: Socio-Economic Status
- 4SDI quintiles: This study works with 5 SES classes (quint=5)
- 5Probst et al., Lancet Public Health 2020: https://doi.org/10.1016/S2468-2667(20)30052-9
- 6WHO Europe Tobacco Report 2023: https://www.who.int/europe/publications/i/item/9789289059283
- 7Eurostat tobacco statistics: https://ec.europa.eu/eurostat
- 8OECD Obesity Update 2024: https://www.oecd.org/health/obesity-update.htm
- 9ObÉpi-Roche (Inserm / Kantar Health): https://www.obepi.org/
- 10Diabetes by income: https://www.vzinfo.nl/diabetes-mellitus/inkomen
- 11Maastricht 2024: https://repository.ubn.ru.nl/bitstream/handle/2066/313262/313262.pdf
- 12Lorant et al. 2005 – European comparative study: https://doi.org/10.1192/bjp.187.1.49
- 13Lorant et al. 2021 – Socioeconomic disparities in suicide: https://doi.org/10.1371/journal.pone.0243895

Haven't looked at the original study, but
As Tobacco use 5.0% vs 4.5%
Obesity 0.9 % VS 0.7 %
Diabetes 2,0 % vs 2,0 %
I assume that the data for a number of variables were not known (obesity in France <1%, for example, is not compatible with what you can see in the French street scene). So either they pretended not to have been exposed (to obesity, smoking, etc.) for anyone who had not reported exposure, or they left the missing data out of their model (selection bias). Or both.
All in all, of course, this doesn't matter. Whatever the cause of the result, it is a story that is too good to be true (vaccination not only prevents mortality but also falls, suicide, etc, etc, etc). The conclusion is: Garbage in
Garbage out. And that under the banner of Science.
French arithmetic…
This is very reminiscent of that Dutch study that also yielded nothing useful, with similar propaganda in the media (this one on Euronews).
Of course, it cannot be true that vaccination greatly reduces the risk of all types of accidents. The healthy vaccine bias due to better social status is apparently enormous and insufficiently corrected.
In raw figures, there was 0.12% (not 0.2%) more mortality among the unvaccinated. But only about 0.004% of this was due to excess mortality due to Covid.
If the mRNA vaccines (over that period for people under 65) were three times more deadly than Covid itself, then that was very bad (especially because vaccination did little or nothing to protect the elderly). But such an excess mortality of 0.012% in that age group is still an order of magnitude smaller than the difference in mortality due to social differences, and “disappears into the noise” (is not measurable) – contradicting the article's claim.
Okay. I'm honest about it. I really enjoy reading the articles here, with great interest and usually I understand the content. This one is a bit more difficult to follow... My knowledge of statistics is insufficient. Still, I hope you will persevere, every article is still a confirmation of my decision not to have my injection. Thanks for that!
Sorry! Without substantiation, nothing remains of the statement 'it is not correct'.
But that Cox story isn't really necessary.
The most important thing is that you explain differences, for example, by the fact that one group consists of car drivers and the other of motorcyclists, while your research results also show that both groups have the same number of helmets at home. Something like that cannot be reconciled.