Dr Clare Craig

A new paper in Medical Research Archives examines all-cause mortality across 15,000 US nursing homes from May 2022 to June 2023, stratified by vaccination status. The authors – Denhaerynck, Mead and Wolfinger – use forward and reverse lag models to show that mortality rose following weeks with more COVID-19-positive residents. Their headline finding: mortality elevations persisted for about one week in unvaccinated residents, three weeks in the partially vaccinated, and five weeks in the fully vaccinated.

The authors interpret this as evidence of prolonged vulnerability among the more vaccinated. They discuss immune suppression, IgG4 class switching, and the non-specific effects of non-live vaccines. All worthy of attention. But what caught my eye was something they didn’t discuss.

The 76% question

Their own data show that 76% of all detected infections occurred in the “partially vaccinated” group: 6,828 cases, compared with 877 among the fully vaccinated and 1,253 among the unvaccinated. That is a striking distribution.

Who are the “partially vaccinated”? According to the CDC definitions used in the study, this includes anyone within 13 days of their latest dose – the window during which post-dose immune suppression is well documented. To know whether these people were at disproportionate risk of infection, you would need to know how many residents were in that window at any given time. The authors tracked weekly vaccination percentages by status –  they say so in their methods – but they never report the breakdown. Table 1 gives only a binary split: 88.2% vaccinated, 11.8% unvaccinated.

Why would you write a paper like this and not report the proportion in each vaccination category? Is it not relevant? Or does its absence help obscure an awkward question – namely, whether the recently dosed were at disproportionate risk of infection?

They named it and walked past it

The most remarkable thing is that the authors come close to identifying this problem themselves. In their discussion of immortal time bias, they write that the 14-day classification window means that recently vaccinated residents “cannot be coded as vaccinated despite elevated short-term vulnerability, thereby lowering mortality numbers among vaccinated.”

They have named the mechanism. They have acknowledged that it biases their data. And then they move on – without connecting it to their own finding that 76% of detected infections fall in exactly that category. The three-week mortality tail they observe for the partially vaccinated may not be a tail after infection at all. It may be a tail after injection.

It seems that long-term immune suppression is a publishable finding. Short-term post-dose immune suppression – the kind visible in their own data – remains, apparently, a step too far.

Who is frail and who is healthy?

The authors’ discussion of confounding is also internally contradictory. They cite indicator bias – the sickest residents were prioritised for vaccination, meaning the vaccinated group carries higher baseline mortality risk. They also cite healthy-vaccinee bias – residents judged too frail or terminal were excluded from vaccination, meaning the vaccinated group is comparatively healthier. Both are offered as limitations. Both cannot dominate simultaneously in the same population. The authors do not attempt to resolve this contradiction – they simply list both possibilities and move on.

The 2020 baseline

The authors adjust for baseline mortality using monthly county death counts from 2020, described as “a reference year reflecting the pre-intervention baseline.” The purpose is to control for geographic confounding – the possibility that nursing homes with lower vaccination rates were in areas with higher underlying mortality.

But 2020 was not a clean baseline. It was a year of ventilator protocols, lockdown harms, and widespread iatrogenic injury in care homes. That harm was not evenly distributed. Poorer areas with less stable healthcare infrastructure – the same areas where vaccination uptake would later be lowest – were hit hardest. Using 2020 as the reference therefore overestimates the underlying mortality in those areas, and the model adjusts unvaccinated mortality down more than it should. The net effect is to make the unvaccinated look relatively better in 2022–23 than they might otherwise appear.

This is worth noting, though it is not a fatal flaw. It means the authors’ choice of baseline actually favours the unvaccinated group – and by extension could exaggerate the apparent mortality contrast with the vaccinated. But what alternative baseline could they have used? 2019 is pre-COVID but also pre-pandemic disruption. No available year is pristine.

The Overton window

It is also worth noting the choice of study period. The analysis begins in May 2022 – safely past the initial vaccine rollout, past the booster campaigns of late 2021, past the period where temporal alignment between mass vaccination and mortality spikes was most visible and most politically charged. By starting in the Omicron era, with a milder variant and high prior exposure, the authors pre-select for a weaker signal.

Perhaps that is where the Overton window can move to – but no earlier yet.

The things you have to do to be heard

What makes the paper’s conservatism all the more striking is who wrote it. Mead is at the McCullough Foundation. Denhaerynck is affiliated with Stichting Voor Waarheid – the renamed Viruswaarheid, one of the most prominent COVID-critical organisations in the Netherlands. Wolfinger, listed here as an independent researcher, is in fact a Distinguished Research Fellow at SAS (the statistical software company), a Fellow of both the American Statistical Association and the American Association for the Advancement of Science, and literally wrote the textbook on the mixed-model methods used in this analysis. These are not people who would shy away from the implications of their own data out of institutional timidity.

Which leaves the most likely explanation: this is a paper designed to survive. Frame the findings conservatively, list the biases as limitations rather than confronting them, avoid the post-dose question, start the clock in spring 2022, and conclude with a call for “further individual-level investigation.” Get it into the literature. Live to be cited. The stronger claims – the ones the data are crying out to support – would have sunk the paper before it reached print.

That is the landscape we are in. The data is there. The questions are obvious. But even sympathetic researchers with serious statistical credentials have to self-censor to get published. The Overton window is moving – but it moves at the pace of what reviewers will tolerate, not at the pace of what the evidence demands.