top of page

Explanation of "Healthy Vaccinee Effect" and Confounding

Healthy vaccinee effect (HVE) has become frequent topic of discussion these days.


I am glad to see this, as I have been talking about this phenomenon for the past 2.5 years, and for a long time no one seemed interested in discussing or understanding it.


It has recently become fashionable to blindly use it as a hammer to summarily dismiss all vaccine effectiveness studies, which is inappropriate.  HVE can play a role in some settings and is important to think about, but it is far from the only relevant factor to consider when evaluating a study and is certainly not sufficient to dismiss the vast literature on vaccine effectiveness. 


HVE is really an example of confounding, or more specifically a conglomeration of various residual confounders, so before we talk about what it is and what it isn’t we have to understand what confounding is and why it is a critical factor in interpreting observational studies.


I will try to explain these principles as clearly and succinctly as I can, hopefully making these concepts accessible. The focus of this post is expository.


Understanding confounding in observational studies

One key challenge in observational studies of vaccination is that people are not randomized to vaccinated and unvaccinated groups, so there may be systematic differences in the vaccinated and unvaccinated populations that make the groups inherently different.   Any differences between the groups that are also related with the outcome of interest, whether infection, hospitalization, or death, is called a confounder.  Failure to account for confounders can lead to completely wrong inference about vaccine effects.


The most obvious example is age confounding.  The vaccinated subpopulation tends to be much older than the unvaccinated population.  Given older people are more likely to be hospitalized or die if infected, age is a confounder because it differs between vaccinated and unvaccinated groups and is a strong predictor of risk of hospitalization and death.


As a result, a simple comparison of death rates of vaccinated to unvaccinated that does not adjust for age will invariantly find the vaccinated to have a higher death rate, leading one to inappropriately conclude that vaccines are “killing people” or “not protecting against death” even if the vaccines are harming no one and are having a tremendously strong effect in reducing risk of death. 


Thus, we must adjust for age as a confounder if we want to estimate vaccine effects on a factor like death.  This adjustment can be done various ways, by pairwise matching of vaccinated and unvaccinated individuals based on age, stratifying by age groups, reweighting the data to get age adjusted death rates, or by covariate adjusting for age in the modeling.  I won’t spend any more time discussing the many approaches for trying to adjust for confounders, but there are many possibilities.

Age is one of the key confounders, but not the only relevant one.  There are many others – here are a few other examples.  On one side, people with more medical comorbidities or living in care homes are more likely to be vaccinated and have higher risk of death, and on the other side people from lower SES and people who have less access or utilization of healthcare are less likely to be vaccinated and may have a higher baseline risk of death.  We can see there are many potential confounders, and some tending to make vaccinated more likely to have an event, and others tending to make vaccinated more likely to have an event. This is not contradictory, it is simply a nuance of observational data.


Any strong and rigorous observational study will have individual-level data with many of the key confounders measured, and appropriately methods used to adjust for their effects.          


However, no matter how many confounders are taken into account or how rigorously they are adjusted for, there is always a concern that there are other confounders affecting the outcomes that were not measured and may still affect the adjusted vaccinated vs. unvaccinated comparison.  This is called residual confounding.    This is much more difficult to diagnose and deal with.  This is what can drive what is called a “healthy vaccinee effect” which more generally is sometimes called the “healthy user effect”. 


So, what then is the “HVE”? 


This term is confusing to many, given that in most populations, the individuals who are older and have more pre-existing comorbidities are more likely to be vaccinated, and thus the vaccinated tend to be less healthy not more healthy.  This definitely confuses Fenton and Neil and collaborators, who think the ONS is being self-contradictory when they simultaneously claim that vaccinated cohorts tend to be older and have more comorbidities (they clearly are!) and that there are other factors at work leading to a “healthy vaccinee effect”.  Both are clearly true, and this is not contradictory, in spite of how it might seem to those who don’t really understand confounding or the principle that different factors can be simultaneously at work.


How can this be?  Well, after adjusting for the clear age and comorbidity effects (in that vaccinated tend to be older with more comorbidities), we may still see some residual confounding driven by other factors that were not measured or adjusted for in the modeling.  If this residual confounding makes the vaccinated less likely to have poor outcomes, their collective effect could be called a “healthy vaccinee effect” (HVE).

Another source of confusion for people (including Fenton and Neil) is that the “HVE” is a conglomeration of various types of confounders operating in different groups and on different time frames.  I think it is useful to partition HVE into two major parts:

1.     a time-varying component that primarily affects the first month or so after vaccination,

2.     a time-invariant or inherent component that could hold over longer periods of time.


Time varying Healthy Vaccine Effect

The effect called the "time-varying HVE" comes from the observation that in many studies, the all cause death rate is artificially low compared with the overall population in the weeks immediately following the shot.This is thought to be driven by the fact that individuals who are high risk of imminent death, e.g. in hospice, in the ICU, with end stage cancer, or recently having serious accident or some other acute medical incident, will tend not to get vaccinated while in that condition.


These individuals comprise a very small percentage of the total population, but may comprise a substantially high percentage of deaths in the population.  By removing these individuals from the vaccinated cohort, this makes the death rate of the vaccinated cohort artificially low.  Conversely, it can make the group not receiving the vaccines have an artificially high death rate, a phenomenon I call the “straggler effect” and will mention later.  


This effect washes out over time as these high-risk individuals either die or recover, with the death rate typically returning close to baseline death rates in 3-5 weeks, which is why it can be called a time-varying HVE. 


While the increased risk of some of these individuals might be captured by measured confounders, it will tend not to capture them all since some subset of these individuals did not have any long term pre-existing condition recorded, but have just experienced an acute event that suddenly put them at higher risk of imminent death.  That is why in spite of the fact that those older and with more comorbidities are more likely to be vaccinated (what Fenton and Neil derisively calls “the unhealthy vaccinee effect”), we see a marked decreased risk of death in recently vaccinated that is not captured by the modeled confounders.


As I mentioned, the time-varying HVE produces a complementary “straggler effect, in which the death rate tends to spike in a specific vaccine dose group when it becomes vanishingly small as most others have moved forward to the next dose.  For example, those remaining at one dose once the vast majority of their population have received a second dose are the “stragglers”.  These “stragglers” are enriched for the individuals who don’t move forward to 2nd dose because of high risk of imminent death, whose omission from the next dose group produces a time-varying HVE in the group of those recently receiving 2nd dose.


This “straggler effect” is clearly evident in every single age and dose group in the UK ONS all cause death by vaccination status data set, and it is also evident in every data set of deaths by vaccination status over time I’ve seen, including England, Hungary, Czechoslovakia, New Zealand, Switzerland, and Chile.


BTW, I explained this effect in blog posts on the UK ONS data in early 2022.  Professor Fenton and Neil and others did not understand this phenomenon at the time and summarily dismissed my explanation, not even willing to engage in a discussion or respond substantively to my explanations.  They insist on this narrative to this day, willing to die on this hill no matter how many times the ONS makes it clear that they did not classify deaths in the first 20 days after the shot as unvaccinated, or the deaths in the first 20 days after 2nd shot as 1st shot etc.  This should have been obvious to them given the fact they clearly specified a category <21 days after 1st shot (what do they think this category contains if they insist that deaths in the first 20 days are categorized as unvaccinated?) and that all of the ONS documentation, including letters explicitly responding to their claim, make this crystal clear.


They continue to insist that the only plausible explanation for these spikes of deaths during the dose rollouts was fraudulent misclassification of vaccinated deaths as unvaccinated, and completely dismiss the time-varying healthy vaccinee effect, and are so confused about it they don’t think that such an effect could be present if the older and frailer were prioritized for early vaccination.


They still don’t understand this effect.


So for what settings does this time-varying HVE affect a VE study, and how can it be taken into account?


First, this effect is especially pronounced in nonspecific outcomes such as all-cause deaths, all-cause hospitalizations, or all-cause ICU stays.  In these studies, it is best to stratify the vaccinated/unvaccinated comparisons over time, e.g. splitting into separate analyses for weeks 1-5 and weeks 5+, and taking the results from week 1-5 with a grain of salt because we know the time-varying HVE may be a major factor


For more specific outcomes such as covid-19 infections, hospitalizations, and deaths, this factor does not play as large of a role.  However, it could affect some of the more severe outcomes like covid-19 hospitalizations and death to some degree, so again I would recommend time stratification.


Time-invariant or inherent Healthy Vaccinee Effect

The second type of HVE, which I call the “inherent HVE” and others call “time invariant HVE”, does not so strongly depend on timing of the vaccination and can affect vaccinated/unvaccinated comparisons over a longer period of time.  When evident, it is driven by inherent or long term factors in the vaccinated that predispose them for lower risk of severe disease or death, but are unmeasured or not taken into account in the modeling.


This effect is thought to be driven by the idea that those who get vaccinated also tend to follow other public health lifestyle and healthcare recommendations that reduce their risk of severe disease and death.  This includes factors such as diet, exercise, smoking, usage of alcohol or other substances, preventive care, disease screening, regular checkups and likelihood to promptly seek medical care when a potential issue arises. 


Some studies will adjust for these factors by including confounders such as BMI, substance use, and various healthcare utilization measurements like past vaccinations, screenings, and checkups in their modeling.  But if they either don’t have these data or don’t take them into account, they result in residual confounding that can produce a “healthy vaccinee effect”.


So for what settings does this effect apply, how can it be detected and what can be done about it?


Like the time-varying healthy vaccinee effect, this factor has the largest effect for non-specific outcomes, and especially for severe outcomes like hospitalization, ICU, and death. 


One way to detect residual confounding is to consider “negative control outcomes” (NCO), which are outcomes for which there is no plausible causal effect of vaccines.  One can take the same analysis used for the primary outcome of interest and repeat it for these NCO and, if there is a strong vaccine effect evident, then there is evidence of residual confounding, and then results on the outcome of interest should be taken with a grain of salt.


This approach is popular and effective, but one must be careful to choose valid negative control outcomes.


We see a lot of recent discourse from Prasad, Hoeg and others about using “non-covid deaths” in this way, but that is only a valid negative control outcome if none of these deaths are directly or indirectly related to COVID-19.  It is invalid if the “non-covid deaths” include any undocumented COVID-19 deaths or deaths from post-COVID-19 sequelae such as cardiovascular or other events that have been documented in numerous studies to have elevated risk for up to a year after infection. 


Thus, any analyses treating “non-covid deaths” as a negative control outcome should be interpreted cautiously and is likely very conservative, since they are likely to include some unknown number of deaths from undocumented COVID-19 infections or post-infection sequelae.



At the end of the day, it is all about confounding.  That is why studies that measure and adjust for many confounders are considered higher quality, and why potential residual confounding should always be assessed and taken into account in the limitations.  The HVE is one type of residual confounding that needs to be taken into account, but is far from the only issue that must be considered when trying to interpret observational data, and it cannot be used as a blanket concern to dismiss all vaccine effectiveness studies.

502 views8 comments


su xeko
su xeko

Excellent website. Reading your articles was fun. This was, for me retro games, a quite entertaining read. I am anxious to read more from this saved source. Keep up your fantastic work!


In fact, they always stay ready to give an incomparable love treat to their customers. In a session with our Escort Service in Goa, you can try out different things. Do not get shy. Just speak with our escorts and you get the opportunity to fulfill all your sexual desires.



So the main factors that influence the effectiveness of this vaccine are age and immune factors. Thanks for the explanation



Not to mention a patients adherence to post-op recovery plans. I met with a company early this year that has developed sensor technology. Their tech is currently used in knees and is capturing huge volumes of data. Tremendous resource that provides daily feedback on how the patient is functioning with the new knee, i.e. distance walked, stride length... I think will eventually be applied across numerous device types and be a great data resource for these AI systems.


Nir Tsabar
Nir Tsabar

Indeed, healthy vaccinee effect should always be considered. This is why medical policy should ‎be decided only after conclusive (and inclusive) large randomized clinical trials. ‎


Yes — and that is why it is angood thing that fda approval depends on sufficiently large randomized clinical trials based on pre determined criteria being successfully completed.

But it is always necessary to conduct post approval safety monitoring and assess real world effectiveness studies using observational studies to learn about rare events and performance in real world settings that could not be learned from the randomized trials.

That is why the take home point is not that observational studies are fatally flawed so should be ignored or dismissed.

It is that it is important to take HVE and other potential confounding and biases into account using appropriate design and analysis when doing observational studies to be sure the observational…

Post: Blog2_Post
bottom of page