The EPI-PHARE Study: Professor Rentier Sheds Light on the Matter

The EPI-PHARE/JAMA Network Open^[1] study concludes that there is a reduced risk of “all-cause mortality” as well as COVID-19 among people aged 18 to 59 who have been vaccinated with an mRNA vaccine. However, its design introduces several major biases from the outset that prevent the conclusion of “no long-term risk,” let alone an overall protective effect on mortality.

The study examined a cohort of 28.6 million people aged 18 to 59 who were alive as of November 1,2021, including 22.7 million who received an mRNA vaccine between May 1, 2021, and October 31, 2021 (79.4%) and 5.9 million unvaccinated individuals (20.6%).

The follow-up period averages 45 months but is subject to a surprising measure: it began only 6 months after the vaccination date (or a date assigned to the unvaccinated), under the pretext of avoiding “immortal time.”

Immortal time: a follow-up period during which, by design of the study protocol, an individual cannot yet experience the event under study (e.g., “death among the vaccinated”), which skews the comparison between groups. “Avoiding immortal time” involves defining the start of follow-up and exposure in such a way that no group artificially benefits from a period during which the event cannot occur by definition. This obviously implies that a definitive decision is made from the outset that a death occurring within the first 6 months after vaccination cannot under any circumstances be attributed to the vaccine.]

Cox models were weighted by propensity scores, which included sociodemographic characteristics and 41 known comorbidities.

In biostatistics, a propensity score is the probability that an individual will receive a treatment (or be exposed) given their observed characteristics. In other words, it is a function of the baseline variables (age, sex, comorbidities, risk factors, etc.) that summarizes in a single number between 0 and 1 the “propensity” of that individual to be in the treatment group rather than the control group. The propensity score is primarily used in observational studies to reduce selection bias and to approximate, as closely as possible, the conditions of a randomized trial. By matching, weighting, or stratifying subjects based on this score, the goal is to make the distributions of covariates comparable between the treated and untreated groups, so that the observed difference in outcomes can be interpreted more credibly as a causal effect of the treatment. The causal interpretation of analyses based on the propensity score relies in particular on the assumption of conditional independence: once conditioned on the covariates included in the score, treatment assignment is assumed to be “ignorable” (no other unmeasured confounding factors). There must also be satisfactory common support (overlap of scores between groups), so that each treated subject has comparable controls and vice versa.

The article does not provide a sufficiently detailed description of the propensity score: neither the choice of covariates, nor the quality and granularity of the coding, nor the residual imbalances, nor the balance diagnostics after matching or weight^[2]ing.

Without these elements (e.g., standardized mean differences, variance ratios, density plots), it is not possible to assess to what extent the compared groups are truly balanced with respect to the measured factors.

What conclusion do the authors draw?

The authors conclude that, in their national cohort of French adults aged 18–59, people who received at least one dose of an mRNA COVID-19 vaccine had a lower risk of death—both from all causes and from severe COVID-19—than those who remained unvaccinated during the study period.

• They report a 74% reduction in the risk of death from severe COVID-19 requiring hospitalization among the vaccinated and a 25% reduction in the risk of all-cause mortality over a follow-up period of up to 4 years, after adjustment by propensity score.

• They indicate a lower risk of death in the vaccinated group “regardless of the cause of death considered,” including after excluding deaths due to severe COVID-19.

• The authors state that these results “rule out an increased risk of all-cause mortality at 4 years” among individuals in this age group who received mRNA-based COVID-19 vaccines. They conclude that their study reinforces the long-term “safety profile” of widely used mRNA vaccines and that vaccination is not associated with excess mortality within this cohort.

Major structural biases are immediately apparent in this study

1. The selection of the unvaccinated and the “healthy vaccinee bias.”

• The unvaccinated in France in the fall of 2021 constituted a highly heterogeneous subgroup: it includes an expected overrepresentation of people in precarious situations, unregistered migrants, individuals who have dropped out of care or suffer from severe psychiatric disorders, untreated medical conditions, etc., who are often poorly or belatedly recorded in the National Health Data System (SNDS).
• The propensity score adjusts only for observable and correctly coded variables (41 reported comorbidities), leaving intact a potentially massive “residual confound,” to the detriment of the unvaccinated.
• The fact that a “benefit” is found for virtually all causes of death, including those that a priori have no plausible link to vaccination (accidents, certain early-onset cancers, violent deaths) typically suggests a “healthy vaccinee” selection bias rather than an overall biological effect.

2. Start of follow-up at 6 months

• Follow-up did not begin until 6 months after vaccination under the pretext of avoiding a classic “immortal time” bias (see above). However, this results in the exclusion of all early deaths from the main analysis, even though this is precisely where post-vaccination adverse effects (cardiovascular disorders, myocarditis, arrhythmias, etc.) might be concentrated.
• These deaths occurring within 6 months post-vaccination are analyzed in a separate sub-study and are therefore decoupled from the 4-year risk estimate; the “absence of excess at 4 years” therefore says nothing about the acute or subacute risk window, but merely sums up what happens beyond 6 months.

3. Change in vaccination status

• It happened that unvaccinated individuals who were more health-conscious got vaccinated during the follow-up period. They were excluded from the follow-up on the date of vaccination, whereas vaccinated individuals remained in the follow-up with cumulative doses and boosters. There is no explicit modeling of these repeated exposures as time-dependent covariates. This accentuates the contrast between a “compliant” vaccinated group and a residual group of unvaccinated individuals who are more socially and medically vulnerable, not captured by the propensity score.

4. Significant residuals of socioeconomic and behavioral confounding

• EPI-PHARE notes in its reports that its studies rely on the National Health Data System, which only imperfectly captures many major determinants of health (income, education, isolation, housing conditions, lifestyle habits, occupational exposure, etc.). Yet all these variables are strongly correlated not only with vaccination choice but also with the risk of all-cause mortality.
• The fact that vaccination appears to reduce mortality from nearly all causes is consistent with a behavioral and/or socioeconomic gradient that is not measured at all, rather than with a universal pharmacological effect.

5. Limited information on detailed causes of death and overall excess mortality

• Detailed analysis of causes of death is available only through the end of 2023, whereas vital status follow-up extends through March 2025; this prevents the evaluation of cause-specific profiles across the entire 4-year window.
• The study does not address overall excess mortality in the French population for 2021–2023 (all causes, all ages) or its relationship to the vaccination schedule, which is at the heart of the scientific debate on indirect or delayed effects.

6. Excessive Media and Political Interpretation

• Several news articles and institutional press releases present these results as “putting an end to doubts about long-term risks” or “demonstrating the absence of danger,” whereas the study is limited to all-cause mortality in a French cohort aged 18–59, with the biases noted above.
• Moving from “no increase detected” to “long-term safety demonstrated” is an unjustified logical leap because this is a single observational study, even if it is very large, especially given the high likelihood of residual confounding.

What is “all-cause mortality”? All-cause mortality refers to the total number of deaths occurring in a given population during a specific period, regardless of the cause of death. It is an epidemiological metric that counts every death, whether due to an infectious disease, cancer, an accident, a heart attack, suicide, or any other cause. For example, if a region with 100,000 residents records 1,200 deaths in a year, regardless of the cause (COVID-19, flu, heart attack, car accident, etc.), the all-cause mortality rate for that year is 1,200 deaths, or 12 deaths per 1,000 inhabitants. All-cause mortality is considered a robust epidemiological indicator because it does not depend on the accuracy of death certificates or medical diagnoses. It avoids biases related to incorrect attribution of the cause of death, underreporting (undetected or misattributed deaths), changes in registration criteria over time, and competition between causes when multiple diseases are circulating. Analyzing all-cause mortality allows us to identify the overall impact of a public health intervention or event on the total number of deaths, without getting lost in the details of specific causes. This is particularly relevant during a pandemic or a vaccination campaign: even if deaths from a given disease are reduced, we must ensure that we do not increase all-cause mortality (i.e., that we do not replace one cause of death with another, or that we do not cause excess mortality through other mechanisms). In the current context of post-marketing surveillance (pharmacovigilance), all-cause mortality is a critical evaluation criterion.  A vaccine effective in reducing deaths from a target disease should not increase the risk of death from other causes. This is why public health authorities examine this indicator. In the EPI-PHARE study, the analysis of all-cause mortality by vaccination status raises the following question: “Does the pattern of deaths (all causes combined) differ systematically between vaccinated and unvaccinated individuals?” This implies that the authors are seeking to determine whether, overall and not just for COVID deaths, the two population groups exhibit comparable mortality trajectories. A significantly higher all-cause mortality rate in one group would raise questions about potential causes—whether related to the vaccine, differences in group composition (selection bias), differences in underlying risk factors, or other epidemiological factors. Although the authors note in the technical report that the study is observational and does not prove causality, this nuance is lost in press releases and media coverage. The public and policymakers take away the simplified message: “vaccines reduce all-cause mortality,” without understanding that this association is likely confounded.

 

A detailed methodological analysis of the propensity score is therefore lacking (variables included, coding quality, residual imbalances, calibration diagnostics).

There was no discussion of unmeasured socioeconomic and behavioral determinants, drawing on the literature on “healthy vaccinee/user bias” in observational vaccine studies.

Finally, a comparative review with other sources on excess mortality in Europe post-2021 is missing, particularly those that stratify by vaccination status or campaign period, to demonstrate that the picture is more nuanced.

In summary

The study does not allow us to rule out an increased risk of death within 0–6 months, nor to draw conclusions about specific rare risks (cardiovascular, autoimmune), which would require dedicated and more robust protocols.

Nor does it allow us to conclude that there is no long-term risk for all age groups, or to extrapolate beyond the French context and the 2021 vaccination campaigns (different variants, different regimens, different epidemic pressure).

It does not provide a sufficiently detailed description of the propensity score: choice of covariates, quality and granularity of coding, residual imbalances, or assessments of balance after matching or weighting. Without these elements (e.g., standardized mean differences, variance ratios, density plots), it is difficult to assess the extent to which the compared groups are truly balanced on the measured factors.

It does not discuss unmeasured socioeconomic and behavioral determinants, even though the literature shows that people who get vaccinated are often healthier, more compliant, and more socially integrated (“healthy vaccinee/user bias”). The lack of analysis of these selection biases and preventive behaviors associated with vaccination status leaves open the possibility of an overestimation of the vaccine’s safety or apparent benefits.

^[7]Finally, it does not put its results into perspective with other studies on excess mortality in Europe after 2021, particularly those that stratify by vaccination status or by campaign^[6][5][4][3] phases. A systematic comparative review of these time series and stratifications would be necessary to show that the overall picture is more nuanced than what the study alone suggests.

Prof. Bernard Rentier

[1] mRNA vaccination against COVID-19 and all-cause mortality

[2] Balance of diagnoses after propensity score matching - PMC

[3] European excess mortality following COVID-19 vaccination Patrick E. Meyer, excerpt from the February 18, 2024 version submitted to

[4] All-cause mortality according to COVID-19 vaccination status: An analysis of the UK Office for National Statistics public data - PubMed 

[5] Understanding excess mortality in Europe during the COVID-19 pandemic - The Lancet Regional Health

[6] Excess mortality across countries in the Western World since the COVID-19 pandemic: ‘Our World in Data’ estimates from January 2020 to December 2022 | BMJ Public Health

[7] Sustained excess all-cause mortality post-COVID-19 in 21 countries: an ecological investigation | International Journal of Epidemiology | Oxford Academic