Prior infections and effectiveness of SARS-CoV-2 vaccine in test-negative studies: a systematic review and meta-analysis

Abstract

Prior infection with SARS-CoV-2 can provide protection against infection and severe COVID-19. We aimed to determine the impact of preexisting immunity on vaccine effectiveness (VE) estimates. We systematically reviewed and meta-analyzed 66 test-negative design studies that examined VE against infection or severe disease (hospitalization, intensive care unit admission, or death) for primary vaccination series. Pooled VE among studies that included people with prior COVID-19 infection was lower against infection (77%; 95% CI, 72-81) and severe disease (86%; 95% CI, 83-89) compared with studies that excluded people with prior COVID-19 infection (pooled VE against infection: 87% [95% CI, 85-89]; pooled VE against severe disease: 93% [95% CI, 91-95]). There was a negative correlation between VE estimates against infection and severe disease, and the cumulative incidence of cases before the start of the study or incidence rates during the study period. We found clear empirical evidence that higher levels of preexisting immunity were associated with lower VE estimates. Prior infections should be treated as both a confounder and effect modificatory when the policies target the whole population or are stratified by infection history, respectively.

Introduction

COVID-19 vaccines reduce the risk of infection and can also ameliorate disease severity when breakthrough infection occurs.^1,2 Ongoing evaluation of COVID-19 vaccine effectiveness (VE) has largely been measured through observational studies, particularly test-negative design (TND) studies, which share some similarities with case-control studies.³ However, there has been substantial variation among reported VE estimates,^4‑7 which may be attributable to differences in study design, the vaccines used, disease incidence, and population characteristics. Importantly, preexisting population immunity due to prior infection could explain changes in COVID-19 VE over time and among populations.^8,9 Among 42 systematic reviews on VE of COVID-19 vaccines, including VE against infection, symptomatic cases, severe diseases, or fatality, none of them examined the impact of preexisting immunity on the VE estimates (Table S1). This suggests that the impact of preexisting immunity on VE estimates is rarely explored.

Infection with SARS-CoV-2 induces an immune response to protect against reinfection.^10‑14 However, reinfection could occur due to waning, naturally induced immunity^15,16 or virus evolution.^17,18 Nevertheless, studies have shown that, compared with persons with no prior infection, vaccination among people with prior infection enhances neutralizing antibody activity as well as cell-mediated responses that can protect against (re)infection,¹⁹ suggesting prior infections may modify the protection from vaccinations. In settings where a large proportion of the population has prior exposure through infection, the unvaccinated will be more protected from infection than in a naïve population, thereby diluting the apparent effectiveness of vaccination. Under these 2 scenarios, prior infection modifies the effect of vaccination (Figure S1).

Prior infection can also alter an individual’s decision to be vaccinated and to present for care. For example, vaccination requirements vary for people with recent prior infection in Hong Kong.²⁰ Moreover, individuals with recent infection may choose not to be vaccinated if they believe they have sufficient preexisting immunity to prevent reinfection and ameliorate the severity of any reinfections that do occur.²¹ Additionally, these individuals may also choose not to present for care, believing their COVID-like symptoms are due to another illness, leading to differential under-ascertainment of previously infected COVID-19 cases in surveillance data. Other individual-level factors may also affect the decision to vaccinate and engage in infection-risk behaviors, such as perceived risk of severe disease after infection.^22‑24 Acting in this way, prior infection may create a confounding bias along the lines of vaccination-COVID-19 association (Figure S1).

Here, we systematically review and meta-analyze published data to characterize the potential impact of preexisting population immunity on VE estimates for completed primary vaccination series against COVID-19. We also conducted meta-regression to account for the influence of key design features such as vaccine types and circulating virus strains.

Methods

Search strategy and selection criteria

This systematic review was conducted following the Preferred Reporting Items for Systematic Review and Meta-analysis (PRISMA) statement.²⁵ The review was not registered. A standardized search was done in PubMed, Embase and Web of Science, using the search terms (“test negative” OR “effectiveness”) AND (“vaccine”) AND (“COVID-19” OR “SARS-CoV-2”). The search was done on July 11, 2022, with no language restrictions. Additional relevant articles from the reference sections of identified articles were also reviewed. Two authors (X.H. and C.W.) independently screened the titles and full texts, and extracted data from the included studies, with disagreement resolved by consensus together with a third author (T.K.T.). Studies identified from different databases were deduplicated.

Studies that reported using a test-negative approach in which all cases and noncases were tested were included,^26,27 regardless of whether the studies assessed the role of preexisting immunity on VE estimates. We included published TND studies with participants recruited from the general population and reported VE estimates for completed primary vaccination series (2 doses for most vaccines; 1 dose for the Janssen vaccine) against at least 1 of the following endpoints: (1) positive test result, (2) symptomatic disease, (3) hospitalization, (4) intensive care unit admission, (5) severe COVID-19, and (6) death. We excluded articles if (1) the study participants were recruited from a specific sub-population, such as health care professionals; (2) studies that summarized or reanalyzed already published data; (3) studies that only reported pooled VE estimates for different vaccines; (4) the study was a preprint and, thus, was not peer-reviewed; or (5) the full text was not available.

Data were extracted from included studies using a standardized data collection form (Table S2) that collected information about the (1) study period; (2) region(s); (3) population; (4) the use of clinical criteria for enrollment; (5) whether participants with prior SARS-CoV-2 infection were included, and the corresponding proportion of participants with infections if available; and (6) method to determining vaccination status. For each study, confounder-adjusted VE estimates with CIs were extracted separately for each endpoint (eg, infection, hospitalization), vaccine, and the circulating virus. In some studies, VEs specific to time intervals after vaccination were reported. Therefore, we extracted VE estimates for the first available time interval at least 14 days after vaccination (ie, 14-30 days after vaccination), because antibodies have been shown to peak by then in naïve persons.²⁸ For studies that reported multiple estimates, such as by age group or type of vaccine, all subgroup-specific estimates were included, but the overall estimates were excluded.

Study quality was assessed by using the Risk of Bias in Non-Randomized Studies of Interventions (ROBINS-I) tool.²⁹ The certainty of evidence presented in an article included that the meta-analysis was graded using the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) approach.³⁰

Meta-analysis

In all identified studies, VE was defined as 100% × (1 − odds ratio [OR]). The extracted VE estimates were meta-analyzed to estimate pooled VE. The VE estimates were transformed to the ORs scale, meta-analyzed, then back-transformed to the VE scale for interpretation. The pooled OR was estimated by random-effects meta-analyses using the inverse variance method and restricted maximum likelihood estimator for heterogeneity.^31‑34 Heterogeneity was assessed using Cochran’s Q and the I² statistic.³⁵ We considered an I² value >75% to be indicative of high heterogeneity.³⁶ We also conducted a sensitivity analysis using fixed-effects meta-analyses.

The main study feature of interest was inclusion or exclusion of participants with prior infection. Severe disease was based on whether the estimate was limited to patients who required hospitalization, who required intensive care unit admission, and those who died. Otherwise, the estimate was classified as VE against infection, which referred to estimates of VE against test-positive or symptomatic infection (without hospitalization).

Pooled estimates were additionally disaggregated by the probable circulating virus and vaccine administered. Most studies did not report variant-specific VE estimates but did report study periods and the general prevalence of variants during that period. If this information was not reported, we used the Nextstrain³⁷ to determine the dominant variant in the study period in the study location. The variant with the highest proportion was considered to be the dominant variant. Therefore, estimates were grouped according to the predominant circulating virus, as follows: (1) Omicron; (2) late-Delta, which was the period with co-circulation of Delta and Omicron; (3) Delta; and (4) pre-Delta, which included ancestral strains and variants preceding Delta. The type of vaccine was categorized as follows: (1) mRNA vaccines, including vaccines produced by Moderna and Pfizer-BioNTech; (2) adenovirus vector vaccines, including vaccines produced by AstraZeneca, Janssen, and Gamaleya; and (3) inactivated virus vaccines, including vaccines produced by Sinovac Biotech and Sinopharm.

Meta-regression

To evaluate the impact of preexisting immunity on VE estimates, we used a meta-regression approach. Three proxies of prior immunity were explored: (1) inclusion vs exclusion of participants with prior infection; (2) cumulative incidence of COVID-19 since December 2019 in each of the study countries or regions before the study; and (3) the incidence rate of COVID-19 in the country or region during the study period. For this, we downloaded population denominator data and daily COVID-19 case data from the World Health Organization website.^38,39 We first used correlation analysis, including Pearson (r) and Spearman (ρ) correlation coefficients, to determine the association between preexisting immunity and VE estimates. Meta-regression models were adjusted for age group (all ages, child only, adult and elderly, elderly only), types of vaccines used, predominant circulating virus, the use of clinical criteria for enrollment, the time interval after vaccination (14-30 days after vaccination, unrestricted), and the method to determine vaccination status. A sensitivity analysis was conducted to additionally adjust for location and duration of the study.

The fitted meta-regression model estimated the ratio of ORs (ROR) for each of the prior immunity proxies explored. On the OR scale, values closer to 0 indicated a more effective vaccine, and values closer to 1 indicated a less effective vaccine. This was counter to the VE scale, on which values closer to 0 indicated an ineffective vaccine. Therefore, using inclusion vs exclusion of participants with prior infection as an example, if ROR > 1, then the OR estimated from studies including participants with prior infection was higher than that from studies excluding participants with prior infection. On the VE scale, this translates to lower VE for studies that included participants with prior infection than studies that excluded these participants.

We plotted the expected change in VE estimate to visualize the impact of each prior immunity proxy based on the ROR obtained from meta-regression. To illustrate the change in VE scale, we showed the change in estimate based on the ROR assuming VE for the reference group of 80% against infection and 90% against severe disease. Statistical analyses were conducted using R, version 4.0.5 (R Foundation for Statistical Computing), using the metaphor for meta-analyses and the robvis package for risk-of-bias visualizations.

Results

We identified 6904 studies, among which 2929 were duplicates. Title and abstract screening of the remaining articles identified 479 for full-text review, of which 66 met our inclusion criteria^{4‑7,40‑101} (Figure 1; Tables S3-S8). Studies were set in 17 countries or regions. Most were from the United States²⁸ and United Kingdom.¹⁰ Fifty-one studies provided 173 VE estimates against infection (Figure 2), and 40 studies provided 91 estimates against severe disease (Figure 3). Among all 66 studies, 40 included and 28 excluded participants with prior COVID-19 infection (including 2 studies that provided VE estimates including and excluding participants with COVID-19 infection^4,88). A summary of study characteristics and the corresponding number of estimates, including handling of participants with prior infections, enrollment criteria, vaccine types, and circulating virus are provided in Tables S3-S7.

Figure 1.

Selection of studies for the systematic review. VE, vaccine effectiveness.

Figure 2 .

Continues

Figure 3 .

Continues

Figure 2.

Estimates of vaccine effectiveness (VE) against infection from identified studies that excluded (A) or included (B) participants with COVID-19 infection history. Multiple estimates could be due various factors. These are labeled as follows: for vaccine, a(1) for Pfizer, a(2) for Moderna, a(3) for ChAdOx1, or a(4) for rAd26-rAd5 (Sputnik V); as for the control, b(1) for syndrome-negative), and (b2) for test-negative control; and for different duration between first and second dose: (c1) for 19-29 days, (c2) for 30-44 days, (c3) for 45-64 days, (c4) for 65-84 days, and (c5) for ≥85 days. VOC, variant of concern.

Figure 3.

Estimates of vaccine effectiveness (VE) against severe disease from identified studies that excluded (A) or included (B) participants with COVID-19 infection history. Multiple estimates could be due to various factors. These are labeled as follows: for vaccine, a(1) for Pfizer, a(2) for Moderna, a(3) for ChAdOx1, or a(4) for rAd26-rAd5 (Sputnik V); as for the control, b(1) for syndrome-negative), and (b2) for test-negative control; and for different duration between first and second dose: (c1) for 19-29 days, (c2) for 30-44 days, (c3) for 45-64 days, (c4) for 65-84 days, and (c5) for ≥85 days. VOC, variant of concern.

VE against infection and severe disease

The 173 VE point estimates against infection ranged from 14% to 98%, with I² = 100%, indicating considerable heterogeneity (Figures 4 and 5). Ninety-five VE estimates (55%) were higher than 80%. The 91 VE point estimates against severe disease were also considerably heterogeneous (I² > 97%), ranging from 20% to 100% (Figures 4 and 5). Among them, 70 (77%) were higher than 80%. For both outcomes, we observed declining VE over time from early 2021 to mid-2022 (Figure S2).

Figure 4.

Vaccine effectiveness (VE) point estimates from identified studies based on prior infection (A), predominant circulating virus (B) and vaccine type (C). Each point represents the VE point estimate. Estimates are jittered to enable visualization. Black points represent the pooled VE estimate from meta-analysis, with black lines representing the 95% CI around the pooled estimate. The shaded area is the violin plot, showing the smoothed density of the VE point estimates. VE estimates against infection and severity disease (D) vs the cumulative incidence of cases before the study (in log scale) (I) in the study regions are shown; VE estimates against infection and severity disease (E) vs the incidence rate of cases during study (in log scale) (J) in the study regions are shown. Pearson (r) and spearman (ρ) correlation coefficients are provided.

Figure 5.

Pooled vaccine effectiveness (VE) estimates against infection and severe disease by circulating virus, vaccine types, and the inclusion or exclusion of participants with prior COVID-19 infection from random-effect meta-analysis.

Impact of type of vaccine and circulating viruses

Our meta-analysis (Figure 5) indicated that pooled VE against infection for a primary course of mRNA vaccines was 86% (95% CI, 84-88), compared with 69% (95% CI, 64-73) for adenovirus vector vaccines and 67% (95% CI, 34-84) for inactivated virus vaccines. When we examined differences in pooled VE by the circulating virus, we found that VE against infection during the Omicron period was far lower (52%; 95% CI, 45-59) than during the pre-Delta (89%; 95% CI, 87-91), Delta (78%; 95% CI, 58-88), and the late-Delta (79%; 95% CI, 74-92) periods. Similarly, VE against severe disease during the Omicron period was 64% (95% CI, 52-73), which was lower than for pre-Delta (92%; 95% CI, 89-94), Delta (87%; 95% CI, 76-93), and late-Delta (91%; 95% CI, 88-93) periods. The results were similar when further disaggregated by including or excluding prior infection (Table S9) or using fixed-effects analysis (Figure S3).

Role of prior infection on VE estimates

In general, VE estimates derived from study participants with lower preexisting immunity were higher. The pooled VE against infection for studies that excluded participants with prior COVID-19 infection was higher (87%; 95% CI, 84-89) than from studies that included these participants (76%: 95% CI, 71-81). Similarly, pooled VE against severe disease from studies that excluded participants with prior COVID-19 infection (94%; 95% CI, 92-95) was higher than from studies that included these participants (86%: 95% CI, 82-88), with considerable heterogeneity among the estimates (I² > 99%).

In meta-regression adjusting for vaccine type, circulating virus, and enrollment criteria (Table S10; Figure 6, Figure S4), the OR against infection from studies that included participants with prior COVID-19 infection was 1.62-fold higher (95% CI, 1.32-1.98) than the OR from studies that excluded these participants (ie, with generally lower preexisting immunity). Therefore, if the VE against infection in a study that originally excluded participants with prior COVID-19 infection was 80%, the expected estimate was 68% (95% CI, 60-74) had they included those participants. Similarly, the OR against severe disease from studies that included participants with prior COVID-19 infection was 2.34-fold higher (95% CI, 1.60-3.40) than from studies that excluded these participants. Assuming a baseline VE against severe disease of 90%, the expected VE when participants with prior infection were included would be 77% (95% CI, 66-84). When using the proportion of individuals with prior infections, the results were similar to those obtained when the variable used was whether or not the study included participants with prior infections (Table S10). The results were similar with adjustment for location and duration of study (Table S11).

Figure 6.

Predicted vaccine effectiveness (VE) for a group of individuals based on the estimated ratio of odds ratios estimated from meta-regression, and the VE for the individuals in the reference group. Predicted VE against infection (A) and severe disease (B) are shown. Prior COVID-19 infection, cumulative incidence of cases before the study, and the incidence rate of cases during study are considered. Models adjusted for age group, type of vaccine, predominant circulating virus, and enrolment criteria, corresponding to the estimate in model 1.

Impact of population incidence in the study locations

There was a modest negative correlation between prestudy cumulative incidence of cases (Figure 4) as a proxy of preexisting population immunity (Figure S5), and VE against infection (r = −0.42 [95% CI, −0.54 to −0.30]; ρ = −0.32 [95% CI, −0.45 to −0.18]) and severe disease (r = −0.41, 95% CI, −0.56 to −0.22; ρ = −0.49 [95% CI, −0.64 to −0.30]). There was also a modest negative correlation between the during-study incidence rates of cases with VE against infection (r = −0.38 [95% CI, −0.50 to −0.24]; ρ = −0.48 [95% CI, −0.59 to −0.34]) and severe disease (r = −0.50 [95% CI, −0.64 to −0.33]; ρ = −0.42 [95% CI, −0.58 to −0.23]).

The Pearson and Spearman correlation coefficients were modestly negative between prestudy cumulative incidence of cases (Figure S5), and VE against infection (Figure 4D) and severe disease (Figure 4E). Similarly, we estimated a negative correlation between during-study incidence rates of cases (Figure S5), and VE against infection (Figure 4I) and severe disease (Figure 4J).

In meta-regression, adjusting for vaccine type, circulating virus, and enrollment criteria (Table S10; Figure 6, Figure S4), the ROR against infection associated with twice prestudy cumulative incidence of cases was 1.13 (95% CI, 1.05-1.23). Therefore, if the baseline VE against infection from a study was 80%, then the corresponding VE for a setting with twice prestudy cumulative incidence of cases would represent a 2-percentage point reduction in VE (VE = 77%; 95% CI, 75-79 for the initial doubling). The ROR against severe disease for each doubling of prestudy cumulative incidence of cases (higher preexisting immunity) was 1.53 (95% CI, 1.31-1.77). Assuming a baseline VE against severe diseases of 90%, the corresponding VE for a setting with twice the prestudy cumulative incidence of cases would represent a 5 percentage point drop in VE (VE = 85%; 95% CI, 82-87 for a doubling).

Similarly, we estimated that the ROR against infection for each doubling of during-study incidence rate of cases was 1.16 (95% CI, 1.07-1.26). If the baseline VE against infection from a study was 80%, then the corresponding VE from a study with twice during-study incidence rate of cases would be 77% (95% CI, 75-79). We also estimated that the ROR against severe disease associated with twice during-study incidence rate of cases was 1.22 (95% CI, 1.04-1.43). Therefore, assuming a baseline VE of 90% against severe disease, the corresponding VE for a study with twice during-study incidence rate of cases would be 88% (95% CI, 86-89.6).

Risk of bias

Most studies included in the meta-analysis were judged to be at moderate risk of bias (analyses of VE against infection: 41 [80%] of 51 studies; analyses of VE against severity: 33 [83%] of 40 studies; Figures S6-S7). Eight studies (16%) included in the analyses of VE against infection and 5 studies (13%) in analyses of VE against severity were judged to be at a serious risk of bias. The main sources of bias were potential confounding, bias in the classification of interventions due to self-reported vaccination status, and bias due to missing data (Figures S6-S7). We conducted a sensitivity analysis that removed estimates from studies classified as having serious or critical bias (Figure S8; Table S12), and the impact of prior infections was still similar to one estimated in the main analysis.

Discussion

In this study, we summarized VE estimates from TND studies to understand the impact of prior infections on VE estimates. We found that higher preexisting immunity in the source population, indicated by including participants with prior COVID-19 infection, higher prestudy cumulative incidence of cases, and higher during-study incidence rate of cases, was associated with lower VE.

Prior infection could be a confounder, effect modifier, or both. As a confounder, it could affect a person’s decision to vaccinate and modify their risk behaviors as well as provide protection against reinfection.^12,13,21 Hence, the VE obtained from individuals with or without prior infection would be similar if the influence of the confounding could be controlled in analysis. On the other hand, if prior infection were only an effect modifier (ie, only associated with the risk of (re)infection and not the propensity to be vaccinated), vaccination in settings with higher preexisting immunity would appear to have a relatively modest effect on further increasing protection at the population level, because VE would be lower among previously infected participants, compared with VE among uninfected participants.^12,13 This scenario differs from hybrid immunity,^102,103 which refers to the scenario in which a combination of naturally acquired and vaccine-induced immunity provides stronger protection against SARS-CoV-2 infection than vaccine-derived immunity alone (estimated only on the basis of vaccinated participants).

In reality, prior infection is probably both a confounder and effect modifier and, therefore, studies should consider both an appropriate confounding control, such as through covariate adjustment or stratification, as well as inclusion of interaction terms to explore the potential effect modification. The impact of prior infection could be even larger than observed in our study, because not all prior infections may be documented and, therefore, some participants may have been unknowingly infected and misclassified as infection-naïve. As the pandemic progresses, the proportion of undocumented previous infections could grow. This is likely to be particularly problematic for prior Omicron infections because that variant has lower severity than previous strains.^104,105

When VE was estimated on the basis of studies excluding participants with prior infection, these VE estimates should be interpreted as the VE for a hypothetical population with no preexisting immunity. As of 2023, these estimates have limited practical value because most locations have experienced substantial epidemics. Epidemic forecasting models used to inform public health control policies should separate individuals into different compartments based on infection history to improve the precision of their forecasts. Therefore, groups estimating VE estimates to inform policy should stratify by infection history so their work will be more broadly useful for policy.^106,107

Although 55% of VE estimates against infection and 76% of estimates against severe disease were higher than 80%, heterogeneity was considerable, based on high I² values. Consistent with previous reviews, high heterogeneity could be attributed to differences in effectiveness among vaccine types or the predominant circulating virus in each study.^8,108 However, we continued to observe high heterogeneity when estimating pooled VE against specific vaccine types and the predominant virus. Our meta-regression identified some sources of the heterogeneity, such as preexisting immunity. Further investigation is needed to identify other causes to ensure valid VE estimates are available for ongoing optimization of vaccination strategies.¹⁰⁹

Our review focused on VE of primary vaccination series to answer the research question of whether prior infection may have an impact on VE estimates. Further analysis would be required to determine whether similar issues apply to estimation of VE for booster doses, which are complicated by dosing schedules that mix vaccine types, the number of doses received, greater antigenic differences between the vaccines received and the dominant circulating virus, changes in vaccine formulation including bivalent formulations, and the accumulation of immunity through both vaccination and infection over time.

Our study had some limitations. First, most studies were conducted with adults, so our results may not be generalizable to children. Second, TND studies included in our review were observational. Some confounders were adjusted for in these studies, including age, sex, being a health care worker, or preexisting conditions. Despite our bias assessment to evaluate whether studies adequately addressed confounding and considered other potential sources of bias, such as measurement errors and selection bias, we cannot rule out other unidentified sources of bias. Third, the predominant variant was determined by Nextstrain³⁷ for some studies, which can be inaccurate. Fourth, multiple group-specific estimates reported from studies were included in the analysis (Table S5). If a study reported estimates for multiple factors, such as age group–specific VE estimates and vaccine type–specific estimates, there could be correlations among VE estimates from the same study. Finally, we used incidences of COVID-19 in study countries as a proxy for preexisting immunity. However, we acknowledge that national statistics may not apply to individual study sites.

In conclusion, we observed reduced VE associated with higher preexisting immunity in the population, which is likely to act both as a confounder and effect modifier of the vaccine’s effect. Exclusion of participants with prior infection could artificially inflate VE estimates and affect their generalizability to the wider population. If the goal of a study is to inform policy that applies to the whole population, participants with prior infection should be included and their status included as a covariate for confounder control. However, if decision-makers desire different vaccination policies dependent on infection history, then studies need to stratify or include interaction terms, rather than exclude participants with prior infection. In studies in which prior infection cannot be adjusted for, researchers could consider using external adjustment¹¹⁰ to assess the potential effect of this confounder on their estimates. Optimal design of VE studies remains a research priority. In particular, more work is needed to understand how prior infection influences VE for booster doses and as vaccine formulations change.

Supplementary material

Supplementary material is available at American Journal of Epidemiology online.

Funding

This project was supported by the National Institute of General Medical Sciences (grant R01 GM139926), and the Theme-based Research Scheme (project T11-712/19-N) of the Research Grants Council of the Hong Kong SAR Government. B.J.C. is supported by a Research Grants Council Senior Research Fellowship (grant HKU SRFS2021-7S03) and the AIR@innoHK program of the Innovation and Technology Commission of the Hong Kong SAR Government. The WHO Collaborating Centre for Reference and Research on Influenza is supported by the Australian Government Department of Health and Aged Care.

Conflict of interest

B.J.C. reports honoraria from AstraZeneca, Fosun Pharma, GSK, Haleon, Moderna, Pfizer, Roche, and Sanofi Pasteur. J.N. was previously employed by and owns stocks in Sanofi. S.G.S. reports honoraria from CSL Seqirus, Novavax, and Pfizer. The other authors declare no conflicts.