Background
Neisseria meningitidis is a strictly human bacterium that can cause invasive meningococcal disease (IMD). The most frequent clinical manifestations of IMD are meningitis and sepsis, both serious and rapidly fulminant conditions associated with considerable mortality and sequelae worldwide [
1‐
3]. Six serogroups (A, B, C, W, X, and Y) are responsible for most IMD cases, and the disease incidence is generally highest in infants, showing geographically asynchronous secular trends leading to substantial temporal variability in the number of cases [
2,
4]. Serogroup B is currently a major cause of IMD in the Americas, Australia and Europe [
5].
Pre-licensure efficacy trials for meningococcal vaccines are not feasible because the incidence of IMD is low. Hence, licensure of the four-component serogroup B meningococcal 4CMenB vaccine (
Bexsero, GSK) was based on studies testing its safety and immunogenicity, which is measured through a serum bactericidal assay with human complement that correlates with protection [
6,
7]. Accordingly, real-world studies of vaccine effectiveness (VE) and impact are deemed very important.
The first such real-world study looking into 4CMenB vaccine effect was conducted by Public Health England (PHE). In September 2015, the United Kingdom became the first country to include 4CMenB in its national immunization program offering a reduced schedule of three doses of 4CMenB to all infants: two primary doses at 2 and 4 months of age and a booster at age 12 months [
8]. After 3 years of 4CMenB vaccination, PHE reported a very high vaccine uptake (92.5% for the primary two-dose immunization and 87.9% for the third dose) and a statistically significant 75% reduction of the incidence of serogroup B IMD (95% confidence interval [CI] 64%; 81%). This vaccine impact (VI) was estimated on age groups fully eligible for vaccination (that also included non-vaccinated and partially vaccinated subjects) through a Poisson model [
9].
In contrast, the same study reported VE estimates that were lower than the VI and not significantly different from zero (e.g., in subjects fully vaccinated with three doses, 59.1% [− 31.1%; 87.2%]
95%CI) [
9]. VE was estimated using the screening method, an approach that expresses the effectiveness as a function of the proportion of cases that has been vaccinated and the vaccine uptake in the population [
10‐
12]. VE was estimated from fully vaccinated cases compared with non-vaccinated cases eligible for three doses. VI was estimated from a population that included the same cases used for the VE, with in addition, other partially and non-vaccinated subjects and cases. Therefore, theoretically, VI should not exceed the VE in fully vaccinated. Indeed, when indirect effects of a vaccine are absent or negligible, especially at the beginning of an immunization program, VI should theoretically equal its effectiveness multiplied by the proportion of vaccinated persons (i.e., the vaccine uptake
\(x\)):
\(VI=VE*x\). This is a simple mathematical relation that is valid in the absence of trends in disease incidence and vaccine uptake but can give useful indications for VI and has been previously used to broadly estimate VE from impact and uptake [
13]. We provide its formal derivation and additional details on underlying assumptions in Additional file
1: Section S1. Since
\(x\) cannot be higher than one (i.e., 100% uptake), it follows that VE should be at least as high as VI. Although the measured 75% impact is reassuring of the fact that the immunization program with 4CMenB in England has been successfull, the undefined effectiveness (not significantly greater than zero) does not allow comparing the effectiveness of complete or incomplete vaccination and drawing conclusions about the applied vaccine schedule. Moreover, definite and precise VE estimates are necessary to predict the impact of a similar immunization program in other countries and to parametrize cost-effectiveness models.
The lack of significance for VE estimates warrants a rigorous re-estimation of 4CMenB effectiveness with greater precision. We analyzed real-world data on serogroup B IMD and 4CMenB uptake between September 2011 and August 2018 provided to us by PHE and re-assessed the effectiveness of one to three doses of 4CMenB using an incidence model.
Discussion
In this re-assessment we showed that the effectiveness of 4CMenB vaccine was 80.1% [70.3%; 86.7%]
95%BCI in fully vaccinated infants during its first 3 years of implementation in a national immunization program in England. We demonstrated that our VE estimates are in line with the previously reported VI: a 75% [64%; 81%]
95%CI reduction of serogroup B IMD incidence in age groups fully eligible for vaccination [
9].
The same study [
9] reported non-statistically significant VE, with point estimates more than 15% lower than the VI estimate, even in fully vaccinated subjects. They were calculated using the screening method, a simple and rapid approach where VE is estimated from the proportion of cases that are vaccinated and the proportion of the population vaccinated [
10‐
12]. The study was designed to estimate direct effects of the vaccine by including only cases from the vaccine-targeted population [
9], i.e., it a type I design [
12]. In our re-analysis, we used also cases from the non-eligible population, assuming that indirect effects to be negligible. In other words, our design falls between designs type I and type IIb, where the latter normally estimates direct and indirect effects comparing a vaccinated community with a separate unvaccinated community [
12].
The lack of precision of the estimates based on the screening method compared to our re-assessment may be because the screening method relies exclusively on cases emerging from the population eligible for vaccination. When the disease incidence is low and the vaccine uptake is high—as is the case in this study—the number of unvaccinated cases in the eligible population may be very low or even null (depending on the observational time), thus inadequate to produce sufficiently precise VE estimates [
21]. Instead, our incidence model used non-vaccinated cases from the eligible as well as from the non-eligible population to achieve higher precision, adopting the additional assumption that relative changes in incidence over time are the same in different age groups, an assumption that held for the data analyzed here.
Although all our point estimates for VE fall inside the respective 95% CIs obtained via the screening method, point estimates produced with the screening approach are consistently lower than our results. The reason for this difference may be another consequence of the small number of cases that already caused the lack of precision. Indeed, the screening method was applied by matching the proportion of vaccinated subjects to the observed cases, according to the number of doses received, age and year at disease onset. As the number of disease cases was extremely small compared to the population eligible to the campaign, the matched vaccine uptake may not have been representative of the vaccine uptake in the population, especially with both disease incidence and uptake per dose greatly varying with age, as observed in this immunization program. Therefore, the low number of cases and a possible unlucky uptake matching may have just for chance driven VE point estimates towards lower values. In theory, it could have also led to the opposite effect, i.e., an overestimated VE point estimate, depending on age and time at disease for the limited number of cases to which the uptake was matched. The screening method would have likely required more years of surveillance to have enough cases for a precise evaluation of the proportion of vaccinated population. Instead of only using uptake data relative to cases, we used more robust person-time data of the whole vaccinated and unvaccinated population (mostly composed of non-cases) and could estimate VE in line with the previously reported incidence reduction [
9].
Most relevant from a public health perspective is the overall number of averted cases. We estimated this to be 312 [252–369]
95%BCI, which is in good agreement with the 277 [236–323]
95%CI averted cases estimated by PHE [
9]
. Disease incidence decreased during the first two pre-vaccination years considered here, then slowly inverted its trend, started to rise and continuously increased—in non-vaccinated subjects—after the inclusion of 4CMenB in England’s infant national immunization program in September 2015 (Fig.
1, panel B). Thus the decision to include 4CMenB in the national immunization program was timely.
Although our VE estimates were more precise than those obtained with the screening method and agreed with the incidence reduction, our analysis has some limitations. The data on vaccination used in this re-assessment were originally collected to apply the screening method. Therefore, information on vaccination status at an individual level was collected for cases only. By design, the vaccination status of non-cases, i.e., controls, was derived from uptake statistics that came from a source external to the study [
11]. Also, data on disease that we received were case counts, already aggregated by age, year and doses. Even if we believe that the stratification was optimal for this kind of analysis, we could not test the sensitivity of the results when varying the stratification. Another possible limitation is that we could not test covariates other than age, time and doses received, to find and adjust for potential additional confounders. An intrinsic weakness of our model-based approach is that it assumes incidence rates are constant within each group. Most importantly, we assumed that yearly incidence variation would affect each age group proportionally. In case of a drastic variation of the incidence in specific age groups, as it may happen in case of IMD outbreaks, this model could not hold. In that case, individuals living in geographical areas or specific settings (e.g., schools) particularly affected by outbreaks may be excluded from the analysis, or more complex transmission models to predict incidence trends would be needed. However, we have verified that our model accurately reproduced the data, indicating that the model was appropriate for this context. The approach that we adopted assumes that possible indirect effects due to herd immunity are absent or negligible (as previously observed for this study) [
9], as the number of infants vaccinated up to August 2018 was low compared to the entire England’s population. Nevertheless, in case of vaccine-induced herd immunity, the incidence model and VE estimates could be affected by indirect effects.
Precise and accurate estimates of 4CMenB effectiveness are essential to appropriately inform cost-effectiveness analyses that support public health decision making [
22,
23]. Provided the availability of strain typing data [
24,
25] on the same disease cases used for this re-assessment, the higher precision of our approach would allow further stratifying data and to assessing 4CMenB effectiveness against different strain types. Such estimates, combined with strain typing data from other countries [
26], would in turn allow even more robust predictions of 4CMenB effects in different geographies, improving cost-effectiveness evaluations. In the future, our approach may be adopted in similar settings by using surveillance data at a population level during vaccination programs, if more traditional methods are underpowered.
Our results are also aligned with 4CMenB effectiveness in infants and children reported in two recent observational studies in Italy (two regions: Tuscany and Veneto) and Portugal [
27,
28]. VE estimates reported in these two studies, which were performed in settings with specifications different to England, confirm once again our re-assessment findings.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.