Background
Influenza hemagglutinin (HA) binds to sialic acid receptors on target cells and is the main target of neutralizing antibodies [
1]. As such, HA is the primary standardized antigen in the inactivated influenza vaccine (IIV) [
2]. Serum levels of anti-HA antibodies are measured by the hemagglutination inhibition (HAI) assay. When assay targets are well matched to circulating viruses, there is generally a clear relationship between increasing HAI titer and decreasing infection risk [
3‐
9]; however, some individuals acquire infection or experience influenza disease despite high HAI titers [
4,
10‐
12] or seroconversion [
12]. Neuraminidase (NA) is also present in most influenza vaccines, albeit at nonstandardized concentrations [
13]. As measured by neuraminidase inhibition (NAI) assays, anti-NA antibodies play an independent role in protection from influenza disease and/or in reducing influenza disease severity [
14,
15].
Antigenic drift [
16‐
18] necessitates annual evaluation of influenza vaccine composition [
19]. To optimize the immune response to strains most likely to cause infection, annual influenza vaccination is currently recommended in the United States for all individuals aged ≥6 months (unless contraindicated) [
20]. However, reduced seroresponse can occur after repeated influenza vaccination [
21‐
26]. While this suggests that repeated vaccination may have diminishing benefit to protect against influenza infection or disease, reduced vaccine effectiveness with repeated vaccination has only been observed in some, primarily A(H3N2) predominant, seasons [
27‐
35]. These apparently inconsistent results could be explained by the degree of antigenic relatedness between vaccine and circulating viral strains [
36,
37].
The principal stratification/vaccine efficacy (VE) moderation framework [
38‐
40] (hereafter called the “VE moderation” framework) is a statistical method for assessing how vaccine efficacy varies over subgroups defined by biomarkers measuring immune responses in vaccinees. This framework requires data from a randomized-controlled trial with sufficient endpoint cases and ample immune response measurements from vaccinees in cases and non-cases, as well as variables measured at baseline in both vaccine and placebo recipients (cases and non-cases) that are predictive of post-vaccination immune responses. The latter requirement enables the baseline immunogenicity predictor (BIP)-augmented efficacy trial design [
39,
40] that predicts the immune response to the vaccine that placebo recipients would have had had they been vaccinated, allowing estimation of the VE-by-postvaccination-titer curve. Using this framework, fold-rise in anti-varicella zoster virus (VZV) titers was shown to be strongly associated with VE against herpes zoster disease, whereas post-vaccination titers at 6 weeks were not, implying that baseline titers must be measured to predict VE [
41]. This framework also showed that post-vaccination neutralization titers were strongly associated with VE against virologically confirmed dengue [
42]. We applied this framework to data collected as part of a randomized, placebo-controlled trial of the absolute and relative efficacies of IIV and live-attenuated influenza vaccine (LAIV) [
43] to assess how VE against laboratory–confirmed influenza disease varied across subgroups defined by their HAI or NAI responses to vaccination. We also studied how VE varied with baseline/pre-vaccination data on HAI, NAI, and relevant clinical variables.
Methods
Study design and intervention
The FLUVACS trial enrolled healthy adults aged 18 to 49 years, excluding individuals with a health condition for which vaccination was specifically recommended (including being immunocompromised or being older than 49 years) or for which either vaccine was contraindicated. Participants were recruited from October to November 2007 and randomly assigned to receive IIV (Fluzone, Sanofi Pasteur), LAIV (FluMist, MedImmune), or saline placebo. Surveillance for influenza-like illness was carried out from November 2007 through April 2008. Overall IIV and LAIV efficacy were 68% and 36%, respectively [
43].
Influenza endpoint, cases and controls
The study endpoint was laboratory–confirmed influenza disease, defined as symptomatic acute respiratory illness subsequently confirmed by RT-PCR influenza virus identification [
43]. Participants with an observed endpoint are referred to as cases and participants completing follow-up (with a post-season sample) without experiencing the endpoint are referred to as controls.
For antibody response measurements, serum samples were collected at Day 0 (immediately before intervention administration), Day 30 (approximately), and at the influenza season conclusion (approximately 4 months later). HAI titers were measured in a subset of all participants (including participants in each treatment arm) consisting of all cases and a random sample of participants for whom all 3 serum samples were available [
14]. NAI titers were measured in all cases and in a smaller sub-sample of controls [
14].
HAI and NAI titer variables
The HAI assay measures the highest dilution of serum that prevents influenza virus-induced hemagglutination of erythrocytes [
44]. The reciprocal of this dilution was defined as the HAI titer. In the lectin-based NAI assay [
45], the reciprocal of the highest dilution of serum that inhibits NA activity at least 50% compared to control wells was defined as the NAI titer. Titers below the lower limit of quantification for both assays were set to half this value [
14]. For HAI titers in graphs,
log2(
x) for
x=0,1,2, etc. corresponds to titer (2
x)∗4, and for NAI titers
log2(
x) corresponds to titer (2
x)∗5. Fold-rise in HAI or NAI titer was defined as (Day 30 titer)/(Day 0 titer).
Vaccine efficacy parameters
Overall vaccine efficacy (VE) for either vaccine versus placebo was defined as the multiplicative reduction in the probability of influenza disease occurrence:
$$\text{VE} = 1 - \frac{P(\text{influenza disease}|\text{vaccine})}{P(\text{influenza disease}|\text{placebo})}.$$
VE for a vaccinated subgroup defined by a fixed value
s of the Day 30 or fold-rise in titer was defined as
$$\text{VE}(s) = 1 - \frac{P(\text{influenza disease}|\text{vaccine, titer} s)}{P(\text{influenza disease}|\text{placebo, titer} s)}.$$
The critical feature of the VE moderation framework is that the subgroup
s is defined under assignment to a vaccine group, which is observable for participants actually assigned to a vaccine group, but is counterfactual and hence missing for participants assigned to the placebo group [
39,
40,
46]. Implementation of this framework for estimating VE over subgroups
s requires prediction of the missing counterfactual responses of placebo recipients [
41].
For studying the association of baseline variables with VE, the VE parameter of interest is one minus the ratio (vaccine/placebo) of the disease incidence in subgroups defined by fixed values of the baseline variables.
Objectives
We investigated: (1) if VE varies by Day 0 HAI or NAI titer and/or by other baseline clinical variables; (2) if VE varies by fold-rise HAI titer and by fold-rise NAI titer; (3) if VE varies by Day 30 HAI titer and by Day 30 NAI titer. Baseline clinical variables included age, sex, race, and whether the individual self-reported ever having received an influenza vaccine before (hereafter referred to as “previously vaccinated”).
Statistical analysis
Boxplots and scatterplots with Spearman rank correlations describe the Day 0, 30, and fold-rise HAI and NAI titer distributions. These distributions were compared across groups (treatment arms, vaccination history) using Wilcoxon rank-sum tests, and across age levels by Spearman rank correlations.
Objective (1) was addressed by multivariable logistic regression (with sandwich variance estimates) modeling of how the risk of influenza outcome depended on titer variables and clinical variables, which is valid under the case-control sampling design [
47]. Forward selection stepwise regression based on Wald tests was used to select best-fitting models.
Objectives (2) and (3) were addressed using the same or similar statistical methods as in [
41]. The influenza disease endpoint was treated as a dichotomous outcome (case vs. control); time-to-event methods would not add value given all vaccinations were completed from 10/10/2007 to 11/09/2007, all influenza events occurred between 1/10/2008 and 3/09/2008, and 96.4% of participants completed all scheduled visits in this year.
For Day 30 or fold-rise variables treated as quantitative variables, the Juraska et al. method [
48] was applied, which was also used in [
42]. We describe the method for Day 30 HAI titer; the same method is used swapping in Day 30 NAI titer, fold-rise HAI titer, and fold-rise NAI titer. This method specifies a structural logistic regression model
$${} P(Y(z)\,=\,1|S(z)=s,X=x) \,=\, expit \left(\beta_{0z} \,+\, \beta_{1z} S(z) + \beta_{2z} X \right) $$
where
Y(
z) is the indicator of the influenza endpoint if assigned treatment
z,
z=0 (placebo) and
z=1 (vaccine),
expit(
a)=
exp(
a)/(1+
exp(
a)),
S(
z) is Day 30 HAI titer if assigned treatment
z (
z=0,1), and
X is the baseline covariate age in years at enrollment. Inverse probability-weighting is used in the logistic regression model is used to account for the probabilities participants have the titer data measured. The method incorporates an estimate of the conditional density of
S(1) given
Sb and
X, obtained by nonparametric kernel smoothing with optimal bandwidths selected by likelihood cross-validation [
49], where
Sb is Day 0 HAI titer. A main assumption of this method is that the risk of
Y(0)=1 is conditionally independent of
S(1) given
S(0) and
X, which in our application states that after accounting for age and the Day 30 HAI titer if assigned placebo, Day 30 titer if assigned vaccine does not contain additional information about influenza risk if assigned placebo.
The method outputs point estimates of VE(s), bootstrap-based 95% pointwise and simultaneous confidence bands for VE (s), and a bootstrap-based 2-sided p-value for testing whether VE (s) varies in s, for each of the IIV and LAIV versus placebo.
For regression analyses of objectives (1)–(3), quantitative response variables were analyzed on the log
2 scale. Analyses for objectives (2) and (3) were conducted using the pssmooth R package available at CRAN [
50].
Discussion
Higher pre-vaccination/baseline HAI titers were observed among the previously vaccinated than among the previously unvaccinated, while no difference in NAI titers was observed. These findings suggest that NAI titers decay faster than HAI titers after vaccination or are not boosted as much by vaccination as HAI titers. The former explanation is unlikely considering that post-vaccination HAI and NAI titers have been observed to wane at similar rates [
21,
51], while the latter explanation is supported by the fact that the NA concentration is not standardized in currently licensed influenza vaccines.
There was no evidence that LAIV efficacy depended on previous vaccination or baseline HAI or NAI titer. In contrast, IIV efficacy was significantly modified by previous vaccination and baseline NAI titers in an interactive way. In the previously vaccinated, estimated VE was near zero for individuals NAI seronegative at baseline and increased to about 75% for those with highest baseline NAI titers, whereas in the previously unvaccinated the opposite pattern was observed, with estimated VE about 75% for baseline NAI seronegative individuals and decreased to zero for those with baseline NAI titers at 40 or higher. This result seems to be driven by the placebo group, for which risk was not associated with baseline NAI titers in the previously vaccinated but was strongly inversely associated with baseline NAI titers in the previously unvaccinated. One possible explanation of this difference in risk profiles is that previously unvaccinated individuals with high baseline NAI titers represent individuals with sustained/durable NAI responses, perhaps obtained following natural infection, and there is lower risk of influenza disease. In contrast, for previously vaccinated individuals, the subgroup with high Day 0 NAI titers is a mixture of two subgroups: individuals who could have high NAI responses even without the previous vaccination, and individuals with a boosted NAI response from the previous vaccination, where the risk of influenza disease is not low. Alternatively, Day 0 NAI titers may lose their ability to accurately mark natural protection among individuals who have been previously vaccinated. The finding that previously unvaccinated individuals with high pre-vaccination HAI or NAI titers were unlikely to be protected by IIV may be related to multiple layers of protection conferred through naturally acquired immunity by prior influenza infection, operating not only through humoral responses but also possibly through cellular responses, which contrasts with the IIV-conferred protection that mainly operates via humoral responses to the HA and NA surface antigens. Moreover, the different result for the LAIV vaccine – no evidence of modification of LAIV vaccine efficacy by prior vaccination or by post-vaccination HAI or NAI titer – may strengthen this point, given that LAIV mimics natural immunity. Given the limited precision in the analysis, it would be important to confirm the effect modification findings in other studies. A caveat of this study was that data were not collected on recent influenza history, nor on recent vaccination, precluding the ability to study any moderating impact of recent natural or vaccine immunity on the HAI/NAI correlates of risk and of vaccine efficacy.
IIV efficacy significantly increased within strata of increasing Day 30 NAI titers observed for vaccinated individuals and predicted counterfactually for placebo recipients, and similarly seemed to increase with Day 30 HAI titers. In terms of point estimates, LAIV efficacy trended toward decreasing with both Day 30 NAI titers and Day 30 HAI titers, but with inadequate precision to infer a real decrease – overall there was no statistical evidence that LAIV efficacy correlated with any post-vaccination titer markers.
Mathematical modeling has been one approach to try to quantify the relationship between HAI titer and VE [
3]. Such modeling has predicted that vaccines that elicit higher HAI responses should have higher VE against laboratory-confirmed influenza disease [
52]. Influenza vaccine efficacy studies have also supported that post-vaccination HAI titers are positively associated with VE [
12,
51,
53], although it has been proposed that such titers should not be used as a surrogate endpoint for reliably inferring VE due to the potential importance of cell-mediated immunity and anti-NA antibodies [
12,
14]. Using the Prentice (1989) [
54] surrogate endpoint framework with Dunning et al.’s (2015) scaled-logit model [
55], Dunning et al. (2016) [
56] analyzed a trial of the IIV vaccine at standard dose (as in FLUVACS) vs. high dose in persons 65 years of age and older and inferred that post-vaccination HAI titer of 40 corresponded to 50% protection (identical for vaccine- and natural-immunity for a valid Prentice surrogate) when the assay virus was antigenically well-matched to the circulating virus. Based on our VE moderation framework applied to FLUVACS, an HAI titer of 40 corresponded to estimated IIV VE of about 55% and an NAI titer of 40 corresponded to estimated VE of about 55-60% (Fig.
5), remarkably close to the previous estimates. Thus, use of the VE moderation framework may provide some additional confirmation that a post-vaccination titer response above 40 predicts reasonably high vaccine protection, complementing the result from the Prentice approach. As Dunning et al. did not analyze whether or how post-vaccination HAI fold-rise in titer corresponded to protection, the relative performance of post-vaccination absolute titer vs. fold-rise as correlates of protection in this framework cannot be compared.
The pattern of estimated VE as a function of Day 30 HAI titer in vaccinees was similar to that of the pattern with Day 30 NAI titer, and the analysis did not provide statistical support that one marker was a stronger correlate of VE. While our analysis lacked the statistical precision to discern which correlate was stronger, the fact that the point estimates showed a stronger association for NAI titer may suggest that NAI is as important as HAI. Had the association patterns been largely different by assay the study could have detected it– future studies with greater power would be needed to detect small-to-moderate differences. In addition, our inability to distinguish the predictive power of HAI and NAI titers may have been impeded to some degree by the modest correlation between these two antibody measures. Future studies could also assess how vaccine efficacy depends jointly on HAI and NAI titer levels, as there was insufficient sample size to support these analyses in the present study.
We next discuss some limitations of this work. First, as influenza vaccination was specifically recommended to individuals aged ≥50 years [
57], ethical considerations precluded the enrollment of such individuals in the study, given that some of these individuals would have been randomized to placebo. Thus, we were unable to examine whether our findings held true in older individuals. This is an ongoing issue for all clinical trials of influenza vaccines, given the current recommendations. Given the evidence that T-cell responses may be a better correlate of protection against influenza disease than antibody titers in the elderly (≥65 years; reviewed in [
58]), it is reasonable to hypothesize that we may see differences in the associations of HAI and/or NAI titer with VE, particularly in the elderly; however, there are considerable ethical and methodological challenges involved in testing this hypothesis.
In addition, a limitation of the VE moderation statistical methods used here is that they could give biased results if certain assumptions are violated [
40,
46]. The methods assume that within each treatment arm the risk of the influenza endpoint conditional on the baseline and post-baseline biomarker follows a regression model, which can be directly checked. However, the methods also assume that, for placebo recipients, after conditioning on baseline demographic variables and the baseline and post-baseline biomarker values, influenza risk does not depend on the post-baseline biomarker value that would have been observed had the placebo recipient been vaccinated – this assumption cannot be validated because the vaccine response biomarker is counterfactual/missing. Undiagnosable violations of this assumption could cause bias in the estimated VE curves. A common misperception is that the VE moderation methods require perfect prediction of the counterfactual vaccine responses for placebo recipients; in fact, the methods were designed to account for uncertainty in the prediction and have been shown to provide valid results with partial predictiveness.
Another limitation of our analysis is that there were too-few influenza events to estimate VE by post-vaccination titer curves separately for the previously vaccinated and the previously unvaccinated. Given the significant interaction result that Day 0 NAI positively associated with IIV VE in the previously vaccinated yet negatively associated with IIV VE in the previously unvaccinated, and the positive association of Day 0 and Day 30 titers, we conjecture that the positive association of Day 30 titers with IIV VE may have been driven by the previously vaccinated subgroup. The design of future studies should consider planning for the enablement of such analyses with adequate statistical power.