 Research article
 Open Access
 Published:
Assessing correlates of protection in vaccine trials: statistical solutions in the context of high vaccine efficacy
BMC Medical Research Methodology volume 19, Article number: 47 (2019)
Abstract
Background
The use of correlates of protection (CoPs) in vaccination trials offers significant advantages as useful clinical endpoint substitutes. Vaccines with very high vaccine efficacy (VE) are documented in the literature (VE ≥95%). The rare events (number of infections) observed in the vaccinated groups of these trials posed challenges when applying conventionallyused statistical methods for CoP assessment. In this paper, we describe the nature of these challenges, and propose easytoimplement and uniquelytailored statistical solutions for the assessment of CoPs in the specific context of high VE.
Methods
The Prentice criteria and metaanalytic frameworks are standard statistical methods for assessing vaccine CoPs, but can be problematic in high VE cases due to the rare events data available. As a result, lack of fit and the problem of infinite estimates may arise, in the former and latter methods respectively. The use of flexible models within the Prentice framework, and penalizedlikelihood methods to solve the issue of infinite estimates can improve the performance of both methods in high VE settings.
Results
We have 1) devised flexible nonlinear models to counteract the Prentice framework lack of fit, providing sufficient statistical power to the method, and 2) proposed the use of penalised likelihood approaches to make the metaanalytic framework applicable on randomized subgroups, such as regions. The performance of the proposed methods for high VE cases was evaluated by running simulations.
Conclusions
As vaccines with high efficacy are documented in the literature, there is a need to identify effective statistical solutions to assess CoPs. Our proposed adaptations are straightforward and improve the performance of conventional statistical methods for high VE data, leading to more reliable CoP assessments in the context of high VE settings.
Background
Assessing a vaccine’s ability to induce immune responses that can effectively protect from infection and disease is key. The use of clinical endpoints to assess vaccine efficacy (VE) can be burdensome on the development, licensure, duration and effectiveness monitoring of immunisation trials. Replacing the clinical endpoint of a vaccine by an immunological endpoint can positively impact many of these aspects and considerably reduce costs as a result, as well as facilitate ethical procedures. Indeed if measured appropriately, immunological endpoints are biomarkers that can accurately predict VE on a shorter time scale while using significantly fewer participants compared to clinical endpoint assessments, making them an attractive time and costeffective option [1].
The terms ‘correlate’ and ‘surrogate’ of protection are common in the literature when referring to immunological endpoints, but are often used inconsistently, including by regulators and other prominent authorities. The first formal definition of surrogacy was introduced by Prentice in 1989, and was complemented with a set of criteria based on the concept of mediation [2]. Several statistical methods for evaluating surrogate endpoints soon followed as part of the causal inference [3–5] and metaanalytic frameworks [6–8], on which Alonso et al. provided a useful description of their relationship [9]. A hierarchical framework was proposed by Qin et al. to shed clarity on the profuse topic of immune correlates, and to assess their validity as substitute endpoints [10]. In their proposal, three levels of association are distinguished: ‘Correlate of Risk’ (CoR) (1), level 1 ‘specific’ surrogate of protection (SoP) (2) and level 2 ‘general’ SoP (3), where levels 1 and 2 reflect whether the analysed data comes from single or multiple trials, respectively. Specifically, a level 1 (specific) SoP is an immunological measurement predictive of VE in the same setting as the trial in which the vaccine was investigated, while a level 2 (general) SoP refers to a surrogate that can predict VE across a range of different populations and settings [10]. Metaanalytic approaches have been proposed to evaluate level 2 SoPs using data collected from multiple trials [6–8].
Within level 1, Qin et al. further subdivide this SoP into a statistical or principal category, according to the method used for their validation. A statistical SoP is an endpoint that satisfies the Prentice criteria [2], while a principal SoP is defined using a causal inference framework [3–5, 10, 11]. The latter aims to address postrandomisation selection bias by estimating what the vaccine responses would have been if the nonvaccinated group of a trial had been immunised. Such endpoints can be used to predict VE once they are validated and approved by a regulatory body.
In this manuscript, SoP endpoints are referred to as correlates of protection (CoPs). Specifically, we address CoP levels 1 and 2, based on Qin et al.’s following definitions of a CoR as an "immunological measurement that correlates with the rate or level of a study end point used to measure VE in a defined population", and a CoP as a "CoR that reliably predicts a vaccine’s level of protective efficacy on the basis of contrasts in the vaccinated and unvaccinated groups’ immunological measurements" [10]. Moreover, we address the concept of CoPs in the context of a continuous, rather than a threshold approach [1].
Although not common, vaccines with very high efficacy (95% or above) are documented in the literature [12–17]. These include the salmonella typhi vi coniugate [12], or the combined measlesmumpsrubellavaricella immunisation [17]. These trials raised the problematic of assessing CoPs in the context of high VE using classical statistical methods. Indeed, a very small number of cases/infections (corresponding to the vaccinated groups) can trigger considerable issues for such statistical models. There is therefore a need to adapt statistical methods for CoP assessment to the context of high efficacy vaccines. To the best of our knowledge, such tailored approaches are lacking in the literature. The aim of this manuscript is to present statistical solutions and to generate adapted methods to assess CoPs based on Prentice criteria and metaanalytic frameworks (by randomized subgroups such as centers and regions) in single trial setting (STS) with high VE.
Methods
Statistical methods for assessing CoPs
The Prentice criteria and metaanalytic approach are two classical statistical methods used for assessing vaccine CoPs. The following sections describe both methods, and our specific adaptations as statistical solutions for high VE settings. The results section shows the performance of our proposed adapted models using simulations.
The prentice criteria
The following set of notations will be used throughout the manuscript: T_{j} and S_{j} are random variables denoting the true binary and the surrogate endpoints for subject j=1,...,n and Z_{j} is a binary treatment indicator.
Key concepts, including the hypothesistesting approach to the validation of substitute endpoints using randomised clinical trial data, were introduced by Prentice [2]. His four criteria for the validation of a surrogate endpoint can be adapted for vaccine trials as follows:
Protection against the targeted disease is significantly related to having received the vaccine, where the corresponding logistic model (Prentice criterion 1) is given by:
The substitute endpoint is significantly related to the vaccination status (Prentice criterion 2):
where ε is the zeromean normally distributed error term.
The substitute endpoint is significantly related to protection against the clinical endpoint (Prentice criterion 3):
The full effect of the vaccine on the frequency of the clinical endpoint is explained by the substitute endpoint, as it lies on the sole causal pathway (Prentice criterion 4).
Therefore, criterion 4 is met if the null hypothesis H _{01}:γ_{Z}=0 is rejected and the null hypothesis H _{02}:β_{S}=0 is not rejected.
Although Prentice’s definition and criteria have been the subject of much debate [1, 4, 18], we decided to apply this approach for its simplicity and frequent usage, as well as its close relation to many of the methods proposed later on. These include the proportion of treatment explained [19], the proportion of information gain [20], and the individuallevel surrogacy measured by the information theoretic approach [21].
The metaanalytic framework
In this paper, we consider the metaanalytic framework in the single trial setting (STS), in which the units are randomized subgroups such as centers or regions. The metaanalytic approach can be represented by a bivariate mixedeffects model as follows:
where μ_{S} and μ_{T} are fixed intercepts, α and β the fixed effects of treatment on the endpoints, m_{Si} and m_{Ti} the random intercepts, and a_{i} and b_{i} the random effects of treatment on the endpoints in subgroup i [6]. For simplicity, we assume no random intercepts here (reduced model).
When the full bivariate mixedeffects approach is used to assess surrogacy, computational issues often occur. One simple solution is to use a fixed effect metaanalysis on aggregated data (twostage approach) [6]. This means performing separate regression of S on Z and then T on Z for each of the subgroups and then doing a weighted linear regression of the T slope (\(\hat \beta _{i}\)) on the S slope (\(\hat {\alpha _{i}}\))
with weights given by \(w_{i}=1/\hat Var(\hat \beta _{i})\). In this case, the trial level surrogacy is given by the R^{2} of the weighted linear regression. More sophisticated regression models can be used, such as the bivariate random effects model [22, 23].
Statistical solutions for high vaccine efficacy
Statistical methods for the analysis of rare events are extensively described in the literature [24]. VE can be expressed as follows:
where P(T=1Z=1) and P(T=1Z=0) are the probabilities of disease among vaccinated and unvaccinated individuals, respectively. In the context of high VE where a small number of events are observed in the vaccinated group, methods tailored for rare events can be applied in this specific setting. The following sections detail our proposal for statistical solutions that allow reliable CoP assessments of high efficacy vaccines. Both adapted methods are compatible with standard statistical software including R and SAS.
Flexible models for prentice criteria framework
The model assessing Prentice criterion 4 includes the surrogate and the treatment as covariates. When the number of events is small, this model can encounter issues due to lack of fit, leading to erroneous conclusions. To solve the problem of lack of fit, flexible link functions [25–27], could be used within Prentice framework. In this paper, we consider the classical logistic models with flexible (nonlinear) effect of the surrogate
where f(S_{j},θ) is a nonlinear function, such as polynomials or smoothing splines. This flexible model is popular for several reasons including: known properties, interpretability of parameters, easy to fit and implemented in many standard softwares.
The metaanalytic approach using penalised likelihood
The metaanalytic approach can be applied when multiple randomized subgroups are available for analysis. However, when applying this method in a high VE setting, maximum likelihood (ML) subgroupspecific VE estimates may be infinite, causing classical metaanalytic methods that combine subgroupspecific VE to potentially fail. To overcome this issue, we estimated subgroupspecific VE using the penalised likelihood method. Penalisation, which is equivalent to using proper priors on coefficients, solves the problem of infinite coefficient estimates. To achieve this we applied two approaches: the Firth method [28], and the weakly informative prior (WIP) proposed by Gelman et al. [29]. Firth showed that his method is equivalent to the use of Jeffreys’ invariant prior. Gelman et al. on the other hand proposed a WIP distribution (Cauchy prior with scale 2.5), which relies on the assumption that a typical change in an input variable is unlikely to correspond to a change as high as 5 on the logistic scale. As part of a twostep approach, we first independently executed the Firth method and Gelman approach using the logistf and bayesglm R packages respectively [30, 31]. In a second step, we evaluated the performance of both methods as part of a metaanalysis in the context of high VE, by running simulations.
Results
Flexible models for the prentice criteria framework
To evaluate the impact of the lack of fit corresponding to Prentice criterion 4, we simulated data using the Dunning regression model [26] in an ideal CoP setting, where the treatment effect is fully explained by the surrogate (full mediation) as follows:
Here, π is interpreted as the probability of being exposed to the disease. Irrespective of the interpretation of π, this is a valuable, monotone, skewed, flexible and nonlinear model to generate the type of data described above.
Simulations were run using the following parameter assumptions: Total sample size n=5000, 1:1 randomization, π=0.1, p_{0}=P(T=1Z=0)=0.05, μ_{1}=E(SZ=1)=4.5,4,3.75,3.33, μ_{0}=E(SZ=0)=3, VAR(SZ=1)=VAR(SZ=0)=0.2, γ=log(1−0.95), μ=8.3. A range of VE values were considered (VE = 0.4, 0.75, 0.85 and 0.95), and 5000 datasets were simulated for each scenario. We fitted Prentice model 4 on the simulated data using classical logit regression shown in Eq. (1), the proposed nonlinear model depicted in Eq. (3) with a quadratic term
and the scaled logistic model [26]. Table 1 shows the outcome of these simulations.
Table 1 shows that using a flexible model considerably increases the power to meet Prentice criterion 4 when the VE increases. In fact, the simple linear logistic model does not control the typeI error of the treatment effect (p(Z)<α) when VE is high. This is due to the lack of fit of the linear effect which is absorbed by the treatment effect, thereby considerably reducing the power to meet Prentice criterion 4. We can see that the scaled logistic model is slightly conservative. Standard errors of this model should be computed by bootstrap [27].
The metaanalytic approach using penalised likelihood
We considered the metaanalytic approach in a single trial setting. The single trial was split into several relatively small randomized subgroups (such as geographical regions or centers), and these small subgroups were used as units for the metaanalysis. For illustration purposes, we analysed a publicly available simulated dataset containing both continuous outcome and surrogate endpoints [21]. This dataset consists of 50 subgroups characterised by a 1:1 randomization and sample size of 20 per subgroup.
Figure 1a shows the results of the twostage metaanalytic approach with a continuous outcome. Here, a strong correlation between the treatment effect on the true outcome (\(\hat \beta _{i}\)) and the treatment effect on the surrogate outcome (\(\hat \alpha _{i}\)) is observed, with an estimated R^{2} of 0.77. When artificially dichotomising the true outcome as Y=1 if T<−2.87 and Y=0 if T≥2.87, the resulting VE on this binary outcome is 95%. Figure 1b shows the results on this true binary outcome, where several β values fall around 10. These values are extremely high for a logistic regression and they are due to the lack of events in the treatment group, thus generating a small R^{2} value (0.17). Figure 1c shows the twostage metaanalytic approach, where the treatment effect on the binary outcome is estimated using the penalised likelihood approach proposed by Firth [28]. Here, we observe that the problem of infinite estimates is solved, and so the R^{2} value is much higher compared to the classical approach. Similar results were obtained using the penalised likelihood approach proposed by Gelman, as shown in Fig. 1d [31]. To better understand the results it is useful to look at summary statistics from the different logistic models by number of events in control and in vaccinated groups. Table 2 shows that when there are no events in the two groups (n_{V}=n_{C}=0) then the estimated effect is zero (\(\hat \beta =0\)) and the estimated variance is “infinite” for the logistic model while it is relatively small for the penalized methods. When there are no events only in the vaccinated group (n_{V}=0 and n_{C}>0) then the effect and the variance estimated by the standard logistic model are “infinite”, while the penalization of the likelihood prevents infinite estimates and variances. This is the reason why the penalized methods outperform the standard logistic approach in the case of high VE.
To confirm these results, additional data was simulated with a true binary outcome and a continuous surrogate, using the reduced model in Eq. (2) without random intercepts. This dataset consists of 25 subgroups and n =40 participants per subgroup with a 1:1 randomisation. We simulated data using the following parameters: μ_{S}=4.609; μ_{T}=−2.2401; α=5.458; β=(−1,−2,−4); Var(a_{i}) =10; Var(b_{i}) =4. The correlation between the treatment random effects is \(\rho = {Cor}(a_{i}, b_{i})=\sqrt {0.9}\), with an R^{2} value of 0.9. The R^{2} estimated by different methods as a function of VE is presented in Table 3.
Table 3 shows that penalised approaches (Firth and Gelman’s WIP) outperform the standard logistic model in terms of Mean Square Error (MSE), especially in case of high VE where there is a high chance of having subgroups with zero events in the vaccination group. In fact, when the VE is 0.75, 0.82 and 0.95, the average number of subgroups with zero events in vaccination groups are 9, 13 and 20, respectively. Both penalised approaches show very similar results.
Discussion
Despite recent advances in immunology, we are only beginning to understand how vaccines work best, and how we can improve vaccine design for higher protective efficacy [32]. Although not common, vaccines with a high efficacy, are documented in the literature [12–17, 33]. These include the salmonella typhi vi conjugate [12], or the combined measlesmumpsrubellavaricella immunisation [17]. Rare events data obtained in high VE trials make it challenging for statisticians to apply classical methods used for CoP assessment due to the lack of available information. These include ML estimators, where bias, infinite estimates, multicollinearity and convergence issues can arise and negatively impact Prentice criteria and metaanalytic frameworks commonly used to assess vaccine CoPs, as shown in this paper [24, 26, 27].
To overcome this problem, we evaluated the impact of high VE using two classical statistical approaches: the Prentice framework and the Metaanalytic framework applied on randomized subgroups (e.g. geographical regions). We chose these methods for their common usage in CoP assessments, and their userfriendly characteristics. We performed data simulations with high VE to illustrate the problems and to evaluate the proposed solutions.
By working on the Prentice framework, we show that it is critical to both design and evaluate flexible and adaptable models that are tailored to high VE cases, as the lack of fit of a model leads to substantial loss in power. Accordingly, we propose to analyse data using a logistic model with nonlinear surrogate effect. This popular model is flexible, with known properties, easy to fit and implemented in many standard softwares. The number of additional parameters should be small to avoid overfitting. Other models with flexible link functions have also been proposed that can be used within the Prentice framework [26, 27]. Model selection can be done using the Akaike Information Criterion (AIC) approach. Furthermore, adjustments for baseline covariates can play an important role in improving model fit.
Regarding the metaanalytic framework, we demonstrate that penalised likelihood approaches (such as Firth or Gelman’s WIP) outperform the standard logistic model when VE is high, as they solve the problem of infinite estimates. This problem can occur when VE is high where there is a high probability of observing zero cases in certain subgroups of the vaccinated group, as we have also shown. For simplicity, we used a twostage approach where treatment effects were estimated for each subgroup using a penalised likelihood approach, followed by a (fixed effect) metaanalysis to combine results from different subgroups. Another possibility is to use a mixed model with WIP or Jeffrey priors. For example, it is straightforward to implement the bivariate model, depicted in Eq. (2), with WIP for the covariance matrix of the treatment randomeffects using a Bayesian framework (e.g. WinBugs, JAGS or Stan). Additional simulation studies, comparing one and twostage penalised approaches, would therefore be worth pursuing to help overcome these problematics in the context of high VE.
It is noteworthy that the concept of a vaccine CoP often refers to the establishment of a protective immunogenicity threshold as alluded to earlier, above which disease acquisition is unlikely to happen. However, relating immunological biomarkers to disease risk and therefore VE can also be made possible as part of a continuous approach, without the assumption of a threshold titre. This manuscript addressed this type of (continuous) approach that employs fitted regression models on antibody titres in vaccinated and nonvaccinated individuals to show the statistical association between antibody titres and disease incidence [1, 26, 34, 35].
Although this study was limited by its use of simulated data only, our results suggest that the solutions we propose substantially increase the power of classical statistical approaches for CoP assessment, when dealing with high VE. Furthermore, they are straightforward and compatible with standard statistical software.
Conclusions
Following our observation that CoP assessments for high VE vaccines comes with statistical issues using standard methods, we devised flexible nonlinear models to counteract the lack of fit in the Prentice framework, and propose penalized likelihood approaches for metaanalysis. These statistical solutions are easytoimplement adaptations to both conventional methods for application in high VE cases. Such statistical challenges associated with high VE may have so far been overlooked due to their low occurrence, yet high VE cases exist. For binary surrogates it may be interesting to explore how the individual causal association [9] and the surrogate predictive function [36] perform in the setting of high VE. Finally, evaluating the impact of high VE on the Principal stratification approach should be beneficial to the field, towards improving CoP assessments of vaccines [3–5, 10, 11].
Abbreviations
 AIC:

Akaike information criterion
 CDF:

Cumulative distribution function
 CI:

Confidence interval
 CoP:

Correlate of protection
 CoR:

Correlate of risk
 LL:

Lower limit
 ML:

Maximum likelihood
 MSE:

Mean squared error
 SE:

Standard error
 SoP:

Surrogate of protection
 UL:

Upper limit
 VE:

Vaccine efficacy
 WIP:

Weakly informative prior
References
 1
NguipdopDjomo P, Thomas SL, Fine PEM. Correlates of vaccineinduced protection: methods and implications. WHO/IVB/10.00. 2013; 181:1–55.
 2
Prentice RL. Surrogate endpoints in clinical trials: definition and operational criteria. Stat Med. 1989; 8(4):431–40.
 3
Follmann D. Augmented designs to assess immune response in vaccine trials. Biometrics. 2006; 62(4):1161–9.
 4
Frangakis CE, Rubin DB. Principal stratification in causal inference. Biometrics. 2002; 58(1):21–9.
 5
Gilbert PB, Qin L, Self SG. Evaluating a surrogate endpoint at three levels, with application to vaccine development. Stat Med. 2008; 27(23):4758–78.
 6
Buyse M, Molenberghs G, Burzykowski T, Renard D, Geys H. The validation of surrogate endpoints in metaanalysis of randomized experiments. Biostatistics. 2000; 1:49–67.
 7
Daniels MJ, Hughes MD. Metaanalysis for the evaluation of potential surrogate markers. Stat Med. 1997; 16:1965–82.
 8
Gail MH, Pfeiffer R, Houwelingen HCV, Carroll R. On metaanalytic assessment of surrogate outcomes. Biostatistics. 2000; 1:231–46.
 9
Alonso A, Van der Elst W, Molenberghs G, Buyse M, Burzykowski T. On the relationship between the causalinference and metaanalytic paradigms for the validation of surrogate endpoints. Biometrics. 2015; 71(1):15–24.
 10
Qin L, Gilbert PB, Corey L, McElrath MJ, Self SG. A framework for assessing immunological correlates of protection in vaccine trials. J Infect Dis. 2007; 196(9):1304–12.
 11
Rubin DB. Causal inference using potential outcomes: Design, modeling, decisions. J Am Stat Assoc. 2005; 100(469):322–31.
 12
Mitra M, Shah N, Ghosh A, Chatterjee S, Kaur I, Bhattacharya N, Basu S. Efficacy and safety of vitetanus toxoid conjugated typhoid vaccine (pedatyph) in indian children: school based cluster randomized study. Hum Vaccines Immunotherapeutics. 2016; 12(4):939–45.
 13
Lin FYC, Ho VA, Khiem HB, et al.The efficacy of a salmonella typhi vi conjugate vaccine in twotofiveyearold children. N Engl J Med. 2001; 344(17):1263–9.
 14
Wei M, Meng F, Wang S, Li J, et al.Twoyear efficacy, immunogenicity, and safety of vigoo enterovirus 71 vaccine in healthy chinese children: a randomised openlabel study. J Infect Dis. 2017; jiw502:56–63.
 15
Phua KB, Lim FS, Lau YL, Nelson EAS, et al.Rotavirus vaccine RIX4414 efficacy sustained during the third year of life: a randomized clinical trial in an asian population. Vaccine. 2012; 30(30):4552–7.
 16
Black S, Shinefield H, Fireman B, Lewis E, et al.Efficacy, safety and immunogenicity of heptavalent pneumococcal conjugate vaccine in children. Pediatr Infect Dis J. 2000; 19(3):187–95.
 17
Prymula R, Bergsaker MR, Esposito S, Gothefors L, et al.Protection against varicella with two doses of combined measlesmumpsrubellavaricella vaccine versus one dose of monovalent varicella vaccine: a multicentre, observerblind, randomised, controlled trial. Lancet. 2014; 383(9925):1313–24.
 18
Burzykowski T, Molenberghs G, Buyse M. The Evaluation of Surrogate Endpoints. New York: Springer; 2005.
 19
Freedman LS, Graubard BI, Schatzkin A. Statistical validation of intermediate endpoints for chronic disease. Stat Med. 1992; 11:167–78.
 20
Qu Y, Case M. Quantifying the effect of the surrogate marker by information gain. Biometrics. 2007; 63(3):958–63.
 21
Alonso A, Molenberghs G. Surrogate marker evaluation from an information theory perspective. Biometrics. 2007; 63:180–6.
 22
Houwelingen H. C. v., Arends LR, Stijnen T. Advanced methods in metaanalysis: Multivariate approach and metaregression. Stat Med. 2002; 21(4):589–624.
 23
Tibaldi F, Abrahantes JC, et al. Simplified hierarchical linear models for the evaluation of surrogate endpoints. J Stat Comput Simul. 2003; 73:643–58.
 24
Del Paal B. A comparison of different methods for modelling rare events data. PhD thesis, Ghent University, Ghent, Belgium. 2013.
 25
Kim HJ. Binary regression with a class of skewed t link models. Commun Stat. 2002; 31(10):1863–6.
 26
Dunning AJ. A model for immunological correlates of protection. Stat Med. 2006; 25(9):1485–97.
 27
Dunning AJ, Kensler J, Coudeville L, Bailleux F. Some extensions in continuous models for immunological correlates of protection. BMC Med Res Methodol. 2015; 15(1):107.
 28
Firth D. Bias reduction of maximum likelihood estimates. Biometrika. 1993; 80:27–38.
 29
Gelman A, Jakulin A, Pittau MG, Su YS. A weakly informative default prior distribution for logistic and other regression models. Ann Appl Stat. 1993; 2(4):1360–83.
 30
Heinze G, Ploner M, Dunkler D, Southworth H. logisf: Firth’s bias reduced logistic regression. R package version 1.21. 2013; 1.
 31
Gelman A, Su YS. arm: Data analysis using regression and multilevel/hierarchical models. R package version 1.86. 2015;1.
 32
Slifka MK, Amanna I. How advances in immunology provide insight into improving vaccine efficacy. Vaccine. 2014; 32(25):2948–57.
 33
Naud PS, RoteliMartins CM, De Carvalho NS, Teixeira JC, de Borba PC. Sustained efficacy, immunogenicity, and safety of the HPV16/18 AS04adjuvanted vaccine: final analysis of a longterm followup study up to 9.4 years postvaccination. Hum Vaccin Immunother. 2014; 10(8):2147–62.
 34
Siber GR. Methods for estimating serological correlates of protection. Dev Biol Stand. 1997; 89:283–96.
 35
Chan IS, Li S, Matthews H, Chan C, Vessey R, Sadoff J, et al.Use of statistical models for evaluating antibody response as a correlate of protection against varicella. Stat Med. 2002; 21(22):3411–1430.
 36
Alonso A, Van der Elst W, Meyvisch P. Assessing a surrogate predictive value: a causal inference approach. Stat Med. 2017; 36(7):1083–98.
Acknowledgements
The authors would like to thank Prof J.C. (Hans) van Houwelingen (LUMC, Leiden University) for his valuable advice and guidance, and Martine Douha for her contribution to the data analysis. Medical writing services, and editorial assistance and publication coordination, were provided by Sonia Norris and Sophie Timmery (XPE Pharma & Science on behalf of GSK) respectively.
Funding
GlaxoSmithKline Biologicals SA was the funding source and was involved in all stages of the study conduct and analysis. GlaxoSmithKline Biologicals SA also took responsibility for all costs associated with the development and publishing of the present manuscript.
Availability of data and materials
Not applicable.
Author information
Affiliations
Contributions
AC and FT equally contributed to all steps of the manuscript’s development, and approved its final version. All authors read and approved the final manuscript.
Corresponding author
Ethics declarations
Ethics approval and consent to participate
Not applicable.
Consent for publication
Not applicable.
Competing interests
AC and FT are employees of the GSK group of companies and hold shares in the GSK group of companies.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License(http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver(http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
About this article
Cite this article
Callegaro, A., Tibaldi, F. Assessing correlates of protection in vaccine trials: statistical solutions in the context of high vaccine efficacy. BMC Med Res Methodol 19, 47 (2019). https://doi.org/10.1186/s128740190687y
Received:
Accepted:
Published:
Keywords
 Vaccine clinical trial
 High vaccine efficacy
 Surrogate endpoint
 Correlate of protection