Efficacy and Duration of Pneumococcal Conjugate Vaccines
Efficacy and Duration of Pneumococcal Conjugate Vaccines
We identified intervention studies reported by Fleming-Dutra et al in a recent systematic review of PCV vaccination schedules, which was based on data published between 1994 and September 2010, with post-hoc supplementation of studies published from 2011. We searched for any additional study published between 2011 and May 31, 2014 using a similar strategy as reported by Fleming-Dutra et al, using EMBASE and MEDLINE databases. Details are provided in Appendix, Supplemental Digital Content 1, http://links.lww.com/INF/C158.
We considered the following initial criteria for inclusion: (i) intervention studies, (ii) providing NP carriage estimates in vaccinated and unvaccinated children and (iii) with children vaccinated as per routine schedule, including 3 primary doses ("3 + 0" schedule) or at least 2 primary doses with a booster dose ("2 + 1" and "3 + 1" schedules). We further restricted our analysis to studies of either 7-valent, 10-valent or 13-valent licensed vaccines (PCV7, PCV10 and PCV13) or unlicensed vaccines (eg, PCV9 and PCV11) linked to similar carrier proteins as licensed vaccines, including the Corynebacterium diphtheria toxin mutant 197, meningococcal outer membrane protein complex or the nontypeable Haemophilus Influenzae derived protein D. Studies based on vaccines conjugated to other proteins or for which immunological equivalence is unclear (such as tetravalent and pentavalent vaccines) were not included.
Given that PCVs are not known to affect carriage clearance that the average duration of VT carriage in infants and young children is somewhere around 2 months but may vary by setting and serotype, and that 2–4 weeks are required for the antibody response to peak after vaccination, we excluded any data collected earlier than 4 months after complete vaccination, when the prevalence and serotype distribution was considered nonstationary, as detailed elsewhere.
All but 4 studies were PCV7 trials, with 3 other trials based on PCV9 and 1 on PCV10. We extracted data on the group of PCV7 serotypes, and each individual PCV7 serotype (4, 6B, 9V, 14, 18C, 19F, 23F). We also extracted data on serotype 6A, one of the most common serotypes, which shares immunological traits with 6B but is not included in PCV7, PCV9 or PCV10, to explore possible cross-reactive protective efficacy. Other potential cross-reactive serotypes, such as 19A, were not studied, because of limited data.
We defined the vaccine efficacy against carriage acquisition (VEC) as the relative reduction in the rate of carriage acquisition among vaccinated compared with unvaccinated children, in trial conditions. Although acquisition events cannot directly be observed, it is possible to obtain a robust estimate of VEC from cross-sectional data based on 1 - odds ratio (OR), under general assumptions, with the OR defined as the odds of vaccination among the (group of) VT(s) (henceforth, the "target" group) to the odds of vaccination among those not carrying any VT (henceforth, the "reference" group). Hence, in calculating the VEC for each individual PCV7 serotype, we included in the target group all vaccinated and unvaccinated carriers of the particular serotype and in the reference group all non-vaccine serotype (NVT) carriers and noncarriers. Other VT were excluded from the serotype-specific analysis to account for vaccine-induced within-host changes in the pneumococcal flora, as explained elsewhere. We also excluded all VT from the analysis of VEC against 6A. Similarly, in trials based on vaccines with higher valency than PCV7, data on the additional VT were excluded. Further details about the methods and assumptions underpinning the estimation of VEC from cross-sectional data are described elsewhere. The analysis was based on summary data by (group of) serotype(s), rather than individual-level data.
We explored whether the proportion of carried VT out of all VT differed between studies, based on data in unvaccinated children, and used I values to quantify heterogeneity.
We used a Bayesian logistic meta-regression model to estimate the aggregate and serotype-specific VEC and its waning. In the model, for each study i
where
and
are the proportion of vaccinated individuals in the reference and target groups, respectively, θ i is the study-specific natural logarithm of the OR, and β 1 represents the coefficient by which the log(OR) changes for each increase in the natural logarithm of time t since the peak VE C (ie, 4 months after vaccination), such that log(OR t ) = θ i + β 1 *log( t i), with time in months.
We used a random effect model taking the between-study heterogeneity into account by assuming that θi were independent and sampled from a normal distribution centered around the mean log(OR) of carriage (μ) with a precision τ, such that θi ~ N(μ,τ) and τ = 1/σ, where σ is the between-study variance. A fixed effect was assumed for β1.
The VEC at time t can, therefore, be expressed as follows:
We assigned uniform priors to α [unif (−10; 10)], μ [unif (−10, 0)], σ [unif (0,10)] and β1 [unif (0,10)]. The time coefficient β1 was constrained to positive values, with the assumption that the efficacy should be declining. This assumption was further tested in a sensitivity analysis, by placing an unconstrained prior on β1 [unif (−10,10)].
Some studies provided more than 1 estimate. However, we did not adjust for the lack of independence because of the limited number of estimates from each study.
We explored the impact of schedule [booster (3 + 1 or 2 + 1) versus nonbooster (3 + 0)] by including schedule as a covariate in a multivariable model and assigned a normal uninformed prior to its coefficient [ ~ N(0,10)]. We used an interaction term between schedule and time to look for a difference in the waning by schedule, with a normal uninformed prior on the interaction coefficient [β3 ~ N(0,10)]. Studies in which a 23-valent polysaccharide vaccine (PPV23) booster dose was provided after a primary schedule (as given by Russell et al and Lakshman et al) were considered part of the 3 + 0 group, given the lack of effect of PPV23 on carriage.
Finally, we conducted sensitivity analyses to explore the impact on our pooled VEC estimates of omitting any 1 study. We also analyzed 2 additional models of waning VEC, including a model where time was included as a linear covariate and another model with an asymptotic function in which the VEC of carriage approaches 0 as time approaches infinity. Models were compared using the Deviance Information Criterion (DIC), a likelihood-based model fitting statistic for Bayesian models similar to the frequentist Akaike Information Criterion. Further details are presented in Appendix, Supplemental Digital Content 2, http://links.lww.com/INF/C159.
Posterior distributions were obtained through a Markov Chain Monte Carlo Gibbs sampling algorithm based on 2 chains of 100,000 iterations running in parallel, after a burn-in of 5000 iterations. The model was implemented in R using the jags package.
Methods
Search Strategy
We identified intervention studies reported by Fleming-Dutra et al in a recent systematic review of PCV vaccination schedules, which was based on data published between 1994 and September 2010, with post-hoc supplementation of studies published from 2011. We searched for any additional study published between 2011 and May 31, 2014 using a similar strategy as reported by Fleming-Dutra et al, using EMBASE and MEDLINE databases. Details are provided in Appendix, Supplemental Digital Content 1, http://links.lww.com/INF/C158.
Inclusion Criteria
We considered the following initial criteria for inclusion: (i) intervention studies, (ii) providing NP carriage estimates in vaccinated and unvaccinated children and (iii) with children vaccinated as per routine schedule, including 3 primary doses ("3 + 0" schedule) or at least 2 primary doses with a booster dose ("2 + 1" and "3 + 1" schedules). We further restricted our analysis to studies of either 7-valent, 10-valent or 13-valent licensed vaccines (PCV7, PCV10 and PCV13) or unlicensed vaccines (eg, PCV9 and PCV11) linked to similar carrier proteins as licensed vaccines, including the Corynebacterium diphtheria toxin mutant 197, meningococcal outer membrane protein complex or the nontypeable Haemophilus Influenzae derived protein D. Studies based on vaccines conjugated to other proteins or for which immunological equivalence is unclear (such as tetravalent and pentavalent vaccines) were not included.
Given that PCVs are not known to affect carriage clearance that the average duration of VT carriage in infants and young children is somewhere around 2 months but may vary by setting and serotype, and that 2–4 weeks are required for the antibody response to peak after vaccination, we excluded any data collected earlier than 4 months after complete vaccination, when the prevalence and serotype distribution was considered nonstationary, as detailed elsewhere.
Data Extraction
All but 4 studies were PCV7 trials, with 3 other trials based on PCV9 and 1 on PCV10. We extracted data on the group of PCV7 serotypes, and each individual PCV7 serotype (4, 6B, 9V, 14, 18C, 19F, 23F). We also extracted data on serotype 6A, one of the most common serotypes, which shares immunological traits with 6B but is not included in PCV7, PCV9 or PCV10, to explore possible cross-reactive protective efficacy. Other potential cross-reactive serotypes, such as 19A, were not studied, because of limited data.
Analysis
We defined the vaccine efficacy against carriage acquisition (VEC) as the relative reduction in the rate of carriage acquisition among vaccinated compared with unvaccinated children, in trial conditions. Although acquisition events cannot directly be observed, it is possible to obtain a robust estimate of VEC from cross-sectional data based on 1 - odds ratio (OR), under general assumptions, with the OR defined as the odds of vaccination among the (group of) VT(s) (henceforth, the "target" group) to the odds of vaccination among those not carrying any VT (henceforth, the "reference" group). Hence, in calculating the VEC for each individual PCV7 serotype, we included in the target group all vaccinated and unvaccinated carriers of the particular serotype and in the reference group all non-vaccine serotype (NVT) carriers and noncarriers. Other VT were excluded from the serotype-specific analysis to account for vaccine-induced within-host changes in the pneumococcal flora, as explained elsewhere. We also excluded all VT from the analysis of VEC against 6A. Similarly, in trials based on vaccines with higher valency than PCV7, data on the additional VT were excluded. Further details about the methods and assumptions underpinning the estimation of VEC from cross-sectional data are described elsewhere. The analysis was based on summary data by (group of) serotype(s), rather than individual-level data.
We explored whether the proportion of carried VT out of all VT differed between studies, based on data in unvaccinated children, and used I values to quantify heterogeneity.
We used a Bayesian logistic meta-regression model to estimate the aggregate and serotype-specific VEC and its waning. In the model, for each study i
where
and
are the proportion of vaccinated individuals in the reference and target groups, respectively, θ i is the study-specific natural logarithm of the OR, and β 1 represents the coefficient by which the log(OR) changes for each increase in the natural logarithm of time t since the peak VE C (ie, 4 months after vaccination), such that log(OR t ) = θ i + β 1 *log( t i), with time in months.
We used a random effect model taking the between-study heterogeneity into account by assuming that θi were independent and sampled from a normal distribution centered around the mean log(OR) of carriage (μ) with a precision τ, such that θi ~ N(μ,τ) and τ = 1/σ, where σ is the between-study variance. A fixed effect was assumed for β1.
The VEC at time t can, therefore, be expressed as follows:
We assigned uniform priors to α [unif (−10; 10)], μ [unif (−10, 0)], σ [unif (0,10)] and β1 [unif (0,10)]. The time coefficient β1 was constrained to positive values, with the assumption that the efficacy should be declining. This assumption was further tested in a sensitivity analysis, by placing an unconstrained prior on β1 [unif (−10,10)].
Some studies provided more than 1 estimate. However, we did not adjust for the lack of independence because of the limited number of estimates from each study.
We explored the impact of schedule [booster (3 + 1 or 2 + 1) versus nonbooster (3 + 0)] by including schedule as a covariate in a multivariable model and assigned a normal uninformed prior to its coefficient [ ~ N(0,10)]. We used an interaction term between schedule and time to look for a difference in the waning by schedule, with a normal uninformed prior on the interaction coefficient [β3 ~ N(0,10)]. Studies in which a 23-valent polysaccharide vaccine (PPV23) booster dose was provided after a primary schedule (as given by Russell et al and Lakshman et al) were considered part of the 3 + 0 group, given the lack of effect of PPV23 on carriage.
Finally, we conducted sensitivity analyses to explore the impact on our pooled VEC estimates of omitting any 1 study. We also analyzed 2 additional models of waning VEC, including a model where time was included as a linear covariate and another model with an asymptotic function in which the VEC of carriage approaches 0 as time approaches infinity. Models were compared using the Deviance Information Criterion (DIC), a likelihood-based model fitting statistic for Bayesian models similar to the frequentist Akaike Information Criterion. Further details are presented in Appendix, Supplemental Digital Content 2, http://links.lww.com/INF/C159.
Posterior distributions were obtained through a Markov Chain Monte Carlo Gibbs sampling algorithm based on 2 chains of 100,000 iterations running in parallel, after a burn-in of 5000 iterations. The model was implemented in R using the jags package.