Efficacy of RTS,S malaria vaccines: individual-participant pooled analysis of phase 2 data

Summary Background The efficacy of RTS,S/AS01 as a vaccine for malaria is being tested in a phase 3 clinical trial. Early results show significant, albeit partial, protection against clinical malaria and severe malaria. To ascertain variations in vaccine efficacy according to covariates such as transmission intensity, choice of adjuvant, age at vaccination, and bednet use, we did an individual-participant pooled analysis of phase 2 clinical data. Methods We analysed data from 11 different sites in Africa, including 4453 participants. We measured heterogeneity in vaccine efficacy by estimating the interactions between covariates and vaccination in pooled multivariable Cox regression and Poisson regression analyses. Endpoints for measurement of vaccine efficacy were infection, clinical malaria, severe malaria, and death. We defined transmission intensity levels according to the estimated local parasite prevalence in children aged 2–10 years (PrP2–10), ranging from 5% to 80%. Choice of adjuvant was either AS01 or AS02. Findings Vaccine efficacy against all episodes of clinical malaria varied by transmission intensity (p=0·001). At low transmission (PrP2–10 10%) vaccine efficacy was 60% (95% CI 54 to 67), at moderate transmission (PrP2–10 20%) it was 41% (21 to 57), and at high transmission (PrP2–10 70%) the efficacy was 4% (−10 to 22). Vaccine efficacy also varied by adjuvant choice (p<0·0001)—eg, at low transmission (PrP2–10 10%), efficacy varied from 60% (95% CI 54 to 67) for AS01 to 47% (14 to 75) for AS02. Variations in efficacy by age at vaccination were of borderline significance (p=0·038), and bednet use and sex were not significant covariates. Vaccine efficacy (pooled across adjuvant choice and transmission intensity) varied significantly (p<0·0001) according to time since vaccination, from 36% efficacy (95% CI 24 to 45) at time of vaccination to 0% (−38 to 38) after 3 years. Interpretation Vaccine efficacy against clinical disease was of limited duration and was not detectable 3 years after vaccination. Furthermore, efficacy fell with increasing transmission intensity. Outcomes after vaccination cannot be gauged accurately on the basis of one pooled efficacy figure. However, predictions of public-health outcomes of vaccination will need to take account of variations in efficacy by transmission intensity and by time since vaccination. Funding Medical Research Council (UK); Bill & Melinda Gates Foundation Vaccine Modelling Initiative; Wellcome Trust.


Introduction
The increasing application of interventions for malaria control over the past 10 years has been linked to reductions in morbidity and mortality associated with malaria infection. 1,2 A vaccine for malaria could have an important role in further reduction of the burden of disease. The candidate malaria vaccine RTS,S/AS01 is now in phase 3 clinical trials, for which preliminary data for the fi rst 12 months of follow-up are available. 3 Effi cacy against clinical malaria was 55·8% (97·5% CI 50·6-60·4) among children age 5-17 months. Combined effi cacy against severe malaria for children aged 5-17 months and 6-12 weeks was 34·8% (95% CI 16·2-49·2).
RTS,S protects at pre-erythrocytic stages of the parasite's lifecycle. It is partly eff ective and has been described as a leaky vaccine 4 -ie, no individual is protected consistently against every episode of exposure, but the risk of acquiring infection after any single episode of exposure is reduced. In fi eld trials, RTS,S has been given with either of two diff erent adjuvant systems: AS01 or AS02. Although RTS,S/AS01 seems to be more immunogenic than RTS,S/AS02, effi cacy trials of RTS,S/ AS01 and RTS,S/AS02 have not resulted in defi nitively powered comparisons. 5,6 Furthermore, the variation in vaccine effi cacy over time remains unknown, with confl icting evidence from individual trials.
Vaccine effi cacies are usually summarised with point estimates. However, if vaccine effi cacy is heterogeneous by subgroups within the population, this effi cacy fi gure will be a mean of the vaccine effi cacy in the various subgroups, weighted according to the proportion of the population. 7 For instance, if vaccine effi cacy is higher in older children then the overall effi cacy in a particular trial will depend on the proportion of older children to younger children that are vaccinated.
Analysis of phase 2b data to date shows variations in measured effi cacy between trials. 6,[8][9][10][11][12][13][14][15][16] These diff erences might be attributable to the vaccine formulation, intensity of transmission, length of follow-up, or agerange of participants. To ascertain which covariates are associated with variations in vaccine effi cacy, we did a pooled analysis of data from phase 2b trials.

Data collection
We identifi ed phase 2b trials 6,[8][9][10][11][12][13][14][15][16] of RTS,S from the GlaxoSmithKline Biologicals registry of trials (data on fi le), and raw data were provided by GSK Biologicals to three academic investigators (PB, MTW, and ACG). One of us (PB) checked data for completeness by comparing data summaries with the primary publications; all investigators analysed the data. Characteristics of the trials, done at 11 sites in total (from six countries), are summarised in table 1. In the identifi ed trials, healthy adults or children were recruited after clinical and laboratory screening to exclude participants with clinically signifi cant disease. Five trials of children or adults used active case detection for Plasmodium falciparum infection (ACDi), 6,[8][9][10][11] one used active case detection for clinical malaria (ACDc), 12 and two used passive case detection (PCD) for clinical malaria 10,13 (one trial used both ACDi and PCD). Trials using ACDi and ACDc included assessment of participants who presented with acute illness between scheduled visits, which is usually referred to as PCD in protocols. For simplicity, in our analysis here we use ACDi to refer to the combination of ACDi and PCD, ACDc to refer to combined ACDc and PCD, and PCD to refer to exclusive use of PCD. ACDi was done after antimalarial treatment during the vaccination course and was then monitored with regular fi nger-prick blood smears. Deaths and severe malaria episodes were monitored in all fi ve trials in which children were enrolled.

Procedures
We used four primary endpoints in our analysis: infection, clinical malaria, severe malaria, and death. In the trials we identifi ed, infection was defi ned as any detectable P falciparum parasitaemia, with or without a measured fever. We defi ned clinical malaria as the presence of 2500 or more P falciparum parasites per μL of blood in association with reported or measured fever (≥37·5°C). 17,18 We deemed clinical malaria episodes occurring within 28 days of a previous episode to be part of the same episode. We did not censor time of monitoring according to antimalarial drug use or reported absences from the study area.
We analysed episodes of infection identifi ed by ACDi as one dataset. We combined clinical malaria episodes identifi ed by ACDc and PCD and analysed these as a second dataset. ACDc and PCD were included in the initial study protocols, except for one trial, 13 in which the effi cacy assessment was included after a protocol amendment as an exploratory objective, for which some data were extracted retrospectively.
We identifi ed episodes of severe malaria from safety data reporting. Criteria for severe malaria were derived from the WHO defi nition 19 and were applied by the clinical investigators at every site, comprising asexual P falciparum parasitaemia, no alternative (or more probable) cause of illness, and either severe malaria  Our primary analysis was vaccine effi cacy, which was assessed per-protocol. Hence, cohorts monitored for infection or clinical malaria included all participants who had received three doses of vaccine, from 2 weeks after the third dose. We did not judge adults at risk of severe disease or death from malaria. Analysis of severe malaria or death was based on intention to treat and included all children who received at least one dose of vaccine, from the time of the fi rst vaccination.
In Mozambique, 10,14,16 one cohort (cohort 2) fi rst underwent ACDi in a double-blind phase then subsequently underwent PCD for clinical malaria in a singleblind phase. We included data from the double-blind phase in the ACDi dataset and those from the singleblind phase in the clinical malaria dataset, taking the start of the single-blind phase as the initial time of monitoring for clinical malaria.
We recorded the following covariates across the seven trials: sex, age at the time of vaccination, country, bednet use, adjuvant used (ie, AS01 vs AS02), and clinical disease surveillance method at the site level (ie, ACDc vs PCD only). To ascertain transmission intensity, we used estimates from the Malaria Atlas Project (MAP) for prevalence of asymptomatic parasitaemia in children aged 2-10 years in 2007 (PrP 2-10 ), 20 identifi ed with the geopositioning coordinates of the trial sites. We refer to this measure here as the local parasite prevalence.

Statistical analysis
We summarised unadjusted vaccine effi cacy for the four endpoints of infection, clinical malaria, severe malaria, and death with Kaplan-Meier curves and effi cacy estimates with unadjusted Cox proportional-hazard models for fi rst or only event. To analyse ACDi and combined ACDc and PCD, we assessed the eff ect of covariates with adjusted Cox proportional-hazard models. We analysed multiple episodes of clinical malaria with Poisson regression, adjusting for the time of follow-up as an off set variable, implemented as one observation per participant.
Rather than present subgroup analyses according to strata (which would be necessarily narrow and be typically confounded by other important covariates), we pooled individual participant data and estimated the linear and non-linear eff ect of covariates in the data and the interactions of these covariates with vaccine effi cacy. We used these empirically observed functions to estimate effi cacy in subgroups by multiplying the fi xed eff ect of vaccination (ie, the estimate of the eff ect of vaccination among those with the baseline value of the covariate) by the interaction term (ie, the estimate of how vaccine effi cacy varies for each diff erent level of the covariate). We added variances and the covariance matrix to calculate SEs. All covariates were included in an initial model, and we excluded covariates or interaction terms with p greater than 0·05 to produce a fi nal model. To examine the possibility that analyses of clinical malaria risk were biased by unequal durations of follow-up in some subgroups, we did an additional analysis restricted to 1 year of follow-up.
We calculated vaccine effi cacy as either 1 minus the hazard ratio or 1 minus the incidence rate ratio. We modelled the non-linear eff ects of age at vaccination and local parasite prevalence as multiple fractional polynomials, according to the method of Royston and colleagues. 21 We fi tted changes in vaccine effi cacy over time as an interaction between time and vaccination status in Cox regression models, using the Anderson Gill modifi cation, 22 with clustering by individual to include multiple episodes.
We examined parametric survival models to fi t a γ distribution to the unmeasured heterogeneity in exposure. We used a Gompertz survival distribution for parametric models, since this method fi tted the data better than the alternatives (exponential, log normal, or Weibull) and gave hazard ratios for vaccination that

Poisson regression for all episodes of clinical malaria (ACDc/PCD)
Hazard ratio (95% CI) p Hazard ratio (95% CI) p Incidence rate ratio (95% CI) p Age (  were nearly identical to those estimated using Cox proportional-hazards.

Role of the funding source
We did this pooled analysis after a call for proposals initiated and facilitated by GlaxoSmithKline Biologicals. Employees of GSK Biologicals were investigators on the original phase 2 studies and were authors on the primary reports. Employees of GSK Biologicals reviewed the analysis plan and commented on early drafts of the pooled analysis, but were not required to give fi nal approval of the manuscript. The corresponding author had full access to all the data in the study and had fi nal responsibility for the decision to submit for publication.

Results
We analysed pooled data for 4453 participants in seven trials (table 1). 1376 participants received all three vaccinations, were given curative antimalarial treatment, and underwent ACDi. 3184 participants received all three vaccinations and were monitored for episodes of clinical malaria, either by ACDc or PCD. 465 adults received one or more vaccination; these data were excluded from the analysis for severe malaria or death. 3988 children (ie, younger than 6 years at vaccination) received one or more vaccination and were included in intention-to-treat analyses for severe malaria or death.
Unadjusted effi cacy against severe malaria was 37% (95% CI 6 to 58, p=0·023); data were from 39 children with severe malaria from a total of 2080 RTS,S vaccinated people, versus 58 children with severe malaria from a total of 1908 controls. Effi cacy against death was 48% (-8 to 75, p=0·081); 11 deaths occurred in the 2080 people receiving vaccine and 19 deaths happened among the 1908 controls. We judged the frequency of severe malaria and death to be too low to justify further multivariable analysis.
The survival plot of time to infection during ACDi by vaccination status shows convergence after the initial divergence (fi gure 1A), and the plot of time to clinical malaria by ACDc and PCD shows a gradual slowing in the rate of divergence (fi gure 1B). An interaction between effi cacy and time gave similar goodness of fi t (judged by Akaike's information criterion) for various powers of time (2, -1, -2, 0·5, 0·25) and linear and log functions. We therefore selected a linear fi t for simplicity and to make interpretation of the interaction terms more intuitive (fi gure 4).
In unadjusted analysis of ACDi, effi cacy seems to wane rapidly (fi gure 4A), but after adjustment for local parasite prevalence and for a γ-distributed shared frailty ( =0·96, p<0·0001, in dicating that signifi cant evidence exists for  Every row represent coeffi cients from a single model. The fi xed eff ect of vaccination refl ects the hazard ratio associated with vaccination at 0 years. The interaction term refl ects the change in hazard ratio associated with every year since vaccination. *Transmission intensity is a fi xed eff ect to account for known variation in exposure to malaria, and γ-distributed shared frailty accounts for unknown variation in exposure to malaria. ACDi=active case detection for infection. θ pronounced heterogeneity of risk), the eff ect of vaccination did not vary by much over time (fi gure 4B,  table 3). In unadjusted and adjusted analyses of ACDc and PCD, including single and multiple clinical episodes, the estimated vaccine effi cacy fell over time, from 36% effi cacy (95% CI 24 to 45) at time of vaccination to 0% (95% CI -38 to 38) at 3 years (fi gures 4C-E).

Discussion
The fi ndings of our pooled analysis show that the RTS,S malaria vaccine is protective against infection and disease. However, unadjusted effi cacy against clinical malaria was lower than previous estimates in children age 5-17 months 12 and substantial heterogeneity was noted in effi cacy between population subgroups and over time. Vaccine effi cacy against clinical malaria was lowest at high (70%) transmission intensity, and it was reduced for the AS02 adjuvant compared with AS01. Weak variation in effi cacy was noted according to age on Poisson regression, which was not signifi cant on Cox regression. Vaccine effi cacy did not vary by gender or bednet use. Results for effi cacy from Cox regression for fi rst episodes and Poisson regression for all episodes were similar, although CIs suggested greater precision when all episodes were included.
A higher vaccine effi cacy with PCD versus ACDc might indicate bias resulting from a prophylactic eff ect of antimalarial drugs administered for episodes of malaria that do not meet the case defi nition. These malaria episodes are likely to be more common in unvaccinated children and hence could result in an underestimate of vaccine effi cacy on ACDc. However, no sites used PCD and ACDc alongside each other, hence there is confounding by site and the diff erence might refl ect other variations between sites that were not measured by the available covariates. To examine whether additional unmeasured factors that segregate by site might lead to varying effi cacy, we fi tted a post-hoc interaction term between vaccination and stratifi cation by site, in addition to the previous model (table 2). These additional interactions signifi cantly improved model fi t (p<0·0001, by likeli hood-ratio test), indicating that other unmeasured factors cause vaccine effi cacy to vary between sites.
Transmission intensity (as measured by local parasite prevalence in children age 2-10 years) had a non-linear eff ect on clinical malaria incidence. 23 The incidence of clinical malaria reached a peak in areas with a local parasite prevalence of 40%. This fi nding could be accounted for by children who acquire greater immunity with increasing exposure, which off sets the rises in incidence of clinical malaria that otherwise might be seen at a higher local parasite prevalence.
RTS,S can be regarded as a leaky barrier to infection, because it protects against some infectious bites but not against others. 24 The probability of protecting a par ticipant exposed to two infective bites during the course of a night against a subsequent episode of clinical malaria is half the probability of protecting a participant exposed to one bite. This statistic suggests that vaccine effi cacy will be lower at high transmission intensity, which accords with our observations. We used MAP estimates of age-adjusted prevalence of asymptomatic malaria (PrP 2-10 ) to gauge transmission intensity. These approximations were based on several thousand surveys in the countries where trial sites were located. We chose these standardised independent measures rather than within-trial factors, such as incidence of malaria among controls, because monitoring was not the same between trials. MAP estimates do not account for changes over time, but transmission intensity at our sites is likely to be stable enough over a few years for these data to be a reasonable approximation. Variations in seasonal transmission have a modest eff ect on the relation between entomological inoculation rate and asymptomatic parasitaemia, but in view of the limitations of using data from 11 sites, we did not feel that more complex characterisations of transmission intensity were warranted.
Our fi nding that vaccine effi cacy is not aff ected by use of insecticide-treated bednets but is diminished at higher levels of transmission intensity might seem contradictory, since bednet use might be expected to reduce exposure and, hence, enhance vaccine effi cacy. However, individual use of insecticide-treated bednets might be only modestly protective (compared with greater mass eff ects at reducing transmission when whole com munities use bednets). Furthermore use of insecticide-treated bednets was not distributed evenly by site, varying from 4·5% to 100% of children. Tests for variation in vaccine effi cacy by bednet use are, therefore, vulnerable to ecological confounding by site.
We identifi ed signifi cant interactions between time and vaccine effi cacy. We can confi dently reject the null hypothesis that vaccine effi cacy is constant over time (p<0·0001), but we cannot be confi dent about the shape of the plotted decline, which is refl ected in the wide CIs surrounding estimates of effi cacy at later timepoints (fi gure 4). We chose a linear interaction for simplicity of presentation, although power functions of time fi t the data slightly better. Therefore, we cannot extrapolate beyond the data to longer durations of follow-up, since the shape of the line is determined by statistical convenience rather than biological understanding.
A fall in vaccine effi cacy over time might be attributable to systematic bias in estimates obtained using survival analysis because of heterogeneous exposure from a partly eff ective vaccine, as previously described. 25 Com parison of fi gure 4A with fi gure 4B suggests that systematic bias resulting from heterogeneous exposure can account for the apparent waning of effi cacy in the ACDi dataset, rather than a genuine biological waning of effi cacy taking place. However, similarity between fi gures 4C-E suggests that no clear systematic bias exists in estimates of effi cacy over time when using data from ACDc or PCD for clinical malaria. The diff erence could be because ACDi was monitored during one trans mission season, with only 60% of participants having an episode during this period. A few unexposed individuals can lead to a biased estimate of rapidly declining effi cacy. 25 However, data for clinical malaria included 4 years of monitoring during many transmission seasons. Furthermore, individual exposure could vary from year to year. 26 Hence, a discrete unexposed popu lation is less likely to exist with ACDc or PCD compared with ACDi. Heterogeneous exposure, therefore, seems to be a suffi cient explanation for the observation that effi cacy wanes more rapidly in the ACDi dataset than it does in the ACDc and PCD dataset.
We report here all phase 2 data for RTS,S malaria vaccines (panel), including effi cacy outcomes for clinical malaria (ACDc and PCD) and for malaria infection (ACDi). Some phase 2 trials also included crosssectional surveys for asymptomatic parasitaemia. In the Mozam bique trial, substantial protection against asymptomatic parasitaemia was noted 45 months after vaccination in cohort 1, 14 which is longer than would have been predicted by our analysis. On the other hand, protection against asymptomatic parasitaemia was not noted in Mozam bique cohort 2 at 21 months after vaccination. This fi nding could be explained by diff erential acquisition of blood-stage immunity between these two cohorts. 27 Our data do not allow us to distinguish waning vaccine-induced immunity from delayed acquisition of blood-stage immunity, but analysis of the eff ect of the booster vaccinationplanned as part of the phase 3 trial-is likely to be highly informative. A booster dose can restore vaccineinduced immunity but will not have an immediate eff ect on immunity to blood-stage parasites. Furthermore, the larger sample size in the phase 3 trial will provide more accurate point estimates for effi cacy in the age-groups assessed (ie, 6-12 weeks and 5-17 months) than is possible in a meta-analysis of phase 2b data.
In summary, we noted signifi cant variation in estimated vaccine effi cacy by population subgroups and a signifi cant decline in protection against clinical malaria over time. One might argue that the unadjusted pooled estimates of effi cacy nevertheless refl ect what was actually seen in the population tested. However, the unadjusted pooled effi cacy is merely a weighted mean of effi cacies seen in the component subgroups of the population and, therefore, cannot be generalised to other populations. For instance, if the vaccine is more eff ective at lower transmission intensity, the pooled vaccine effi cacy will depend on the proportion of children recruited in sites at low transmission intensity. Predictions of public health outcomes of vaccination will need to take account of these variations in effi cacy by transmission intensity and by time since vaccination.

Contributors
PB, MTW, and ACG designed the study and were involved in interpretation and writing. AO, KB, and STA were involved in data collection and interpretation. All other authors contributed to data collection, interpretation, and writing.

Confl icts of interest ACG acknowledges support from the Bill & Melinda Gates Foundation
Vaccine Modelling Initiative and MRC Centre funding. GSK Biologicals did not sponsor the investigators to do the analysis, but did fund transport expenses for investigators to attend a meeting to discuss their progress.