This re-analysis of a large Kolkata cholera vaccine trial showed that inclusion or exclusion of the large outbreak in the fourth year of the trial and treating age group as a fixed or time-dependent covariate could have large impacts on the estimated covariate effects on culture-confirmed cholera. The preferred model fit is a hybrid model fit to the outbreak-free data that lets the age group effect be associated with time-dependent age group and the vaccine effect be associated with age group at baseline. Our re-analyses provide additional evidence that the efficacy of OCV varies between different age groups, and suggest that the efficacy among the older children and adults lasted five years in this cholera-endemic setting. We also looked for evidence of waning of efficacy, and we found no evidence, at least among adult vaccinees.
The trial had not originally been powered to estimate differences in VE across age groups, but our re-analysis found more definitive evidence supporting age group-dependent VE than the previous analysis. We found VE to be higher among older children than among young children (p = 0.002) or adults (p = 0.106). The overall significance is 0.007. The same trend had also been observed in an earlier analysis using only 2 years of follow-up but that result was not statistically significant (overall p = 0.07), probably due to the smaller number of events observed [8]. In the previously published analysis using all 5 years of follow-up, adults had the highest estimated VE [7]. The lower VE among older children observed in that study, compared to both the 2-year analysis and ours, can be explained by the inclusion of the March–April 2010 outbreak, which appeared to have disproportionately affected older children (Additional file 1: Table S3). Overall, our age-specific OCV efficacy point estimates of 38%, 85%, and 69% for children under 5 years old, children older than 5 and younger than 15, and individuals 15 years old or older, respectively, look more similar to the published estimates that used the first two years of follow-up (49%, 87%, and 63% [8]) than those using all five years (42%, 68%, and 74% [7]).
The age of participants changed substantially during the 5 years of follow-up. When age enters into a model nonlinearly, e.g., as an age group variable, whether the age is treated as a fixed or time-dependent variable impacts the outcome of the analysis [18]. In most published analyses of clinical trials, age has been treated as fixed. One reason for this may be that the duration of most trials is not long enough to make a material difference to whether or not the age variable is treated as a time-dependent variable. The Kolkata trial analyzed here has 5 years of follow-up and our analyses have demonstrated the importance of treating age group as a time-dependent variable.
An innovative aspect of our analyses is that we decouple the age group that impacts the vaccine effect covariate and the age group that impacts the natural risk. There are four possible combinations: B/B, B/T, T/T and T/B, where the first letter denotes whether VE depends on baseline or time-dependent age group and the second letter denotes whether natural risk depends on baseline or time-dependent age group. A priori, it seems biologically more plausible that the natural risk should depend on the time-dependent age group and not on the age at vaccination. This is because as people age, their behavior and life history change and those are likely to impact cholera risk. It also seems more likely that vaccine efficacy depends more on the age group at vaccination than on the time-dependent age group, since the vaccine-elicited immune responses differ by the age at the time of vaccination [19]. Based on these reasonings, the B/T is the most biologically reasonable model while the T/B model is the least. Results from the four fitted models (Table 1 and Additional file 1: Table S4) support the idea that the B/T model is the best model, and the fitted B/T model gives us additional insights over the more conventional B/B model. The estimated hazard ratio between the middle and youngest age group was 0.22 by the B/T model, but the B/B model gave a more attenuated estimate 0.35. This attenuation can be explained by the fact that the younger children group as defined by the B/B model is actually a mixture of the younger and older children at the time of infection.
The occurrence of a major outbreak during follow-up presents an analytical challenge. Our choice of removing the outcomes and censoring events occurring during the outbreak is a simple approach, and it allows us to cleanly assess vaccine efficacy and the effects of other covariates under non-outbreak situations. For example, removing the outbreak from the analysis set results in an upwards shift of the estimated VE among the older children (ages 5 to 15 years) from 0.71 to 0.85. Furthermore, in combination with the use of the B/T model for encoding age group, the evidence for the age group-vaccine treatment interaction grows from borderline (p-value 0.082) to highly significant (p-value 0.007). It should be noted that vaccinated individuals were protected during the outbreak and exclusion from the analysis does not imply a lack of efficacy during acute outbreaks. Exclusion of the censored and infected subjects during the outbreak has some drawbacks. For example, it is difficult to determine how it may have affected the randomization. We had considered other approaches. One temptation is to use an indicator variable to denote the outbreak period and include the variable in a model to analyze the full dataset. The advantage of this approach is that all data is put to use, but the model is difficult to interpret because the outbreak is not an attribute of the subjects, but rather an outside force that happens at a particular time and place. One may also envision a competing risk model where the events during the outbreak are treated as coming from an alternative cause. As the outbreak only occurs in a relatively short period, there is probably little benefit to take this more complicated approach.
An interesting observation from our analysis is that the cholera risks of a vaccinated older child (ages 5 to 15 years) and a vaccinated adult (15 years old and older) are very close, despite the fact that unvaccinated adults have a lower risk of cholera than an unvaccinated older child. Using young children (ages 1 to 5 years) from the placebo group as reference, the risks experienced by both aforementioned groups are 30-fold lower. This suggests that the vaccine reduced the risk level of those 5 years old and older to a similarly low level.
The results presented here may be relevant for future routine OCV deployment plans in endemic regions where cholera strikes frequently. Our analyses show that OCV could provide 5 years of high-quality protection for those over 5 years old with no evidence of waning among adults, suggesting that they might not need to be re-vaccinated frequently to maintain protection, making mass vaccination more cost-effective in cholera-endemic regions. However, younger children are at high risk of cholera in this population and OCV may confer lower protection, thus re-vaccination when they are older may be desirable.
The duration of protective efficacy of OCV may differ in endemic and epidemic regions. Repeated exposure to V. cholerae may serve as natural boosters for vaccinees, prolonging their protection [7, 20]. Therefore, vaccine protection may appear to wane more quickly in populations exposed to cholera less frequently. OCV use outside Asia has largely been to prevent or mitigate outbreaks in populations that do not see cholera every year [21]. OCV was recently used in reactive campaigns during outbreaks in Guinea and Haiti, and the VE estimated during these outbreaks was consistent with the estimates from Kolkata reported here [22, 23], but the duration of protection in these populations is not yet known. Additional studies are needed to estimate how long two doses of OCV can confer protection in populations outside well-studied South Asian endemic regions.
