Pneumococcal conjugate vaccine (PCV) effectiveness (VE) against invasive pneumococcal disease (IPD) varies across Streptococcus pneumoniae serotypes (STs). Combining data from multiple sites helps obtain reliable serotype-specific VE estimates but introduces a hierarchical structure. Ignoring this structure can bias results and overstate statistical significance. This study compared Bayesian hierarchical model (BHM) and frequentist fixed-effect model (FFEM) to estimate serotype-specific VE using multisite surveillance data. The indirect cohort method was used to estimate serotype-specific VE by immunization status and time since the last dose. IPD cases among children aged 2-59 months were identified according to laboratory and clinical criteria in four Canadian provinces from 2010 to 2019. The FFEM adjusted for province as a fixed effect, whereas the BHM modeled province as a random intercept to account for province-level variability. Intraclass correlation coefficient (ICC) quantified the proportion of variance attributable to province. A total of 931 IPD cases were included in the analysis. Based on BHM, ICCs indicated that province-level variation accounted for more than 20% of total variance for ST-7F and ST-19F and approximately 5% for ST-3 and ST-19A. Under PCV13 ≥1 dose, VE was 90% (95% credible intervals CrI 68 to 97) for ST-19F, 86% (95%CrI: 37 to 97) for ST-7F, 63% (95%CrI: 26 to 81) for ST-19A and 56% (95%CrI: -9 to 82) for ST-3. VE waned over time, particularly for ST-3, decreasing from 80% (95%Crl: 39 to 94) within 12 months after the third PCV13 dose to -44% (95CrI: -272 to 43) thereafter. When compared with FFEM, BHM produced more stable VE estimates in data-sparse strata and yielded more conservative credible intervals. FFEM, by contrast, often failed to converge. PCV vaccine effectiveness varies across serotypes. Bayesian hierarchical modeling is preferred for estimating serotype-specific VE from multisite studies with sparse or clustered data.
Zhou et al. (Fri,) studied this question.