We identified datasets from IPD surveillance systems that report rates through two approaches. First, we identified datasets gathered from a comprehensive systematic literature review on PCV dosing schedules [11]. In that systematic literature review, a search for English language publications on the immunogenicity, and direct and indirect effects of various PCV schedules on nasopharyngeal (NP) carriage, IPD, and pneumonia among children was performed using 14 databases (i.e., African Index Medicus; BioAbst/Reports, Reviews, Meetings; Biological Abstracts; Cochrane Library; EMBASE; Global Health; Index Medicus for Eastern Med. Region; Index Medicus for South-East Asia Region; IndiaMed; Latin America and Caribbean Health Sciences Information; Pan-American Health Organization; Pascal Biomed; PubMed; and Western Region Index Medicus) as well as meeting abstracts of the International Symposium on Pneumococci and Pneumococcal Disease (ISPPD) and the Interscience Conference on Antimicrobial Agents and Chemotherapy (ICAAC). The search included studies published between 1994 and 2010. The complete list of database-specific and Medical Subject Headings (MeSH) search terms used in the literature search is detailed by the authors. We reviewed those publications with IPD as an outcome; these publications needed to include at least one “narrow vaccine” search term as well as an IPD related search term, i.e., (“Invasive disease” [all fields]), (“invasive pneumococcal disease” [all fields]), and/or (“invasive bacterial disease” [all fields]).

Second, we solicited potential datasets from experts in pneumococcal disease, WHO headquarters and regional offices, and by reviewing references from publications.

We solicited datasets from investigators using a standardized format, requesting IPD case counts for up to 5 years before and 10 years after PCV7 introduction, stratified by age groups (0–1, 2–4, 5–17, 18–49, 50–64, and ≥65 years old), clinical syndrome (overall IPD and meningitis specifically), hospitalization status, and serotype (Text S1). Meningitis was defined as isolation of pneumococcus from cerebrospinal fluid by culture. We requested age- and year-specific catchment population denominators to estimate rates, and we solicited descriptions of the PCV7 vaccination program, IPD surveillance system, changes to surveillance methodology or clinical practices, and potential IPD outbreaks.

Two coordinators conducted a quality check of datasets included in the analysis using a checklist (Box 1). Any requests for data clarification were emailed to the contributing investigator and the data were updated as applicable.

The inclusion criteria of the datasets for collection and analysis are given in Figure 1.

In datasets where serotypes 6A and 6C were not differentiated, we distributed these serotypes according to the known distribution of 6A and 6C in the same geographic region or globally in the pre- and post-PCV7 periods [12]. First, we calculated the percentage of 6A isolates out of all 6A and 6C isolates, using datasets where 6A and 6C isolates were distinguished. The percentage of true 6A isolates was calculated for all datasets, as well as by region for datasets from Europe and North America. Estimates of the percentage of true 6A isolates were weighted by the size of the site and calculated in four different time periods: pre-PCV introduction; 1–2 years post-; 3–4 years post-; and 5+ years post-PCV introduction. In sites that did not differentiate 6A and 6C serotypes, 6A/6C isolates were then redistributed according to the estimated regional (for North America and Europe datasets) or using all datasets’ (for datasets from sites outside of Europe or North America) distribution of differentiated 6A and 6C isolates.

After redistributing serotype 6A with VT (serotypes 4, 6B, 9V, 14, 18C, 19F, and 23F) and 6C with NVT (all other serotypes), remaining isolates with unknown serotype were redistributed. Specifically, isolates with known serotype were classified into four groups: VT serotypes (PCV7 serotypes and 6A); serotypes 1 and 5; serotypes 3, 7F, and 19A; and all other NVT serotypes. Serotypes 1 and 5 were grouped together to allow for modeling expected rates absent the potential influence of outbreaks of these two serotypes. The remaining additional serotypes included in higher valency PCVs–3, 7F, and 19A—were grouped together for analyses of changes over time as they, along with serotypes 1 and 5, are likely to be prevented by introduction of the higher valency PCVs. Non-typeable isolates were added to the category of all other NVT serotypes. We calculated the percentage of each of the four groups of known serotypes out of all known serotypes. Unknown isolates that were not serotyped were then redistributed into the four serotype groups per the calculated proportional distribution. Redistribution was performed by site, age group, year, and syndrome.

To minimize the effect of temporal and geographic differences in blood culturing practice among children in the outpatient setting, we restricted our analysis to hospitalized cases for children <5 years [13],[14]. Since IPD among adults is almost always a severe illness among inpatients, we assumed all cases among older persons were hospitalized. This assumption was confirmed in the few sites that did capture data on hospitalization status among adults with IPD [59],[60]. We excluded persons aged 5–17 years because case counts were too small for meaningful analysis.

Because IPD rates were changing before PCV7 introduction in some sites, we used the pre-PCV7 IPD trends (excluding the year of introduction) to predict future years’ IPD rates, absent PCV7 use [15],[16]. We used Poisson regression to model expected rates of VT, NVT, and overall IPD absent PCV7 introduction. We assumed that overall IPD was a more stable indicator of pre-PCV7 trends than either VT or NVT, which could be affected by outbreaks of a single serotype [16]. Therefore, we used the regression intercept and slope of the pre-PCV7 annual rates of overall IPD to estimate future rates, absent PCV7, for overall, VT, and NVT IPD. Because serotypes 1 and 5 rates can fluctuate annually owing to outbreaks, we excluded them from the regression estimation of pre-PCV7 trends, but included them in the actual rate estimates on the basis of the trends. Separately, we calculated the pre-PCV7 average proportions of IPD caused by VT and NVT and applied each to the expected overall IPD rate to generate the expected VT and NVT IPD rates. The annual surveillance population denominator was included as an offset variable, and the slope of the modeled expected rates was assigned a value of zero from 4 years post-PCV7 onwards, assuming stabilization of any pre-PCV7 IPD surveillance trends by then.

For children aged <5 years, expected rates for 11 of 19 sites (58%) were modeled. Among the 15 sites included in the IPD analysis for adults aged 18–49 years, 50–64 years, and ≥65 years, expected rates were generated using modeling for 10 (67%), 5 (33%), and 7(47%) sites, respectively. For age strata with an annual pre-PCV7 average of <20 IPD cases or <3 years of pre-PCV7 data, we felt that pre-PCV7 rates were unreliable to define surveillance trends because of small sample size or too few years. For these strata, expected IPD rates absent PCV7 introduction were estimated by averaging annual IPD rates before PCV7 introduction.

We estimated the change in IPD rates following PCV7 introduction by calculating rate ratios (RRs), dividing the observed IPD rate by the expected IPD rate for each post-PCV7 year. We calculated 95% confidence intervals around RRs through simulation of observed and expected case counts and the delta method [17]. The delta method can be used to approximate the variance of a ratio and has previously been applied to estimate the variance of the log RR [17],[18]. To estimate the variance of the log RR, we simulated 200 observed and expected case counts using the Poisson distribution with the actual observed and calculated expected number of cases as the mean. We converted these simulated observed and expected case counts to rates. From these simulated rates we calculated the variance of the observed and expected rate, as well as the covariance between these rates using STATA Version 12.1 (StataCorp.).

Using the delta method formula below, we combined the variance of the observed and expected rate to estimate the variance of the log RR.

Where σ² is the variance; Y is the observed rate; X is the expected rate; and COV (X,Y) is the covariance between the observed and expected rate.

We included the covariance in the calculation of the variance of the log RR because for a few strata the covariance was greater than zero and so we were unable to assume independence between the observed and expected rates. The square-root of the variance of the log RR was used to estimate the standard error of the log RR. The standard error of the log RR was calculated separately for each site, age group, serotype combination, and post-PCV7 year.

A value of 0·5 cases was assigned as a continuity correction to each stratum (i.e., site-age group-serotype group) with zero cases reported [19] so as to avoid undefined RRs (when zero cases of IPD were expected in a year) or undefined variances (as Poisson simulation would generate missing values for zero cells).

Because the impact of PCV7 was expected to be heterogeneous across sites, we used random effects meta-analysis to pool the site-specific RRs [20]. Meta-analysis was performed for each age and serotype group for each of the 7 years after PCV7 introduction, generating a summary RR with 95% confidence intervals. Meta-analysis of RRs was performed both including all datasets available for each year post-PCV7, as well as including only those datasets with at least 7 years of post-PCV7 data, which was the last year with enough datasets for robust meta-analysis (i.e., five datasets). The same analysis comparing observed and expected rates was performed limited to meningitis cases.

We performed several sensitivity analyses for IPD. First, we used a continuity correction of 0·1. Second, we performed an analysis completely excluding serotypes 1 and 5 from both pre- and post-PCV7 IPD rates. Third, we performed the analyses with the expected IPD rate as the observed average pre-PCV7 introduction IPD rate for all site-age group-serotype group strata (i.e., no modeling of expected IPD rates).

Additionally, we performed an analysis comparing observed and expected IPD rates for two separate NVT serotype groups: NVT serotypes in the higher valency pneumococcal conjugate vaccines that are not in PCV7 (i.e., serotypes 1, 3, 5, 7F, and 19A) and NVT serotypes not in the higher valency vaccines. The RR of the observed over the expected rates in the years after PCV7 introduction and 95% CI were calculated for each site, age, and year stratum for both of these categories of NVT. A summary RR for both NVT categories was obtained for each age group in each post-PCV7 year using random-effects meta-analysis.

To compare the contribution of these two NVT categories to the overall IPD incidence post-PCV7 introduction, we performed a separate analysis restricted to the post-PCV7 period where we defined the RR as the observed rate of IPD due to the NVT included in the higher valency vaccines over the observed rate of all other NVT not included in those vaccines. The 95% CI for this RR was also calculated using the delta method for each site, age, and post-PCV7 year. A summary RR for each age group and post-PCV7 year was calculated using random-effects meta-analysis.

The analysis dataset was generated using SAS Version 9·2 (SAS Institute Inc.). Meta-analyses were conducted using STATA Version 12·1 (StataCorp).

We identified 72 potentially eligible datasets and requested information from the investigators (Figure 1). Of 32 datasets received, 21 from four geographic regions (six North America, 11 Europe, three Australasia, one South America) met the inclusion criteria for analysis (Figure 1). For children, 19 datasets were included in the IPD and meningitis analyses, although two sites were included only for IPD and two different sites were included only for meningitis. For adults, 15 and 11 datasets were included in the analyses of IPD and meningitis, respectively. At least 19 datasets included in the analysis have previously published IPD surveillance data, though not necessarily including the same data used for this analysis (i.e., age group, case population, syndrome, and years of surveillance) [3],[4],[16],[21]–[36].

Specific reasons for exclusion from analysis for 11 datasets received were the following: no denominator provided (one); serogroup 19 not serotyped (one), <70% coverage of the primary PCV7 series by 12 months of age (four) [37]; <2 years of pre-PCV7 data (three); inability to define a proper denominator population (one); and substantial changes over time in case ascertainment of the surveillance system (one). The average annual number of IPD cases pre-PCV7 introduction for the 11 datasets excluded (six Europe, three North America, one Africa, and one Western Pacific) ranged from 8–1,490. Furthermore, among datasets included, two and six site-age group strata were excluded from the IPD and meningitis analyses, respectively, because <50% of isolates in those strata were serotyped (Table 1). In one site, adult strata were excluded from the analysis due to an increase in VT cases in the post-PCV7 introduction period, indicating changes in surveillance or bias that would affect the analysis (Table 1). No sites were excluded due to implausible distributions of serotype 6A/6C isolates.

The PCV7 schedules used included two primary doses plus a booster (nine sites), three primary doses without a booster (one site), and three primary doses with a booster (11 sites); 16 sites had catch-up campaigns (Table 2). All sites achieved ≥70% immunization coverage during the surveillance period and the range of average immunization coverage estimates for all post-PCV7 years was 55%–97% (Table S1).

The annual number of IPD isolates at baseline for children <5 years ranged from 2 to 690 and the median baseline rate was 31·4 cases per 100,000 (range 4·7–280·3) (Figure 2; Table 2). Our meta-analysis showed that the rate of overall IPD decreased significantly by 1 year after introduction (summary RR 0·55, 95% CI 0·46–0·65), which was then maintained out to 7 years post-introduction (RR 0·49, 95% CI 0·35–0·68) (Figure 3; Table 3). Although there was heterogeneity in the effect across sites, as expressed by the I² statistic, the point estimates tended in the same direction with all 19 sites showing a decrease (in 15, these reductions were statistically significant) compared to baseline in overall IPD in at least one post-introduction year (Figure 4). The rate of VT IPD declined significantly by 1 year after introduction (summary RR 0·34, 95% CI 0·28–0·41) and continued to decrease through 7 years (summary RR 0·03, 95% CI 0·01–0·10) (Figures 3 and 5; Table 3). The rate of NVT IPD increased significantly by 2 years after introduction (summary RR 1·34, 95% CI 1·02–1·77) and increased through 5 years, with little change thereafter through year 7 (summary RR 2·81, 95% CI 2·12–3·71) (Figure 3; Table 3). Most sites (seven statistically significant) showed an increase in NVT IPD rate in at least one post-introduction year (Figure 6). To account for the possible confounder of varying numbers of datasets included by year after PCV introduction, we repeated the meta-analysis including only the five sites with 7 years of post-PCV7 data. For VT, NVT, and all serotypes, the summary RRs were similar to those when all sites were included (Tables 3 and S2). The results were also similar when using a continuity correction of 0·1 instead of 0·5 (Table S3) and when excluding serotypes 1 and 5 (Table S4). In the analysis in which all expected rates used the average pre-PCV7 rates (i.e., no modeling of expected rates), the trends of post-PCV7 IPD changes were similar to those from the modeling approach, although NVT summary RRs tended to be slightly higher, as would be expected with no adjustment for increasing surveillance sensitivity over time (Table S5).

In the pre-PCV7 period, the percentage of IPD due to meningitis ranged from 3%–34% by site (Table 2). The meta-analysis results for meningitis were similar to overall IPD, with sustained reductions in meningitis due to all serotypes through 7 years post-PCV7 introduction (RR 0·40, 95% CI 0·25–0·64) (Figure 7; Tables 4 and S6). Due to smaller numbers of meningitis cases, there was more variability by year and wider confidence intervals for the RR point estimates (Figures and 7; Tables 3 and 4).

For adults, the annual number of IPD isolates at baseline ranged from 3 to 4,929 with a median IPD baseline rate of 14·2 cases per 100,000 (range 0·6–101·7) (Figure 2; Table 2). The summary RR point estimates from the meta-analysis showed reductions in overall IPD for most years, though not statistically significant in years 1–6 post-introduction (Figures 3, 8, 11, and 14; Table 3). Among the five sites with data 7 years post-introduction, statistically significant reductions were seen in year 7 for persons 18–49 years (summary RR 0·52, 95% CI 0·29–0·91), for persons 50–64 years old (summary RR 0·84, 95% CI 0·77–0·93), and for persons ≥65 years old (summary RR 0·74, 95% CI 0·58–0·95) (Figures 8, 11, and 14; Table S2). VT IPD decreased significantly for all adult age groups by the second year after PCV7 introduction (Figures 3, 9, 12, and 15; Table 3). In contrast to children, this decrease in VT IPD rates occurred more gradually; not until the fourth year after PCV7 introduction did adults have decreases in VT IPD similar in magnitude to those seen among children in the first post-PCV7 year (Figure 3; Table 3). In adults aged 18–49 years old, there was no significant increase in NVT IPD rates compared to baseline for any year, while for adults aged 50–64 years and ≥65 years, significant increases in NVT IPD were observed from year 2 and year 1 post-introduction, respectively (Figures 3, 10, 13, and 16; Table 3). There was substantial variability in the magnitude of NVT IPD increase by site (Figures 10, 13, and 16). For adults, the meta-analyses using a 0·1 continuity correction, excluding serotypes 1 and 5, limited to the five sites with 7 years of data, and using only averaged pre-PCV7 rates showed similar findings (Tables S2–S5).

Among all adults in the pre-PCV7 period, the percentage of IPD due to meningitis ranged from 0%–8% by site (Table 2). The findings for meningitis were similar to overall IPD for 18–49 year olds, with statistically significant reductions at 7 years post-PCV7 introduction (RR 0·61, 95% CI 0·40–0·95) (Figure 7; Tables 3 and 4). For persons 50–64 years old, in most years the increase in NVT meningitis tended to be higher than for NVT IPD, resulting in some early years (i.e., years 2 and 3) when there was an increase in overall pneumococcal (i.e., any serotype) meningitis, although this significant increase was not sustained in subsequent years (Figure 7; Tables 3 and 4). In contrast to 50–64 year olds, among persons ≥65 years there was less of an increase in NVT meningitis than NVT IPD in most years, resulting in relatively greater reductions in overall meningitis due to all serotypes, although never reaching a statistically significant decrease (Figure 7; Tables 3 and 4).

The magnitude of increases in IPD rates due to the subset of NVT included in higher valency conjugate vaccines but not PCV7 (i.e., serotypes 1, 3, 5, 7F, 19A) was similar to the increases among all the other NVT not in the higher valency vaccines (Table 5). However, the rates due to IPD caused by the five NVT included in higher valency vaccines were higher than rates of the NVT not in the higher valency vaccines in most post-PCV7 years for children (Table 6). In contrast, among adults aged 50–64 years and ≥65 years old, IPD rates of NVT not in the higher valency vaccines were higher than rates caused by the NVT in the higher valency vaccines for most years (Table 6).