Similar to a previous cost-effectiveness analysis [13], we developed a probabilistic model to estimate the health impact and cost-effectiveness of replacing PCV13 with PCV15 in a hypothetical cohort of children. We used Monte Carlo simulation in spreadsheet-based software (@Risk 8.2; Palisade Corporation, Newfield, NY) to predict incremental cases of pneumococcal disease events and deaths averted, costs per quality adjusted life-year (QALY) gained, and costs per life-year gained under PCV15 vs. PCV13 schedules.

We modelled the effects of PCV13 and PCV15 separately on VT IPD, VT pneumococcal pneumonia, and VT pneumococcal AOM. Our model tracked disease incidence for 15 years after the last PCV dose, although costs of sequelae, lost life-years, and quality-adjusted life-years were tracked through life expectancy. The general model structure is shown in Fig. 1. The population of the model represented the 2020 U.S. birth cohort (N = 3,939,295 in year 1) and used life expectancy and background mortality estimates by age from the National Vital Statistics System [14].

Disease burden incidence inputs in the model varied from 0 through 9 years of age and were constant from 9 through 20 years of age. Pneumococcal serotypes were classified into five categories that were not mutually exclusive: PCV13-serotypes (1, 3, 4, 5, 6A, 6B, 6C, 7F, 9 V, 14, 18C, 19A, 19F, 23F), serotype 3 alone, serotype 19F alone, PCV15-unique-serotypes (22F, 33F), and non-vaccine serotypes (NVT, all serotypes not in PCV13 or PCV15). Serotypes 3 and 19F were modeled separately based on reports of reduced PCV13 VE against these serotypes [9]. While serotype 6C is not included in PCV13, it was included within the PCV13-serotypes category due to reported cross protection from serotype 6A [15].

IPD related parameters included incidence per 100,000 people; the proportion of IPD due to specified serotype groups; the proportion of IPD resulting in meningitis; and case fatality rates, all of which were obtained from the Centers for Disease Control and Prevention’s (CDC) Active Bacterial Core surveillance (ABCs) data during 2018–2019 [16]. Estimates of the proportion of individuals experiencing post-IPD meningitis sequelae of long-term disability or deafness, were informed from Olarte et al. [17], and Edmond et al. [18].

All-cause inpatient pneumonia incidence per 100,000 people and case fatality rates were obtained from National (Nationwide) Inpatient Sample (NIS) data during 2018–2019. The NIS dataset is the largest publicly available all-payer inpatient care database in the U.S. [19]. Age-stratified all-cause inpatient pneumonia incidence rates were obtained by identifying all inpatient hospitalizations among children aged < 18 years with a pneumonia-related International Classification of Diseases Diagnosis ICD-10-CM code in any location (Supplementary Material S1). The proportion of all-cause inpatient pneumonia hospitalizations that were pneumococcal pneumonia was informed from CDC’s Etiology of Pneumonia in the Community (EPIC) study by Jain et al. [20], and pneumococcal disease subject matter expert input. As the EPIC study reported no etiology in 20–30% of pneumonia hospitalizations, expert input considered the proportion of pneumococcal inpatient pneumonia reported in the study (4%) to be an underestimate, and thus recommended assumption of a higher proportion of 12% for the current analysis. This higher assumption was informed from studies among adults in North America, which found the proportion of pneumococcal pneumonia to range between 9 and 12% [21–24]. Data on serotype distribution of inpatient pneumococcal pneumonia was not available and was assumed to be the same as that for IPD.

All-cause outpatient pneumonia incidence per 100,000 was obtained from Tong et al., which used MarketScan Commercial Claims and Encounters data from 2014 [25]. We assumed no mortality from outpatient pneumonia. Based on pneumococcal disease subject matter expert input, the proportion of all-cause outpatient pneumonia events that were pneumococcal pneumonia was assumed to be half that of pneumococcal pneumonia in the inpatient setting (i.e., 6%). Similar to inpatient pneumonia, the serotype distribution for outpatient pneumonia was assumed to be the same as that for IPD.

All-cause AOM incidence per 100,000 was obtained from another study by Tong, et al., that also used MarketScan data from 2014 [26]. As the study by Tong et al. only reported the incidence of index AOM events, we estimated the incidence of recurrent AOM based on the proportion of recurrent episodes among AOM episodes, by age group, reported by Maron et al. [27]. We assumed no mortality from AOM. The proportion of children with AOM with consequent tympanostomy tube insertion surgery was obtained from Pichichero et al. [28]. As our estimates of all-cause AOM incidence were obtained from administrative data, we assumed that 60% of AOM events among children aged < 2 years and 70% of AOM events among children aged 2–5 years were true AOM diagnoses, of which 95% were bacterial infections; these proportions were informed by subject matter input. Additionally, in a study by Kaur et al., 24% of clinically diagnosed AOM cases were reported to be due to pneumococcus [7], as such, we assumed that 14% and 16% of AOM cases reported in administrative data were pneumococcal AOM events in children aged < 2 years and 2–5 years, respectively (Table 1). The study by Kaur et al. also reported serotype data on pneumococcal AOM cases and was used to inform the serotype distribution of pneumococcal AOM in this analysis.

During 2018–2020, 92.4% of children had received 3 PCV13 doses and 82.3% had received at least 4 PCV13 doses by 24 months of age [29]. We assumed the same coverage rate for both the PCV13 and PCV15 strategies among infants. For the assessment of the supplementary PCV15 dose among children 2 years and older, we assumed a 50% vaccine uptake.

Vaccine effectiveness (VE) of PCV was modelled as a percentage reduction in incidence of IPD, pneumonia, and AOM due to VT serotypes. Estimates of PCV13 VE against IPD due to serotypes 3 and 19F were obtained from Andrew et al. [9] while estimates of VE against IPD due to remaining PCV13 serotypes were obtained from Moore et al. [30]. The VE of PCV13 against inpatient and outpatient VT-pneumococcal pneumonia was estimated by applying the ratio of VE against VT-IPD to VE against VT-non-bacteremic pneumococcal pneumonia observed in the CAPiTA trial (75%:45%) [31] to IPD VE estimates used in this study. The VE of PCV13 against VT-pneumococcal AOM was obtained from a study by Escola et al. [32]. In order to account for lower VE against AOM due to serotypes 3 and 19F, the ratio of VE estimates against IPD by serotype was applied to the VE against VT-AOM estimate. For all disease outcomes, the VE of PCV15 against serotypes 23F and 33F was assumed to be the same as the VE of PCV13 against serotypes other than 3 and 19F.

During the first year of life when children receive up to 3 PCV doses, we estimated that children would have 75.6% of the full VE. This estimate was informed from a study by Whitney et al. [33]. We assumed full protection from PCV in the second year of life once a child had received the complete 3 + 1 schedule [33]. As there are limited data on duration of protection provided my PCVs, we adopted a conservative estimate that VE started waning 5 years after completion of the 3 + 1 schedule, with linear waning to 0% effectiveness over the next 10 years. Since vaccine protection expired 15 years after last PCV dose in our model, the analytic horizon for the model was selected to be 15 years.

Finally, as previous pediatric PCV introductions have been associated with indirect effects resulting in declines in pneumococcal disease due to VT serotypes among those who are unvaccinated [34], we assumed a similar decline in disease due to serotypes 22F and 33F within our cohort following PCV15 introduction. This indirect effect was incorporated into the model by removing 7.8% of serotype 22F and 33F pneumococcal disease each year, consistent with indirect effect estimates used in a previous PCV cost-effectiveness analysis [13].

Our analysis was performed from the societal perspective, including both medical and nonmedical costs. Medical costs were from MarketScan data [35] while nonmedical costs were informed from Ray et al. [36] and a Morbidity and Mortality Weekly Report [37]. For medical costs, conversions to 2021 dollars were done using the Consumer Price Index for medical care. For nonmedical costs, conversions to 2021 dollars were done using the Consumer Price Index for all items [38]. All outcomes were discounted by 3% annually. The public ($150.83) and private ($226.43) prices of a dose of PCV13 were obtained from CDC’s 2021 vaccine price list [39] and were weighted by public (61%) and private (39%) purchase shares from Pfizer’s (manufacturer of PCV13) internal sales data for 2021 (obtained through email communication). We assumed a vaccine administration cost of $15.04 [40] and travel/caregiver time cost as $33.30 [41]. At the time of this analysis, PCV15 (manufactured by Merck) did not have a published price for use within the pediatric population. Based on the price of PCV15 use among adults and following discussions Merck, we assumed the private price for PCV15 was $216.09 and the public price was $150.83.

To obtain a composite measure that could combine mortality outcomes with less severe health outcomes, we applied QALY decrements to each episode of disease. QALY decrements were informed from Tang et al. [42] with the exception of tympanostomy tube insertions, which were from Delgleeize et al. [43]. No loss of health was indicated by a decrement of 0, movement from perfect health to death had a decrement of 1, and decrements for non-fatal disease episodes were between 0 and 1. The specific decrements per episode of disease are detailed in Table 3.

In the assessment of PCV15 vs. PCV13 among infants, we conducted two univariate sensitivity analyses. First, we assumed that the public price for PCV15 was 5% higher than PCV13 at $158.37. Second, we assumed no indirect effects of PCV15.

In the assessment of a supplementary PCV15 dose for children aged 2–5 years old, we also carried out two univariate sensitivity analyses. First, we assessed impact when assuming a higher proportion of pneumococcal AOM (19% among children aged < 2 years and 23% among children ≥ 2 years). This was done because the high burden of AOM was considered a key driver for PCV cost-effectiveness and a recent study demonstrated a higher proportion of pneumococcal AOM when using multiplex polymerase chain reaction instead of traditional culture methods to confirm diagnosis [44]. In the second univariate analysis, we assumed a higher proportion of sequalae after meningitis (20% among children aged < 5 years and 30% among children aged ≥ 5 years) because base case assumptions of sequalae following meningitis were informed from limited data.

Finally, we conducted multivariate sensitivity analysis using ranges for model inputs indicated in Tables 1–3. We randomly drew VE parameters from Beta-pert distributions over the indicated ranges. Disease and cost parameters were drawn from a normal distribution using the indicated 95% confidence interval around the sample mean. The 95% confidence interval around the sample mean for costs was calculated by bootstrapping the mean from a sample size equal to the total observations for AOM (n = 1,715,182), tympanostomy tube placement (n = 321), meningitis without IPD (n = 56), and meningitis with IPD (n = 41) over 5,000 repetitions with replacement. We drew QALY parameters from a uniform distribution to reflect the uncertainty surrounding QALY values. Our reported 95% confidence intervals for the main results are from 50,000 model iterations. The impact of the most important inputs identified during the multivariate sensitivity analysis were examined using a tornado diagram. A tornado diagram illustrates the impact range of inputs in the model on the cost per QALY gained and is sorted with the most influential (e.g., widest ranging) impacts at the top. The edges of each bar indicate the cost per QALY gained when the given input was drawn from the 10th or 90th percentile of possible values for that input.