Epidemiological model design
We adapted a dynamic transmission model used to assess the impact of QIV in the US, first published by Crépey et al (15). The model is a compartmental model able to simulate infections by two A subtypes or two B lineages at the same time. Individuals can be susceptible to infection (S), exposed but not infectious (E), infectious (I), or recovered (R) from an infection and therefore immune. In addition, individuals can be vaccinated (V) against both B influenza lineages, and either one of the two. The model accounts for cross-immunity against a B lineage induced by vaccine containing the opposite B lineage or induced by natural infection. The population in the model is divided into 8 age groups (0-5mo, 6mo-5yo, 6yo-9yo, 10yo-14yo, 15-19yo, 20-39yo, 40-59yo, and older than 60yo). In addition, the model simulates several epidemic seasons in a row in order to take into account the evolution in time of the immune status of the population. The model was developed in R 3.5.3 (16) and C++ (17).
Economic model design
Our economic model is similar in structure to the one published by de Boer et al (18). It is a decision tree-based model where symptomatic individuals infected with influenza will have various probabilities of having an outpatient visit, being hospitalized, or dying from influenza, depending on their age and whether they are at-risk of severe consequences. The economic model computes health outcomes (outpatient visits, hospitalizations, and deaths), health effects (life year lost, quality adjusted life year lost), medical costs, vaccination costs and indirect costs (productivity losses). Age-stratified outputs of the epidemiological model are used as inputs of the economic model, developed in Excel© 2010.
The proportion of influenza A and B circulating on the period 2010–2017 was extracted from the Brazilian SINAN notification system (19). The split for A/H1N1 and A/H3N2 was obtained for Brazil either from WHO Flunet (20)(2010–2012) or from SINAN (2013–2017), and the split of B lineages in Brazil from Luna et al. (21). We obtained hospitalization data over the same period from the Brazilian public health care system (Sistema Único de Saúde, SUS) (22). As the SUS only accounts for public hospitals, and since approximately 25% of the Brazilian population have access to private hospitals, we extrapolated the number of hospitalizations obtained from the public sector proportionally to estimate the total burden of influenza hospitalizations, assuming that the same incidence is observed in both public and private hospitals. We obtained the proportion of the Brazilian population covered by the private system over the period of analysis from the National Health Agency (ANS) (23).
Following the same approach used by the US CDC for influenza burden estimation (24), we divided the number of hospitalizations by the test sensitivity and by the percentage of tested subjects to account for non-tested and false negative subjects. Next, we multiplied this estimation by a case-hospitalization ratio to obtain a first estimation of the total number of influenza cases seen by the healthcare system (25). Finally, we divided this estimation by the probability to seek for healthcare in order to obtain the total number of influenza cases per year in Brazil. Influenza incidences from 2010 to 2017 obtained thanks to this methodology are shown in Fig. 1. As our epidemiological model requires weekly incidence data, we extracted the weekly number of influenza positive samples reported to WHO FluNet in Brazil over the studied period (20). We synchronized the peaks occurring during each season (average lag of 5 weeks) and then computed the average number of cases for each week to obtain an estimated epidemic profile representative of a typical influenza season in Brazil (Figure S1). We then applied this “typical” epidemic profile to the yearly incidence per age-group previously estimated to obtain weekly incidence from yearly incidence.
Vaccination rates for the period 2010–2017 (Table S2) were obtained from the Information System of the Brazilian National Immunization program (26) and were used for the model calibration process. For the analysis, we applied the coverage rates observed in 2017 for all age-groups as it is more likely to correspond to current and future coverage rates in the country. We varied the coverage rate observed in the 6 m-5yo age group from 50–100% in the sensitivity analysis.
Epidemiological model calibration
Probabilities of infection were estimated for each influenza A subtypes and B lineages and estimated two by two (the two influenza A and the two influenza B). Estimations were performed sequentially for each year, on weekly influenza incidence, following the method described in Crépey et al.(15). This method ensures that the level of immunity in the population for a given year depends on the influenza epidemic dynamics observed the previous years. We improved the calibration process developed in Crépey et al.(15) by estimating separately age-based susceptibility levels allowing to reproduce more accurately influenza incidence by age-groups. As influenza epidemics in Brazil do not start at the same time depending on the latitude (27), epidemic curve shapes at the national level are difficult to fit with a model simulating a single epidemic on an single population. Consequently, in addition to matching the weekly incidence, the model was forced to replicate the yearly incidence in order to ensure that the model outcomes were consistent with the number of influenza cases observed in Brazil.
Health outcomes data
All health outcomes data are detailed in supplementary material (Table S3 & S4). Probabilities of outpatient visits in case of influenza are taken from Prosser et al. (28) and Molinari et al. (29). Although we obtained the number of hospitalizations for influenza in Brazil, the number of influenza cases in Brazil is not directly available and we could not document the probabilities of hospitalization per case (case/hospitalization ratio) in the Brazilian context, hence probabilities of hospitalization in case of influenza are taken from Reed et al. (30) in the US context. Probabilities of death in case of symptomatic influenza for each age-groups were extracted from published CDC estimates and averaged over the seasons 2012–2013 to 2016–2017(31). Influenza test sensitivity and probabilities of being tested were also taken from a study based in the US (32). We used the estimated population size by age group and the life expectancy estimates over the period for Brazil from the Geographic and Statistic Brazilian Institute (33). Due to the lack of utility estimates specific to Brazil, we used data from the US for quality adjusted life years lost and utility loss due to influenza and its consequences(28, 34).
All costs used in the model are detailed in supplementary material (Table S5).
Medical costs and indirect costs
Outpatient cost and medical cost of deaths are taken from SIGTAP(35), while hospitalization costs were provided by DATASUS (36) and averaged over 2010–2017. Treatment cost consider only the public cost of antiviral treatment in Brazil as we did not consider over-the-counter medication for simplification reasons, and since in many cases these represent out-of-pocket expenses. Private costs are detailed in the supplementary material. Productivity loss were estimated based on daily wages in Brazil (33) and inflated to the year 2017. The number of workdays lost due to influenza estimated according to Molinari et al. (29). Productivity loss due to mortality are estimated by computing the loss in earnings for the life years lost.
For the cost of a dose of TIV (0.5 ml), we consider the price published by the Brazilian government of R$15,14 (37). For the QIV price, we considered R$33,89 (0,5 mL) which is the maximum manufacturer price without taxes published by Brazilian Medicines Market Regulation Chamber (CMED) (38). We considered that the cost of a pediatric dose (0,25 ml) was half the cost of the adult dose. We did not account for administration cost in the analysis, as it would not make a difference since TIV and QIV are assumed to have the same coverage rate.
Vaccine efficacy data
We considered vaccine efficacy per age and per influenza A subtypes (A/H1N1, A/H3N2) and B lineages (B Victoria, B Yamagata) as described in Crépey et al. (15) and shown in Table S6. Regarding cross-immunity between B lineages, we considered that a mismatched vaccine conferred 70% of the matched efficacy (39). This cross-immunity estimate was varied in a dedicated sensitivity analysis.
A post-pandemic retrospective time horizon of 8 years from January 2010 to December 2017 was used in the calibration process in order to account for fluctuations in influenza incidence, influenza B circulation, and vaccine mismatch between seasons. However, we decided to not consider the year 2010 in our vaccine impact analysis to avoid the risk of biasing our results with the immediate aftermath of the 2009 pandemic. A high influenza A monovalent vaccination in 2009 may have triggered a proportionally higher influenza B circulation the following year (Table S1), which would have artificially favored the QIV strategy. To assess the vaccine impact, we used the 2011–2017 period and presented the averaged results of the 7-year period. The pandemic year 2009 was not considered as well for not being representative of the current epidemiological context. Incremental cost effectiveness ratios (ICER) were computed by dividing the net incremental cost of the strategy, compared to the baseline, by the net difference in QALYs or LYs. We considered one time the gross domestic product per capita (GDP) as a threshold for a “highly cost-effective” strategy (R$32.747)(40) and three times the GDP (R$98.241) for a cost-effective strategy. The public payer perspective was considered (SUS) but we also presented the societal perspective (including the private direct costs). According to Brazilian economic evaluations guidelines (41), all costs and health outcomes were discounted at a rate of 5%.
We performed sensitivity analysis on vaccination coverage of the pediatric population, cross-protection between B-lineages (from no cross-immunity to 90% of the matched efficacy) and on influenza B circulation (from 20–40%). We also provide a deterministic sensitivity analysis on probabilities and costs of health outcomes in order to identify the main drivers of our results. Finally, to account for uncertainty in probabilities of outcomes and costs, and assess the robustness of our results, we performed a probabilistic sensitivity analysis and provide a cost-effectiveness acceptability curve where 1.000 simulations with different combinations of parameters are displayed.