A COMPARISON OF DIFFERENT APPROACHES IN MODELLING FOR THE CONDUCT OF COST-EFFECTIVENESS ANALYSES: A CASE IN CHILDHOOD FLU VACCINATION
Decision trees have traditionally been the modelling approach used to assess the economic value of vaccinations. However, this approach fails to capture the complexities in transmission dynamics. Alternative approaches that may handle such interactions include agent-based models (ABM) and system dynamics (SD). We compared the performance and results generated by a decision tree, ABM and SD in assessing the cost-effectiveness of two different childhood influenza vaccines.
An existing decision tree comparing intranasal live attenuated vaccine (LAIV) against injectable inactived influenza vaccine was adapted into a SD and ABM structure using the epidemiological pattern of ‘susceptible-infectious-recovered’ on AnyLogic 7.0 . The proportion of infected, expected costs and incremental cost-effectiveness ratio (ICER), as estimated by each modelling approach, were compared. Scenario analyses were conducted to relax the models’ assumptions to determine the impact of the various modelling approaches in assessing the economic value of vaccinations.
Model calibration was successful: all three modelling approaches produced similar estimates when identical parameters and assumptions were adopted. LAIV was found to be the dominant strategy. Scenario analyses revealed that disease transmission and economic value of the vaccination strategies were sensitive to: (1) the proportion and schedule of vaccination under both dynamic models; (2) the network topology, which can be more flexibly modelled in ABM and; (3) heterogeneity from age-specific parameters, which was most easily captured in the ABM.
The clinical and economic estimates differ according to the modelling approach employed and its associated assumptions. ABM, an individual-level model, was the most flexible as it could capture patient heterogeneity and model individuals’ behaviours within their social network. SD, an aggregate-level model, was limited in capturing patient heterogeneity and required an assumption of random-mixing between individuals. The most rigid, though, was found to be the decision tree as it relies on a set of simplifying assumptions.