If fluoxetine is effective, is it also cost-effective? Each trial has taken slightly different approaches to the estimating the cost-effectiveness of the interventions. These have been tailored to their national circumstances. We envisage performing an individual patient data meta-analysis to derive the best estimate of overall cost-effectiveness and that in subgroups of patients. At this stage when it is unclear which data will be available in which trials, we cannot be too specific about the statistical methods to be used.
We plan to carry out within-trial economic analysis of direct resource costs and health outcomes on an intention-to-treat basis. A health service perspective will be adopted for measuring and valuing health service use over a 12-month time horizon, although some data to reflect indirect costs will be available in the EFFECTS trial (Sweden).
We will estimate cumulative costs of inpatient episodes, hospital outpatient visits, care home stays and home care visits. Unit costs will be obtained from each participating country (Australia, New Zealand and Vietnam, Sweden and the United Kingdom) and applied to patient-level data collected within each international setting.
The number and duration of hospital episodes and other secondary care contacts will be recorded using linkage to routine data or information obtained from the case report form. Resource use will be valued using unit costs taken from national datasets and published sources (e.g. Australian National Hospital Cost Data Collection, UK Unit Costs of Health and Social Care, Swedish National Cost per Patient register). Unit prices will be applied to resource use data using 2018 as the reference year with country-specific health-sector price deflators used to adjust costs reported in other years. No discounting of direct resource costs will be conducted as the time horizon will be limited to 12 months for within-trial analysis. Output-based hospital-specific purchasing power parities (PPPs) will be used for conversion of expenditure estimates using a common (currency) unit.
Self-reported health-related quality of life (HRQoL) at six and 12 months of follow-up will be measured using the EQ-5D-5 L preference based scale. EQ-5D-5 L single index values will be calculated using English value sets updated if value sets are reported for other countries where the family of trials are being conducted. We also plan to validate the EQ-5D-5 L by checking the concordance with the mRS.
A standard multiplicative model will be used to estimate QALYs calculated by the area under linear interpolation of the EQ-5D-5 L index trajectory for each individual with survival times, the EQ-5D-5 L utility index score at six and 12 months, and a modelled baseline EQ-5D-5 L utility index value. Multiple imputation with chained equations will be used to impute missing HRQoL data on the EQ-5D-5 L assuming missing at random [
25]. We will perform sensitivity analysis based on cases with complete data follow-up and examine whether the pattern of missingness is informative with respect to key individual characteristics.
The primary treatment effect in the economic analysis will be estimated using an individual level regression model for average (mean) incremental costs and incremental QALY times over 12 months after randomisation. The model will consider the joint distribution of costs and QALYs using a general specification that will allow for different parametric and conditional distributions. We also plan to use a Bayesian model with minimally informative priors for means and large variances [
26]. Model parameter uncertainty will be addressed using probabilistic sensitivity analysis summarised using the cost-effectiveness acceptability curve. We will also conduct a companion analysis of cost-effectiveness where we will truncate the cumulative cost distribution at six months and estimate the incremental costs in relation to incremental differences in the primary outcome measure (mRS at six months).
Secondary analyses will be conducted to address heterogeneous treatment effects. Subpopulations with different average treatment effects will be identified using ‘regression tree’ or ‘recursive partitioning’ methods [
27]. These data-driven analyses will complement pre-specified subgroup analyses examining individual and group covariates of substantive interest such as stroke severity (NIHSS) and the SSV model for prognosis [
4].
Longer run modelling will estimate the distribution of costs and QALYs calculated over the expected patient lifetimes. A microsimulation model will be calibrated using information gained from the within trial analysis of cost-effectiveness combined with additional data from: (1) trials and observational studies reporting longer run costs, survival and HRQoL following stroke; and (2) expert beliefs on the distributions of parameters where information is less readily available. The structural uncertainty in the long run model will be addressed using model averaging methods.
The design of the economic analysis will contribute to a structured overview of treatment effects taking advantage of the common trial protocols and consistent capture of resource use and outcomes in FOCUS, AFFINITY and EFFECTS. Generalisability and, in particular, the assessment of treatment effect heterogeneity will be enhanced by the pooled data across these trials.