Maxime Verhoeven

145 Cost-effectiveness of treat-to-target strategies over 5 years SUPPLEMENTARY FILE Supplementary Data S1 QALYs Effectiveness was expressed as QALYs, measured at baseline, 3, 6, 12 and 24 months using the EuroQol 5-Dimensions 5-Levels (EQ-5D5L) with results expressed as a utility score. Utility is a score ranging from 0 (death) to 1 (full health), and was based on the Dutch tariff. 10 For the post-trial period, during which EQ-5D5L was not measured, and in case of missing EQ-5D5L data, utility was estimated from HAQ and age, using the following formula; EQ-5D5L= 0.82 -0.179*HAQ -0.019*HAQ² +0.002*age. This formula is based on a prediction model reported previously, 1 which was re-estimated in our own dataset using linear regression, to obtain optimal EQ-5D5L estimations for our study. QALYs were calculated as the area under the utility curve using EQ-5D5L measurements (year 1 and 2) or the estimated EQ-5D5L (year 3, 4 and 5) with linear interpolation. 1 = Pennington B, Davis S. Mapping from the Health Assessment Questionnaire to the EQ-5D : The Impact of Different Algorithms on Cost-Effectiveness Results. Value Health. 2014;17:762–71. Supplementary Data S2 Missing data and data imputation Costs (per category, over the last 3 months) were collected at baseline, 6, 12, 24, and 60 months, and in addition utility (and medication costs) at 3, 18, 36, 48 months. To obtain yearly costs andQALYs, linear interpolationwas used over scheduled 3monthly visits within a year. For the costs estimates over the 3rd and 4th year, interpolation between month 24 and 60 (and 36 and 48 for medication costs and utility) was used. Remaining missing information for costs and QALYs per year was considerable. During the trial phase, 12% (n=38) and 13% (n=41) of patients hadmissing information in at least one year for costs and QALYs, respectively, and during the PTFU period, on average 22% (n=50) and 20% (n=45) of patients/of yearly estimates, respectively. As these missing might not be completely ‘missing at random’, we imputed these values usingmultivariable imputation from chained equations. To account for missing cost and QALY values in a year, as well as (population) uncertainty in outcomes, we used the following approach, which has been suggested to be optimal in this situation. 1 As first step, 10,000 bootstrap samples (with replacement) were taken per treatment strategy arm. In the second step, single imputation of the missing yearly QALY and cost estimates was performed per bootstrap sample. The variables DAS28, HAQ, utility, age, gender, rheumatoid factor status (RF), anti-cyclic citrullinated peptide status (anti-ccp), seropositivity for RF and/or anti-ccp, all at baseline, DAS28 and HAQ at year 1 and at year 2, and the estimates of QALYs and costs (per category) over all years (except for the specific variable being imputed) were used as predictors. This resulted in 7