Fig. 2

A graphical comparison of four approaches to reference class forecasting of ITEs (thick blue lines with the pointwise 95% confidence intervals [CIs] shown as shaded blue areas) for patients enrolled in the AQUAMAT study [18]. In each panel the ATE (95% CI) from the original trial (n=5483) is shown by the dashed red line (red shaded area). The left column shows fixed bandwidth predictors (fixed effective sample sizes approximately equal to one fifth of the original sample size), and the right column shows varying bandwidth predictors (varying effective sample sizes). a Risk-based quintile partitioning. This does not interpolate between average risks in each subgroup. There is some minor variation in effective sample size due to ties in the multivariable risk scores. b Exponential tilting with free parameter λ as a global reweighting scheme with varying effective sample size (top x-axis). This is centred around the overall treatment effect corresponding to the value λ=0. c1 Epanechnikov kernel with fixed bandwidth chosen for an effective sample size of n=1097 (20% of the original sample size). c2 Epanechnikov kernel with maximal bandwidth reference class. Note that the 50% risk quantile has an effective sample size reduction of 17% with respect to the original trial sample size due to the decay in weights. This contrasts with panel b, where the ITE prediction at the 50% empirical risk quantile equals that of the ATE