Fig. 2

Graphical visualisation and validation of treatment heterogeneity defined by non-crossover interactions in the SEAQUAMAT trial. Panels a and b show the univariate relationships to the individual predicted treatment effect for total parasite biomass and base deficit, respectively. The thick blue lines show spline fits to the data. Panel c shows the cumulative distribution of the p values for the added benefit of the ML model obtained by repeated data-splitting and stacking of the standard model alongside the ML model. Significance (at the 5% level) is obtained if the black line crosses above the red boundary. Panel d summarises the overall non-crossover interaction found by the random forest model with a pruned regression tree model fitted to the individual treatment effects. The leaves of the tree in panel d show the mean treatment effect (difference in mortality between artesunate and quinine)