Fig. 2

Accuracy of individual long-term forecasting. A, B Examples of individual predictions for different amounts of incoming data with population priors for four participants, one per dose. D2: 0 h. D11: 15 h. D25: 30 h. D33: 60 h, each in four scenarios in which the availability of outcome measures increases for each participant (circles). Note how when MAL data following the 2nd bout of training is available, predictions at 6 months become remarkably accurate. B Same as in A, but here, we do not use population hyper-priors but weakly-informative priors for each participant. Solid lines: model prediction (posterior median). Blue shaded zones: 90% and 95% prediction intervals (from darker to lighter). Triangles: MAL data measured but not used to update the model. Note how, when compared with hierarchical model in A, the forecast at 6 month was largely inaccurate and imprecise (i.e., large prediction interval) when only the baseline data were included. C Comparison of mean BF-RMSEs for each week in the different forecasting scenarios. Dots and Triangles are the mean BF-RMSEs for the hierarchical and random effect (i.e., non-hierarchical) models, respectively. Solid and dotted lines are the best fit lines (log-linear) of the weekly mean RMSEs for the hierarchical and random effect models, respectively. D One-sided p-values for the difference in mean RMSEs between the hierarchical and the random-effect models for each forecasting scenario. Note that whereas the hierarchical model improves short- and long-term forecast until the 2nd bout of training, after the 3rd bout of training, predictions with and without the hierarchy become nearly indistinguishable