Table 1.

Determination coefficients and Average Treatment Effect (ATE) variances for Prognostic Covariate Adjustment (ANCOVA) and PPCT methods

	Linear model	Linear mixed-effect model	Disease Course Mapping
			Leaspy univariate	Leaspy multivariate	Leaspy ordinal
Determination coefficient (R2) for the prediction of the PoolSCA data (cross-validation)
With test set simulating ATRIL	-	-	0.133 [0.050–0.217]	0.142 [0.058–0.226]	0.087 [0–0.206]
Determination coefficient (R2) for the prediction of the ATRIL data
With baseline only	-	-	0.002	0.001	0.130
With pre-inclusion + baseline	0.032	0.087	0.145	0.123	-
Variances
$Var(AT{E}_{\mathit{PPCT}})$	0.395	0.401	0.342	0.351	0.355
$(1-R^2)*Var(AT{E}_{\mathit{classic}})$	-	-	0.352	0.349	0.372

For the PoolSCA data, we selected only timepoints similar to the ones in ATRIL and compared the real SARA score progression for one year and the predictions of the model personalized on “PoolSCA test set simulating ATRIL.” A cross-validation process was used to ensure robust evaluation in PoolSCA, as it served as the training dataset, whereas no splitting was necessary for ATRIL dataset, which was used only for predictions. Bold values indicate the best predictive performance (highest R²) among the compared models. For the ATRIL data, we compared the real SARA score progression over one year, and the predictions of the model personalized either on baseline data only or on pre-inclusion data in addition to the baseline. The variances are computed using the sandwich estimator and Eq. (3). As $Var(AT{E}_{\mathit{classic}})=0.407$, the formula from the Eq. (4) linking R² and the variance of the estimator seems to hold empirically. We used the R² from PoolSCA data to verify Eq. (4), as this formula is primarily applied to reduce sample size during the clinical trial design phase, where clinical trial datasets like ATRIL are not available