Table 4. Accuracy of main prediction methods in analyses of type 2 diabetes in a South Asian cohort.

Model	Weights associated to each predictor		Adjusted R2	P-value for improvement over simpler model
	European PRS	Latino PRS
EUR	0.09001		0.01767	<10-3
SAS		0.08488	0.01556	<10-3
SAS+ANC		0.08821	0.01572	0.28
EUR+SAS	0.08309	0.07746	0.03031	<10-2
EUR+SAS+ANC	0.08138	0.07989	0.02968	0.46
EUR-SAS-meta	0.08695	0.00497	0.02098	NA

We report adjusted R² on the liability scale for each of the 5 main prediction methods, as well as EUR-SAS-meta. We obtained similar relative results using NagelkerkeR², R² on the observed scale and AUC (S14 Table). P-values are from likelihood ratio tests comparing models EUR and SAS to the null model, model SAS+ANC to SAS, model EUR+SAS to EUR, and EUR+LAT+ANC to EUR+SAS. For the EUR model we used R_LD²=0.8 and P_T=10^-3, for SAS we used R_LD²=0.8 and P_T=0.8, and for EUR-SAS-metawe used R_LD²=0.8 and P_T=10^-3. We also report normalized weights, defined as the mixing weight ${\widehat{\alpha }}_k$ (see Methods) multiplied by the standard deviation of the PRS.