. 2020 May 28;107(1):46–59. doi: 10.1016/j.ajhg.2020.05.004

Table 1.

Comparison of Prediction Accuracy in Simulated Genetic Architecture

				5%	NPS R²_NagCompared to
% Causal SNPs	Method	R²_Nagelkerke	% h²Explained	Tail OR	P+T	LDPred	PRS-CS
1%	P+T	0.050	14.8	3.18
	LDPred	0.068	20.6	3.66
	PRS-CS	0.075	22.0	4.02
	NPS	0.085	24.6	4.27	1.68^∗	1.25^∗	1.13^∗
0.1%	P+T	0.136	40.8	6.32
	LDPred	0.080	23.0	4.08
	PRS-CS	0.156	44.8	7.03
	NPS	0.179	51.2	8.09	1.31^∗	2.22^∗	1.14^∗
0.01%	P+T	0.213	61.4	9.92
	LDPred	0.153 (0.268)^a	43.8 (74.6)^a	7.66 (13.37)^a
	PRS-CS	0.228	65.3	10.35
	NPS	0.328	92.6	17.19	1.54^∗	2.14^∗	1.44^∗

Non-parametric shrinkage (NPS) is more robust and accurate compared to other methods in simulated datasets. The simulations incorporate the dependency of heritability on minor allele frequency and clumping of causal SNPs in known DHS elements. The heritability was 0.5, and the prevalence was 5%. The number of markers was 5,012,500. The GWAS sample size was 100,000. Prediction models were optimized in the training cohort of 2,500 case subjects and 2,500 control subjects. R² of prediction was measured in the validation cohort of 50,000 samples. The h² explained stands for the proportion of heritability on the liability scale explained by polygenic scores. The asterisk (^∗) indicates a significant improvement in Nagelkerke’s R² (paired t test; p < 0.05).

The accuracy of LDPred varies widely depending on the convergence of prediction model; thus, we report the maximum R² in parentheses as well as the average performance.