Table 3.

Predictive ability and bias of genomic breeding value (GEBV) for BCWD survival DAYS using four GS models with two genotyping platforms.

Genotyping platforma	GS modelb	Training samplec						Validation sampled
		Phenotyped	Genotyped	Effective SNPsf	πg	Fitted SNPsh	$h_M^2$i	Predictive abilityj	Biask
Chip	ssGBLUP	4492	652	40,710	Nal	Na	0.29	0.49	0.68
Chip	wssGBLUPe	4492	652	40,710	Na	Na	0.29	0.40	0.34
Chip	BayesB	583	583	40,744	0.990	407	0.27	0.39	0.55
Chip	BayesC	583	583	40,744	0.995	204	0.26	0.44	0.63
RAD	ssGBLUP	4492	649	10,052	Na	Na	0.33	0.48	0.63
RAD	wssGBLUP	4492	649	10,052	Na	Na	0.33	0.37	0.32
RAD	BayesB	579	579	10,059	0.975	251	0.28	0.47	0.69
RAD	BayesC	579	579	10,059	0.990	101	0.31	0.46	0.61

^aThe effective number of SNPs used was 40,710 and 10,052 from the Chip and RAD genotyping platforms, respectively.

^bGenomic selection models: single-step GBLUP (ssGBLUP); weighted ssGBLUP (wssGBLUP); Bayesian methods BayesB and BayesC; the GS analysis included only progeny of 2005 families.

^cThe training sample included offspring from 10 full-sib 2005 families each with n = 38–80 offspring; BayesB and BayesC models included only fish that had both genotype and phenotype records (n = 579–583). In contrast, the ssGBLUP and wssGBLUP methods included also non-genotyped fish that had disease records (progeny of 2005 families).

^dThe validation sample included 53 breeders (offspring of 2005 families that were included in training sample).

^eFrom wssGBLUP, iteration 2 results are presented; iteration 2 yielded higher accuracy GEBVs than iteration 3.

^fThe analysis included SNPs and samples with a calling rate ≥0.70 and 0.90 for the RAD and Chip genotyping platforms, respectively.

^gMixture parameter p specifies the proportion of loci with null effect.

^hMarkers that are sampled as having non-zero effect (1−π) are fitted simultaneously in the multiple regression model.

ⁱProportion of genetic variance explained by the markers $\left(h_M^{\mathit{\text{2}}}\right)$.

^jThe predictive ability (PA) was calculated using the mean progeny performance (MPP) of 31 progeny testing families that had as parents a pair of validation fish. The PA of GEBV was estimated as the correlation of mid-parent GEBV with MPP from each progeny testing family, PA_GEBV = CORR(Midparent GEBV, MPP).

^kThe bias of GEBV was defined as the regression coefficient of performance MPP on predicted mid-parent GEBV (β_{MPP.Midparent GEBV}).

^lNa indicates either non-available or non-needed data cell.