I run a GWAS for 660k SNPs from an Affymetrix commercial array in a pig population. Then, I designed a custom SNP array with 27 SNPs, 17 that are also present in that Axiom chip and 10 from other sources to genotype other populations. I first genotyped the same pig population with this custom array and to my surprise, for the SNPs that are present in both arrays the results are different (showing less significant P-val in the custom array).
I always understood that in a GWAS, the SNP p-val are nominal and need to be corrected for multiple testing (e.g. FDR or Bonferroni). I also understood that the nominal p-val does not depend on the number of SNPs that have been analyzed in the GWAS. However, I am not sure if this assumption still holds for the MLMA with the REML and the GRM.