Depending on the trait this does seem to work, but sometimes it only works if I drop the smallest chromosomes from the analysis. Then I have an answer for some chromosomes, but not all of them.
What I would really like to do is the equivalent (I think) of a mlma-loco, where I use all of the chromosomes except one to get a grm while I estimate how much of the variance is explained by the focal chromosome. I'm not as keen on the GWAS that pops out of the mlma-loco, I just want a chromosome specific heritability estimate.
If your purpose is to estimate h2SNP for a particular chromosome, you can fit a model y = c + g_minus_c + e where c is the genetic effect for the focal chromosome and g_minus_c is the genetic effect for all chromosomes except the focal chromosome. For example, the focal chr is chr1, then the multi_grm.txt should be chr1_grm all_but_chr1_grm
all_but_chr1_grm can be computed by merging the GRMs of the other 21 chromosomes.
I also interested in doing this LOCO approach for partitioning variance across chromosomes. For each chromosome, however, the variance explained is very small (e.g. 0.001-0.01) and does not come close to adding up to the total heritability of the trait. I realize this could be due to covariance among chromosomes so I have filtered for LD 0.8 before calculating the "all_but_chr*_grm"'s. Are there other explanations for why the variance explained by each chromosome is so low? Would relatedness among individuals cause this?