----------------------------
IMPUTATION------------------
----------------------------

Here we provide sketch details for some of the other analyses performed in the main text, as well as some additional features of LDAK. If including lower-quality SNPs, then simply add --infos when computing kinships. The genotype scaling can be varied using the option --power. For example, to use the previous default scaling, use --power -1.

After performing REML, the .share file provides relative estimates. These are useful when interested in relative contributions (e.g., of different SNP classes). The .reml file contains the null and alternative (log) likelihoods and a likelihood ratio test (LRT) statistic (the null model corresponds to only covariates). To test significance, we typically computed the difference in LRT statistics between results from partitioned and non-partitioned model (e.g., when comparing the GCTA and LDAK Models, we performed REML using just the genome-wide kinship matrix, then using two kinships, one computed from low-LD SNPs, the other from high-LD SNPs.

To test the contribution of DNaseI hypersensitivity sites (DHS), we downloaded DHS annotations from \url{hgdownload.cse.ucsc.edu/goldenPath/hg19/encodeDCC/wgEncodeRegDnaseClustered/wgEncodeRegDnaseClusteredV3.bed.gz}. From these we created dhs.txt which had four columns, providing a unique identifier, chromosome, start and end basepairs for each. Note that if two-way SNP partitioning results in a very uneven divide of SNPs, instead of computing both kinships from scratch, it can be much quicker to first compute kinships for the small tranche, then subtract these from the (previously-computed) genome-wide kinships using the option --sub-grm. We did this when testing DHS, as these contain less than 20% of the genome. A similar strategy can be used when considering more than two partitions.

------------------------------------------------------------------------------

#To calculate enrichment of DHS, first compute kinships from DHS SNPs, then its complement (i.e., subtract from genome-wide kinships), then perform two-way REML
ldak --cut-genes dhs --speed data --genefile dhs.txt --ignore-weights YES
ldak --calc-kins-direct dhs --speed data --extract dhs/gene_preds.txt --weights sections/weights.all --power -0.25
echo -e "kinships/kinshipALL\ndhs" > listsub
ldak --sub-grm not_dhs --mgrm listsub
echo -e "dhs\nnot_dhs" > listdhs
ldak --reml dhs --mgrm listdhs --pheno phen.pheno --covar covar.covar --keep prune.keep \
--top-preds top.in --speed data

#If including rare variants, then partition based on MAF
#Exact boundaries are not too important, but consider extra boundaries (say at 0.00025) if using very rare SNPs
#Likely some of the rare variants will be low quality, so add '--infos'
awk < data.stats '(NR>1 && $3>0.1){print $1}' > maf1
awk < data.stats '(NR>1 && $3>0.01 && $3<=0.1){print $1}' > maf2
awk < data.stats '(NR>1 && $3>0.025 && $3<=0.01){print $1}' > maf3
awk < data.stats '(NR>1 && $3>0.001 && $3<=0.025){print $1}' > maf4
awk < data.stats '(NR>1 && $3<=0.001){print $1}' > maf5
ldak --cut-kins rare --partition-number 5 --partition-prefix maf
for j in {1..5}; do
ldak --calc-kins rare --speed data --partition $j --weights sections/weightsALL \
--power -0.25 --infos data.infos
done

------------------------------------------------------------------------------