# How to deal with correlated SNPs in GWAS?

I have population SNP data in VCF format, and I found that some SNPs have a great similarity across samples(> 99%). For example:

CHROM_POS   s0  s1  s2  s3  s4  s5  s6  s7  s8  s9
chr1_1  A   G   G   A   G   G   A   A   G   A
chr1_2  A   G   G   A   G   G   A   A   G   T
chr1_3  C   C   C   A   C   A   C   C   C   C


The similarity between chr1_1 and chr1_2 is 0.9, because they have a different SNP in sample s9.

Is there a good way to remove these similar SNPs before I put the data into another pipeline?

Added: These SNPs will be used for GWAS analysis.

• Do you mean that some SNPs have a MAF of <1%? – Emily_Ensembl Sep 5 '17 at 8:11
• @Emily_Ensembl MAF is about allele frequency. But I want to remove similar SNP sequence across sampls . – l0o0 Sep 5 '17 at 8:16
• The question is about if removing common SNPs a good thing. – HelloWorld Sep 5 '17 at 8:28
• @SmallChess, I think similar snps may will provide duplicate information and cost more resource to compute. Maybe similar snp in long distance should be kept? – l0o0 Sep 5 '17 at 8:34
• It wasn't clear from your original question what you meant. Now you've edited it I can see that you're looking at haplotypic blocks with SNPs in LD. No, don't remove them. There may be causal SNPs within these blocks – you won't able to tell which SNP is causal but you can identify that the block is important. – Emily_Ensembl Sep 5 '17 at 8:56

As @Emily_Ensembl said, it is not customary to do this for standard association tests: it is possible that one of the SNPs you remove is causal, or a better proxy for the causal locus, and would give (slightly) better association signal than the other. Even for SNPs in perfect LD, pruning is unwise because it would complicate interpretation: if SNP chr1_2 had some obvious function, but you removed it and are left only with its proxy chr1_1 (which could be far away), you will have a hard time connecting the dots. Typical GWAS tests are handled pretty quickly on modern computers, and are linear in SNPs, so there is no computational reason to prune either.