systematic upward bias with `SNPRelate::snpgdsFst`

[thierrygosselin]

Hi Xiuwen,

I'm really impressed by the speed of SNPRelate really nice work !

I recently tested SNPRelate::snpgdsFst with method = "W&C84" and systematically get upward bias when tested against other software (e.g. GENODIVE, hierfsat).

I discarded these factors as potential cause for the bias:

• missing value (I work with RADseq data...) because I tested with and without imputation
• common markers between populations
• monomorphic markers

Not tested :

• the way SNPRelate::snpgdsFst handles uneven sampling size across populations

Would really like to get to the bottom of this, because your implementation is really the fastest in R!

Cheers Thierry

[zhengxwen]

I will check my implementation of Fst. Please let me know the sample size in each population.

[thierrygosselin]

In the example above I used these:

POP_ID   N 
G100    10
G102    10
G103    10
G108    10
G109    10
G111    10
G118    10
G122     8

But do get similar bias with more than > 30 samples/pop Thanks for looking into this Thierry

[thierrygosselin]

Any update regarding this issue ?

[zhengxwen]

I have compared my implementation of W&C Fst with the implementations in plink_v1.9 and vcftools.

plink_v1.9 and vcftools return both "Weir and Cockerham mean Fst estimate" and "Weir and Cockerham weighted Fst estimate". SNPRelate returns "Weir and Cockerham weighted Fst estimate", which is the same as plink_v1.9 and vcftools.

It seems that "weighted Fst estimate" is a little higher than "mean Fst estimate".

[thierrygosselin]

Thanks Xiuwen, I'll make some tests on my end with different RADseq data and will update results. Thierry

[zhengxwen]

The latest SNPRelate on GitHub also provides MeanFst, and it will be available in the BioC release (in April). see the help document snpgdsFst().