Hello,
I have been using SNPTEST and also BOLT-LMM for some quantitative trait GWAS. Is there any GWAS software out there that can handle ordinal data or possibly a bimodal distribution? I've tried googling / searching, but have not found any answers....
I am now helping a colleague with a GWAS for a quantitative trait derived from ELISA measurements. The trait she has measured has approx 40% of the study population with "non-detectable" levels, and the remaining 60% have a wider range of values (which are important to retain). I'm happy with a rank-based distribution for the detectable ones but don't really know how to incorporate the non-detectable levels (essentially, zeros).
It's important that these subjects are retained and that they have lower values that the detectable ones - in other words, I don't simply want to add "detectable vs not" as a covariate, unless it can be ordered. It might be possible to rank them per quartile/quintile, but I don't know of any software (and am not really confident of the ordinal regression statistics) which can handle an ordinal dataset for association testing. It's a 9.5 million SNP dataset, so computational efficiency is also important.
I'm exploring modelling it as a bimodal distribution (with the non-detectable values modelled as a normal distribution between zero and the lower detection limit of the ELISA), but I don't know of any GWAS software which can handle this for association testing, and other than boot-strapping (which'd be too computationally expensive with our 9.5million SNPs), I don't really know how to model the normal distribution in a population which has no information to rank the subjects (the non-detectable values).
Any pointers massively appreciated!
I am not very expert of this, but maybe you can try with the so-called zero-inflated approaches? Like something discussed in this paper: http://www.scielo.br/scielo.php?script=sci_arttext&pid=S1415-47572011000400008