The best powered statistical test is the test that most closely mimics reality (in genetics or in anything else).
Therefore, if a disease is actually inherited in a dominant fashion, then use a dominant genetic model.
If the condition is actually inherited in a recessive fashion, use a recessive model.
You are probably most used to seeing additive tests because many of the diseases for which we are doing GWAS are complex conditions. Because they do not follow AD or AR inheritance, it is unclear what the best powered test could be.
The liability threshold model is, generally, the idea that people get complex diseases due to "stacking" many risk alleles. Though not strictly proven, this theory has been influential and guides many scientists' thinking. In this context, additive seems like a reasonable choice.
I know my explanation is brief. If you need to read more, start at https://en.wikipedia.org/wiki/Threshold_model, and keep reading.
**Genotype Allelic Dosage
Allele 1 Allele 2 Additive Dominant Recessive
A A 0 0 0
A B 0.5 1 0
B B 1 1 1
**Note: assumes diploid.
Thank you Vincent! So suppose the disease is actually inherited in a recessive fashion, when I try to calculate the PCA for a group of people, should I recode the genotype based on the recessive model (0/1) instead of the additive model(0/1/2)?
the PCA calculations shouldn't have anything to do with the genetic inheritance model selected... you may wish to run QC that looks at the two together, but conceptually they ought to be unrelated apart from QC issues...
the PCA is used to control for confounds based on relatedness. see Price AL et al. 2005. the inheritance model is used to code the case control status...