Before doing any GWAS (genome-wide association study) it is necessary to check for the normality of the phenotypic distribution. If the phenotype is normally distributed only once it is log-transformed, what phenotypic data do I have to use while doing the GWAS? The non-transformed one or the log-transformed?
The correct answer can only be given when one takes a look at the distribution of your continuous phenotypes. Obviously one cannot just go around logging everything without real justification. Things like weight, height, etc., can be logged if it helps to bring them to the normal distribution, but you should be checking each. If logging 'improves' the distribution, then include the logged values in the model. We all know that a good analyst will check the logged and un-logged in any case, and pray that they agree..
The type of phenotypes about which you need to feel some worry are ones that are heavily skewed. For example, if we have measures of weight and 99% of the cohort is under 75kg with the other 1% being >150kg, then that's some level of skew that would have to be managed by, possibly, excluding the 1%.
Note that you can di- or tri-chotomise your continuous phenotypes, although this may leave residual confounding factors or leave out useful information pertaining to the phenotype that could otherwise have been useful, depending on the context (ref: The cost of dichotomising continuous variables).
Added 22nd September 2018:
More information: A: Why quantitative design are preferred GWAS approach