Advantages and disadvantages of data transformation and rank-based tests

Dear Talk Stats Community,

I would like to learn about advantages and disadvantages of transforming non-normally distributed data to achieve normal distribution versus using ranks and subsequent non-parametric tests. With assigning ranks to individual values, we lose some information. With transformation, we change the original distribution type. I need to know which of the two approaches is likely to return more accurate inferences about samples?

Would the following be correct?

If data can be transformed to a normal distribution, use parametric tests on transformed data.
If data cannot be transformed to a normal distribution, use non-parametric tests on raw data.

Thank you.



TS Contributor
Why do you think you need "normally distributed data" ?
Except for the statistiscal significance test of the Pearson
coefficient, no statistical analysis requires normally distributed
variables (at least none comes into my mind).

With kind regards

Dear Karabiner,
Thank you very much for the prompt reply. Let me rephrase my question. Does data transformation have application areas besides linearizing relationships? I use the attached paper as a basis for the assumption that transforming e.g. lognormally distributed data to normal distribution is more appropriate at least in situations when the lower limit of CI falls below zero for such variables as mass, size, concentration etc.
Thank you,