Often I see papers where they use "inverse normal transformation" to transform a not-normally distributed trait. I understand why you'd want to transform. But what is the advantage of using "inverse normal transformation" compared to a simple log-transformation?
Transformations raise at least as many issues as they solve. for example, For example, if you log-tranform your data, you are modelling the geometric mean, not the arithmetic mean, and you are going to have hard time explaining this to readers. You are better off, generally, using a generalised linear model with a log link,
It will depend on your data. If right-skewed you would want to try square root, logarithmic transformation etc. and if left-skewed cube or square transformation etc. This is just a guidance, and as mentioned, the key is to find the one that works best for you data if there's one at all. Whether you should transform at all is another question - for large enough sample you might be better off leaving your values untransformed and use robust test. But it does seems like the use of INTs gained popularity in genetic studies at least, and I found this paper (PMID: 19526352) looking at the benefit of using it - not sure if it's relevant to you.
What transformation to use depends on the type of data you have. Inverse normal, arcsine or logit transformations are often used to deal with variables that are limited to the interval from zero to one (e.g. hit and false alarm rates in psychology, etc.). Using those transformations, the interval [0,1] is transformed to +/-infinity.
Signal detection theory is a common example where an inverse normal transformation is used.
- Badges
- Science method
- Similar topics
- Biological Science
- Ecology