Sometimes it is not useful at all, specially in very skewed distributions, with many values=0. We don´t get the normality, and the problem goes on.
Moreover, it seems "going downstairs" to use a non-parametric test (eg, Rho-Spearman) instead of the parametric tests (eg, r-Pearson), more widely used.
Transforming Variables for Normality and Linearity –
When, How, Why and Why Not's:
http://support.sas.com/resources/papers/proceedings12/430-2012.pdf
Do not log-transform count data:
http://onlinelibrary.wiley.com/doi/10.1111/j.2041-210X.2010.00021.x/pdf
Thank you, Amir.
I have pointed out the question because the log-transformation is not a guarantee to solve the problems of deviations from normality. But it quite easy to develop and that is probably the cause of its quick and widespread use.
On the other hand, the non-parametrics methods don´t need normality; besides, they are robust from outliers or extreme values.
In my opinion, I think the data must be first analysed untransformed and if there is evidence of a departure from assumptions, then, do the choice (and to explain it) among a variety of transformations or non-parametrics procedures.
Best regards.
A part the statistical issues, documented by Box Cox, the log transformation suggest some interesing stuff in economics :
In a regression context, the log-transformation offer a interesting interpretation in term of elasticities.
Specially in time series, prices for example, first difference + log lead to consider returns or growth rates.
Best Regards
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis