Misc. Stats Topics Discussion > Transformation Query
Hi Rich!
Sorry for the delay in my reply, but I've been on my honeymoon for the past few weeks. Your question is a great one, as dealing with variable transformations can be complicated and feel a bit ambiguous. In short, my recommendation would be that you transform only the variables that violate the normality assumptions. Another option would be to avoid the ambiguity of transformations all together by using non-parametric analyses that are robust to those violations. I hope that helps!
- Jeremy (Stats Make Me Cry Guy)
June 2, 2011 |
Jeremy Taylor
Jeremy Taylor
Hi,
I am currently analysing a dataset with 26 variables and intend to perform linear regression analysis on the data and as such am preparing the data to meet parametric assumptions.
I have identified 4 possibly 5 variables that are uncomfortably skewed and log (10) transforming these variables makes them rather more normal. My question is that because I have used log transformation on some of the variables in the data set (including one of the outcome variables) should I transform all the other variables in the same way. If I do this however (log transform all) several variables that were normally distributed appear much more negatively skewed. I have tried a square root transformation and this seems to lessen the negative skew problem but I wonder if I need to transform all of them at all.
I would really appreciate any help,
Kind regards,
Rich