The effects of transforming skewed data

[Zenstone55]

Hi There,

I would like to know more about transforming data. I know that for parametric statistics data needs to be normally distributed, but I am confused about when this does and does not apply.

Doesn't transforming data to make it normally distributed hide any experimental effects you have?

I have ran an experiment which should have resulted in differences in responding to questionnaire items. My hypothesis predicts a certain pattern of responding (e.g. higher ratings vs lower ratings) on a set of items for participants in different conditions. If I normalise the data to get rid of any skew towards lower or higher ratings, am I just going to hide any experimental effects? E.g. if group one averages a score of 7 out of 10 on an item while group two averages a score of 3/10 and then I normalise it so they both have similar means.

I am very confused about when you would transform data given that it seems counterintuitive!

 

[Karabiner]

I know that for parametric statistics data needs to be normally distributed,

No. That is wrong. (Nearly) no procedure assumes normally distributed
data. Some procedures assume normality within subgroups, or normality
of the residuals of the model, but even these assumptions are unimportant
if a sample size is large enough.

E.g. if group one averages a score of 7 out of 10 on an item while group two averages a score of 3/10 and then I normalise it so they both have similar means.

Seemingly you did not normalize, but you just performed
a z-standardization of your data? A simple z-standardization doesn't
normalize anything. Or, did you do something else?

With kind regards

K.

 

[Zenstone55]

I'm talking hypothetically. I haven't done anything yet.

I mean doing a square root transformation when finding skew and kurtosis in excess of 1.

 

[Dason]

I really dislike these general rules of thumb for transforming data. Do you know why you're doing those transformations? Do you know how to interpret the subsequent analysis given that you've transformed the data? If the answers are "no" then you might want to consider getting more input and help. I know you said you're talking hypothetically but if that's what you're talking about then my reply would just be "don't do it".