0

I am new to the field and learned about scaling (by z score transformation) of data. While it seems a super useful and universal technique, I have learned that no technique is applicable to each and every problem.

In what cases would scaling do more harm than good?

Nick Cox
  • 48,377
  • 8
  • 110
  • 156
AKP2002
  • 101
  • 3
  • 2
    Can you clarify what you have in mind? – gung - Reinstate Monica Apr 22 '20 at 05:27
  • @ gung - Reinstate Monica I wanted to know if there are some situations when scaling the variables is not recommended. – AKP2002 Apr 22 '20 at 05:47
  • 4
    The tags you've included refer to very different things. Your question is quite vague. Can you give an example of what you are asking about? – gung - Reinstate Monica Apr 22 '20 at 06:07
  • 2
    I would refute the premise of the question, and argue data should only be scaled when there is a very good reason that it should be (like as inputs for a neural network). Otherwise you are just needlessly modifying and biasing the data. Although if your tag of neural network is to indicate that your question is referring specifically to machine learning (if so I would recommend clarifying the question with that information) then the question becomes less about whether to scale and how to scale. – ajax2112 Apr 22 '20 at 08:42
  • 2
    What exactly do you mean by "scaling"? – Peter Flom Apr 23 '20 at 11:50
  • @Peter Flom by scaling I mean a z-score normalization. – AKP2002 Apr 23 '20 at 15:30
  • @ gung - Reinstate Monica I mean something like mean I have a large dataset with many different quantities such a age, height, weight, income etc. I may be a good idea ( at-least it won't harm) to do a z-score normalization before fitting it in a model, but are there any cases when scaling would create a problem ? – AKP2002 Apr 23 '20 at 15:33
  • 1
    If (value $-$ mean) / SD $=: z$ then by doing that you lose sight of mean and SD and the question is then simply posed, although harder to answer, which is when do you need to know mean and SD? And a simple answer is often. For example, if I tell you that my height is 2 SD above mean, can I go through a doorway 2 m high without ducking or hitting my head? You need mean and SD to answer. Silly example, but the point is much more general. – Nick Cox Apr 24 '20 at 14:35
  • Scaling is used in many ways in statistical science, and not just for linear transformations such as z-scores. So I have edited your title to match your question. – Nick Cox Apr 24 '20 at 14:38
  • I found the answer here ( https://stats.stackexchange.com/a/189655/279588) . Can a mod please close the question as duplicate ? – AKP2002 Apr 25 '20 at 07:08

0 Answers0