I don't have enough karma to upvote or comment on this question but I have the same question as this old unanswered one. stats.stackexchange.com/q/363623/310171
The low engagement with the question makes me think people don't understand why anybody would want to know the answer, so I'll try to explain.
1) I think it's good practice to simulate data and ensure that my model recovers the assumed parameter values. For example, when increasing x1 from its mean to 1.0, the outcome should increase by 0.4 standard deviations.
2) In the real data I am standardizing (zscore transforming) the outcome column, so I should probably do the same transformation to my simulated outcome column.
3) When I do that transformation, I no longer recover the assumed parameter values, because they were generated in terms of what's now raw unstandardized outcome values.
To solve that problem I can not standardize the outcome in the simulated data, but then my simulated data doesn't look like my real data. Maybe that's not a problem but naively it seems bad.
@Sycorax Thanks! By the way I'm quite reluctant to ask anything on the Artificial Intelligence SE, including their chat rooms. A question just got asked on MMSE about whether or not "features" should be normalized. The term "features" seems to be a machine learning term. Are there people here who are familiar with features and whether or not they should be normalized? This was the question on MMSE:
In regression models to predict chemical compounds' activities, fingerprints are often used as features. Should one normalize a fingerprint feature to be in the 0-1 range?
@NikeDattani: A feature's just a predictor, a.k.a. independent variable, in the regression model. The question could do with some explanation of (1) what a fingerprint is & how it's represented in the model, & (2) how the model's being fit - ordinary least squares or maximum-likelihood estimation procedures are scale-equivariant.
@NikeDattani It's unfortunate that you don't feel comfortable using AI.SE. Looking at some of the starred messages in the AI.SE chat, I can see that you're not alone in feeling that way.
2
Luckily, perfectly fine to ask questions about statistics, data analysis, machine learning and some (but not all) aspects of artificial intelligence on this website. In almost all cases, it's best to ask such questions using the Question/Answer format, but this particular question seems distinguished because it can be completely answered in a few sentences.
@Sycorax, I just checked their chat room to see what that was about. It's kind of shocking. It was also interesting to see who is at the center of it all & what the offending behavior was...
In regression models to predict chemical compounds' activities, fingerprints are often used as features. Should one normalize a fingerprint feature to be in the 0-1 range?
