Hi,
I have seen many discussions mentioning PCA. Some say that normality of data is neccessary but some say PCA can be implemented on non-normal data for the extraction of the components. Can anyone please suggest me any literature where PCA has been performed for non-normal data. I have five variables which does not give normality and I want to run PCA on it. I have removed all the multivariate outliers and when i run PCA on it, it gives me components with acceptable reliability.
Anyone having any information regarding this subject please guide me. Sharing of lkteratue will be highly appreciated.
There is no requirement for having normally distributed data when using PCA. There are many applications of PCA, most of which have nothing to do with how your data is distributed. For example, see
https://en.wikipedia.org/wiki/Principal_component_analysis
That reference does list one application (PCA and information theory) where the data is assumed to be normally distributed, but that's an exception.
Hello Chadry,
PCA is a data reduction technique, we reduce many correlated variables to a few a composite variables. We are not using inferential statistics and so no assumption of normal errors in that regard. However, PCA is notoriously sensitive to outliers (not normal, if you like). I see that you have taken outliers into account.
Well done,
David S
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis