In some cases, we may want to use principal component analysis on several time-series variables, or some regionalized random variables (i.e., spatial varying variables). The variability of these variables usually contain large-scale trend, which might result in the variables being correlated. Therefore, the correlation obtained from conventional multivariate statistics might be overestimated. I want to know what is the effect of the overestimated correlation for PCA scores? Whether should I detrend the variables before performing PCA?
Asked
Active
Viewed 18 times
0
ttnphns
- 51,648
- 40
- 253
- 462
emberbillow
- 417
- 3
- 10
-
Overestimated relative what? PCA itself is indifferent whether correlations in the input variables are "true" or "twisted". Sure, the magnitude of correlations [affect PCA results](https://stats.stackexchange.com/q/50537/3277). One should consider carefully potentially useful transformations before PCA. – ttnphns Aug 28 '21 at 08:22
-
@ttnphns Thank you for your answer. "Overestimated" here I mean the absolute correlation coefficient is larger than that for the reference variables (without simulated trend term added). I understand you agree with that I should detrend variables before calculating their correlation coeffcient. – emberbillow Aug 28 '21 at 09:00