0

On the back of an earlier question i am having an issue.

Software = SAS JMP Pro 11

Earlier Question

My rotated factors (post Principal Components Analysis [PCA]) are not at all uncorrelated. I have used Variance Inflation Factor (VIF) scores to initially look at this.

I expected some increase in the VIFs across the board but what i have is pretty dire with VIFS in the 10000s on occasion. The PCA factors unrotated are completely uncorrelated when looking at VIFs (all 1.5 or less). What problems can occur in the rotation step to cause this?

The original variables load onto to many of the unrotated factors to too high a degree to be of much real world use.

I have seen written there is no real difference between orthogonal or oblique rotation and I have also found this to be the case

The only thing i can think is my non positive definitive matrix but I thought (I do not know) that if the unrotated factors are uncorrelated this meant the matrix was handled OK?

Community
  • 1
Samuel
  • 121
  • 10
  • What `earlier question`? Link, please. – ttnphns Mar 03 '14 at 18:05
  • I have linked the previous related question. – Samuel Mar 04 '14 at 09:28
  • I am not sure if this isvalid but i have used a different estimation method to generate the correlation matrix and have now lost the NPD response the software was giving me. I have changed from pairwise (default) to rowwise (misses out rows with missing data). I now no longer get the NPD error, my PCs and rotations of these look slightly different but the topology is the same. Crucially the rotated PCs via variamax no longer are collinear (VIF analysis). Is this valid? – Samuel Mar 04 '14 at 11:36
  • @Samuel: I see nothing wrong with omitting rows with missing data. So the problem is solved? – amoeba Mar 04 '14 at 13:39
  • i think yes i am satisfied in mind (the hardest part) that the problem is solved. I have another question but i think that is best left for another day! – Samuel Mar 04 '14 at 16:10

1 Answers1

1

I have used a different estimation method to generate the correlation matrix and have now lost the Non-Positive Definitive (NPD) error the software was giving me. I have changed from pairwise (default) to row wise (misses out rows with missing data). I now no longer get the NPD error, my Principal components and rotations of these look slightly different but the topology is the same. Crucially the rotated Principal components (via Varimax) no longer are collinear (Variance Inflation Factor [VIF] analysis).

Samuel
  • 121
  • 10
  • Please accept your own answer. Also: maybe you could de-abbreviate NPD and VIF. It could also be useful for the potential google searches if you wrote which software you were using. – amoeba Mar 04 '14 at 16:25