Re: [Gretl-devel] pca bug and issues reloaded

On Sun, 14 Dec 2014, Sven Schreiber wrote:

> > Summary: The 'princomp()' function seems to work fine, but 'pca' > apparently uses the full sample for calculating the pca, even if a > reduced sample is specified. OK, I think I've finally worked out what the issue is here. It's complicated by the fact that the eigenvectors of a 2 x 2 correlation matric are invariant with respect to the (single) correlation coefficient. For a while I thought we were in error in reporting the same eigenvectors for the full dataset and the sub-sample, but that's expected. The real problem was that when we standardized the series for computing the saved PC series we used the means and standard deviations for the full dataset, regardless of the current sample. That's now fixed in CVS.

> - Accessor for the loadings, as suggested by Henrique > (http://lists.wfu.edu/pipermail/gretl-users/2012-March/007346.html) and > in terms of the princomp() function by myself. Allin answered that it's > easy to get them as the eigenvectors of the correlation matrix. This is > of course correct, but first it's a convenience issue, and secondly if > you perhaps want to do some simulations it seems like an avoidable > inefficiency to compute the eigenvectors twice (first implicitly in > princomp, and then explicitly by hand). Would be nice but IMO not a high priority. The efficiency point doesn't strike me as very important unless perhaps one were computing PCs for a huge array of series.

> - Automatic printing of the workfile variables: When using 'pca' to save > some PCs to the workfile, gretl automatically prints all the variables > in the workfile. IMHO this contaminates the script output for no good > reason (I currently have thousands of variables in there, and it really > is a long list in the output...). So could this be switched off? You can switch it off with "set messages off". However, in CVS I've now made it quieter even without setting message off.