principal component analysis

[gogusteriade]

I'm trying to estimate some unobservable factors driving a set of variables by means of first r principal components. I'm interested in the actual values of these components since I want to reestimate the initial equation with the estimated factors as additional regressors. But in the output obtained after running principal component analysis I see only the eigenvalues and the factor loadings. How do I compute the estimated values of the principal components?

Thank you!

[EViews Glenn]

Use Proc/Make Principal Components

[gogusteriade]

In order to run principal component analysis on a set of variables (so i can estimate the unobserved factors driving them), do these variables have to be stationary?

[EViews Glenn]

"Have to be" is an ambiguous expression. I will note that properties of the covariance or correlation matrix do rely on iid'ness.

[gogusteriade]

So if I run principal component analysis on the series in first differences (which are stationary), can I still use the principal components obtained as an estimation of the factors driving the initial series (in levels)? (i need to compare the factors driving the initial series to some actual macroeconomic variables)

Thank you!

[trubador]

Principal components belong to the variable from which they are extracted, since the method changes the coordinate system and transforms the variable accordingly. Therefore, operating on levels and differences will not yield identical results. Even if you decide to work with level series, you should still be careful about scale differences among the variables.

[ank]

I have extracted the principal components of a matrix of data, and I want to see the percent of the variance of each time series explained by each principal component retained. Therefore, I regressed each time series on each principal component and retained the R squared. My question is: if some of the time-series are non-stationary, is there a possibility that the respective regression is spurious (so the R square I get isn't valid?)?
Thank you!

[trubador]

This is a problematic issue. Mixing the stationary variables with non-stationary variables may not be a good idea, since they may not be correlated. Principal components analysis is appropriate (and effective) if there is a significant correlation among variables. Otherwise (in case of orthogonality), each principal component will account for the same amount of variance, which would be meaningless. Moreover, if non-stationary variables do have larger variances, then they will probably recieve larger weights and dominate the principal components as a result. Therefore, unless the variances do indicate the importance of the variable at hand, then you should either normalize or standardize your data.

Of course, the objective of your study and properties of your variables are also important and should be taken into consideration to reach a proper conclusion.

[mczhang]

Use Proc/Make Principal Components

I cannot find "Make Principal Components" in Proc. I have 23 countries' real GDP data in time series (from 1970 to 2009). How can I generate the common factors which influence those countries GDP over time as a new series (like the actual values and can be stored/saved)?
Thank you!

[EViews Glenn]

Make sure you have a group...

[mczhang]

Hi there, I am quite new to eviews. I am trying to estimate the changes of the common factor Z on quarterly real GDP growth (as a new time series that can be stored and for future use) of OECD countries from '70-'08. I have tried the way of "select the GDP series as group, then Proc/Make Principle Componant". But I have only got one constant figure for the principle Componant for the entire samply period and the co-relation coefficients between the countries which are also constant.
Does anyway know how this could be done in Eviews 6?