You can make intercorrelation, and if the values between variables are less than 0.30 exclude them from the analysis. 

Hi Dário,

I think ultimately it will depend on the motivation to do the PCA in the first place.

Usually you use the PCA precisely to describe correlations between a list of variables, by generating a set of orthogonal Principal Components, i.e. not correlated; thereby reducing the dimensionality of the original data set. You can then understand the correlation structure in your data by looking at the way variables load out on each resulting PC, and interpret them ecologically. In this context there is usually no need to remove any a priori.

Note, however, that setting aside variables known to be strongly correlated with others can substantially alter PCA results. Therefore, there may be merit in discarding variables  beforehand, if you think they may be measuring the same underlying ecological aspect.

If the objective of the PCA is not to describe relationships among variables, but to exclude highly correlated variables for a subsequent analysis (i.e. dealing with multicollinearity), you can use the resulting PCs as new variables (in case they have some clear ecological meaning). Alternatively you can get pairwise correlation co-efficients to determine which variables are highly correlated and use only one of them by setting some threshold (often ~0.7), or use specific methods such as "Variance Inflation Factors".

Hope it helps.

Best,

Gonçalo

No, you do not need to do correlation analysis. However, any other decision depends on your ultimate goal. What is it? data interpretation, classification, etc.

The answer is: No,

you do not need to do correlation analysis between the variables before running PCA. PCA is perfectly capable of doing this job for you, so you can just run PCA, see how many PC's you need to describe a significant fraction of the overall variation, and continue from there.

You could however want to do something else first: PCA finds a new coordinate system that eliminates correlations between the variables. However, if your data has clusters in it (or nasty outliers), the PCA will find the direction from the center of one cluster to an other cluster (or to an outlier). Although technically correct, this is typically NOT what you are interested in. It therefore makes sense to first run a cluster analysis (dendrogram) on the full dataset, and subsequently run PCA on each of the clusters individually (if they are present). That essentially combines the best of both worlds.

Good luck,

Bas van der Geer

Rule of thumb, u are not required to, but if you do it, it tells you a little bit more about the variable set which correlate.

You dont need to do correlation analysis, but, its recommended that your variables are corrlated each other,so. i recommend you allways check barlett test and KMO, that's important to know if your data is capable to make PCA.

Yes, that is mandatory. if there is no correlation no need of doing PCA anaysis