High loadings reflect the importance/contribution of a particular item in a specified component. Means are related to the unit and scale of measurement of the item. If SD is standard deviation then it is reflects higher variation in responses. If SD is standard scores then standard scores are not related to item it is a relative position of any respondent, high standard scores possibly lead to outlier behavior of the respondents.

PCA uses the variability of each factor in your data set. You need to normalize your factors before you use PCA. If you don't, each PC will be dominated by one of your factors.

Suppose that you have 3 factors in your experiment: Chemical Conc, Voltage and Temp. If you conc ranges from 0.1M to 0.5M, Voltage ranges from 5V to 10V and Temp ranges from 20C to 50C, and you don't normalize your data, PC 1 will be Temp, PC 2 will be Voltage, PC 3 will be Conc. If you change the range of Conc from M(moles) to uM(Micromoles), Conc now ranges from 10,000uM to 50,000uM. If you run PCA on this data without normalizing, PC 1 will now be Conc, PC 2 temp and PC 3 voltage.  

Dear Riaz, my opinion is that means and SD have no direct effects on PCA. This is a statistical technique that works upon correlation matrix, not on your raw data matrix. For instance, you can enter the correlation matrix (even a polychoric one) as the object to analyze. As most statistical techniques, Correlations, and hence PCA, assume univariate and multivariate normality of the distributions of variables, among a number of other prerequisites (often ignored). If your data are not normally distributed you have a problem, but PCA will run anyway.

I disagree with Andrew because variable ranges have not any effect on correlations per se. In his example (I simulated it), when Conc, Temp and Voltage are intercorrelated, a unique principal component arises. Loadings depend on the strenghth of the relationship between the item and the component: the amount of item variance explained by the component; not on means and SDs.

Your loadings depend on the size of the correlations of each of your (low mean-high SD) variables with the rest of the variables in your study. This may be by chance or precisely because they are ill-distributed. My advise is to take a step back to understand your item distributions. To elliminate multi- and uni-variate outliers and to try to achieve PCA prerequisites, and then perform PCA again.

Hope this will help.

Luis, 

If you don't normalize the data before running PCA, the range/variability of the factors WILL have an effect on your PCA scores. If you normalize first, then there is no issue. A lot of different "textbooks" and publications that use PCA in the sciences claim normalization is not necessary because they feel their data is special. The data set I refer to above came out of one of my "textbooks" The authors didn't feel as thought they needed to normalize the data. When you go back and make each of the changes I stated, the PCA results come out as I stated.

If you do things properly, then there is no issue.    

Andrew, of course normalization is a must, but ranges do not affect. For instance, when y = x + 5 (or any other number), rxy = 1; as well as when y = x; and when y = 5 (or any number) * x. Ranges, means, and variances are quite different, but correlations (the input in PCA) are always 1. So I think one must be sure that data distributions are normal plus other prerequisites (Tabachnick and Fidell have a good handbook to follow) before PCA... I guess we agree but I did not explain properly.