11-24-2016 01:19 AM - edited 11-24-2016 01:34 AM
My dataset has 10 variables and 2000 cases. All variables are continuous. I would like to "standardize" each variable column. Then average those 10 columns, and compare the summary averages for each case, say, sorting from high to low.
I know that several variables data is bi-modal, as opposed to centered, with more data occurring at the extremes.
I'm wondering what the best standardization method might be. SAS offers several. STD, MAD, IQR, ABW, and others.
STD is common -- converting to Z-score: (X1 - mean of X1)/standard deviation of X1. Some of the others are apparently more 'robust,' however, with respect to outliers, and, I suppose, certain other data anomilies.
I'm tentatively thinking of using one of the more esoteric 'robust' ones, such as IQR, based on an example given in SAS documentation.
I'd greatly appreciate hearing your thoughts or suggestions on how best to proceed.
11-27-2016 05:03 AM
With todays computing power, and SAS algorithms, I'm suspecting that an all-around better method of standardization now exists, than traditional std.
True or not? And which one is the new top method?
Thanks for comments.
11-28-2016 08:54 AM
I suspect @Ksharp meant principal components analysis, which can be performed in PROC PRINCOMP.
The principal components are based on the correlation matrix of the original variables, which as he said, means you are effectively using Z-scores. But the great thing about PCA is that it will produce a linear combination that would account for the greatest possible amount of variation among the original variables. That would be in Principal Component 1. Principal Component 2, a second linear combo of the original vars, would account for the largest amount of variation left over after PC1. Etc., etc.