PCA Procedure

computes the principal components from the covariance matrix. If you omit this option, the correlation matrix is analyzed. The COV option causes variables that have large variances to be more strongly associated with components that have large eigenvalues, and it causes variables that have small variances to be more strongly associated with components that have small eigenvalues. You should not specify this option unless the units in which the variables are measured are comparable or the variables are standardized in some way.

names the input data table for PROC PCA to use. The default is the most recently created data table. libref.data-table is a two-level name, where

enables PROC PCA to use graphics processing units (GPUs) to perform model computation on GPU-equipped machines (one GPU is used on each machine). Currently the eigenvalue decomposition (EIG) method, the nonlinear iterative partial least squares (NIPALS) method, and the iterative method based on random projection (RANDOM) support this option. If you specify any other principal component extraction method or specify the BY statement or PARTIAL statement, the GPU option is ignored. If no GPU is available, the CPU is used to perform model computation.

PROC PCA uses CUDA 11.8, and the NVIDIA V100 and P100 GPUs are supported.^[2]

specifies the principal component extraction method to be used. You can specify the following values:

By default, METHOD=EIG. If you specify METHOD=ITERGS, METHOD=NIPALS, or METHOD=RANDOM, the following options in the PROC PCA statement are not supported: COV, NOINT, OUTSTAT=, PARPREFIX=, SINGULAR=, and STD.

specifies the number of principal components to be computed. If you specify METHOD=ITERGS, METHOD=NIPALS, or METHOD=EIG, the default is the number of variables. If you specify METHOD=RANDOM, the default is 1. The number must be an integer greater than or equal to 0. If N=0, the default is used.

omits the intercept from the model. In other words, this option requests that the covariance or correlation matrix not be corrected for the mean. When you specify this option, the covariance matrix and, hence, the standard deviations are not corrected for the mean. If you want to obtain the standard deviations corrected for the mean, you can obtain them by using a procedure such as PROC MEANS in Base SAS.

creates an output data table to contain means, standard deviations, number of observations, correlations or covariances, eigenvalues, and eigenvectors. libref.data-table is a two-level name, where libref refers to the library, and data-table specifies the name of the output data table. For more information about this two-level name, see the DATA= option and the section Using CAS Sessions and CAS Engine Librefs.

If you specify the COV option, the output data table contains a covariance matrix, and the corresponding _TYPE_ variable value is COV or UCOV, depending on the value of the NOINT option; otherwise, the output data table contains a correlation matrix, and the corresponding _TYPE_ variable value is CORR or UCORR, depending on the value of the NOINT option. If you specify the PARTIAL statement, the OUTSTAT= data table also contains R-squares.

For more information about the OUTSTAT= data table, see the section OUTSTAT= Data Table.

controls the plots that are produced through ODS Graphics. When you specify only one plot-request, you can omit the parentheses around it. Some examples follow:

plots
plots=none
plots=(eigen profile)
plots(unpack)=scree
plots=pattern(ncomp=3 circles=25 50 75 1)

ODS Graphics must be enabled before you can request plots. For example:

ods graphics on;
proc pca plots=all;
   var x1--x10;
run;
ods graphics off;

You must specify the PLOTS option in order for the PCA procedure to produce plots. If you do not specify a plot-request, PROC PCA produces the scree plot by default.

You can specify the following global-plot-options:

You can specify the following plot-requests:

specifies a prefix for naming the principal components. By default, the names are Prin1, Prin2, …, Prinn. If you specify PREFIX=Abc, the components are named Abc1, Abc2, Abc3, and so on. The number of characters in the prefix plus the number of digits required to designate the variables should not exceed the current name length that is defined by the VALIDVARNAME= system option.

specifies a prefix for naming the residual variables in the OUTSTAT= data table. By default, the prefix is R_. The number of characters in the prefix plus the maximum length of the variable names should not exceed the current name length that is defined by the VALIDVARNAME= system option.

specifies the singularity criterion, where . If a variable in a PARTIAL statement has an R-square as large as when predicted from the variables listed before it in the statement, the variable is assigned a standardized coefficient of 0. By default, SINGULAR=1E–8.

standardizes the principal component scores in the OUT= data table in the OUTPUT statement to unit variance. If you omit this option, the scores have a variance equal to the corresponding eigenvalue. Note that this option has no effect on the eigenvalues themselves.

specifies the divisor to be used to calculate variances and standard deviations. By default, VARDEF=DF. The following table displays the values and associated divisors:

Table 4: continued

Value	Divisor	Formula
DF	Error degrees of freedom		(before partialing)
			(after partialing)
N	Number of observations	n
WEIGHT | WGT	Sum of weights
WDF	Sum of weights minus one		(before partialing)
			(after partialing)

In the formulas for VARDEF=DF and VARDEF=WDF, p is the number of degrees of freedom of the variables in the PARTIAL statement, and i is 0 if you specify the NOINT option and 1 otherwise.