A cautionary note concerning a commonly applied eigenanalysis procedure

Owing to the complex nature of eigenanalysis programming algorithms, most atmospheric science research utilizing empirical orthogonal functions (EOFs), principal component analysis (PCA), or common-factor analysis (CFA) use the “canned” computer programs on statistical packages such as EISPACK, IMSL, SAS, BMDP, etc. These programs usually encourage truncation of the eigenvector series according to some criterion to exclude noise and facilitate data reduction; the EOFs/PCs/CFs retained are then often physically interpreted (with or without further linear transformations) according to which variables have salient magnitudes. Results presented here illustrate that following this type of analysis procedure can lead to the interpretation of coefficients as being physically significant, when they are purely noise. Several checks to avoid such a pitfall are suggested.