Page 433 - Jolliffe I. Principal Component Analysis
P. 433

14. Generalizations and Adaptations of Principal Component Analysis
                              398
                              individuals and observations to give a mode with np categories) before
                              finding cross-products. Details will not be given here (see Tucker (1966) or
                              Kroonenberg (1983a), where examples may also be found).
                                The substantial literature on the subject that existed at the time of
                              Kroonenberg’s (1983a) book has continued to grow. A key reference, col-
                              lecting together material from many of those working in the area in the late
                              1980s, is Coppi and Bolasco (1989). Research is still being done on various
                              extensions, special cases and properties of the three-mode model (see, for
                              example, Timmerman and Kiers (2000)). One particular extension is to the
                              case where more than three modes are present. Such data are usually called
                              ‘multiway’ rather than ‘multimode’ data.
                                Although multiway analysis has its roots in the psychometric literature,
                              it has more recently been adopted enthusiastically by the chemometrics
                              community. Volume 14, Issue 3 of the Journal of Chemometrics, published
                              in 2000, is a special issue on multiway analysis. The issue contains relatively
                              little on multiway PCA itself, but there is no shortage of articles on it in
                              the chemometrics literature and in the overlapping field of process control
                              (see, for example, Dahl et al. (1999)). In process control the three most
                              commonly encountered modes are different control variables, different time
                              intervals and different batch runs (Nomikos and MacGregor, 1995).
                                Another context in which three-mode data arise is in atmospheric science,
                              where one mode is spatial location, a second is time and a third is a set of
                              different meteorological variables. It was noted in Section 12.2.1 that the
                              analysis of such data, which amalgamates the p locations and n different
                              meteorological variables into a combined set of np variables, is sometimes
                              known as extended EOF analysis.
                                An alternative strategy for analysing data of this type is to consider pairs
                              of two modes, fixing the third, and then perform some form of PCA on each
                              chosen pair of modes. There are six possible pairs, leading to six possible
                              analyses. These are known as O-, P-, Q-, R-, S- and T-mode analyses (Rich-
                              man, 1986), a terminology that has its roots in psychology (Cattell, 1978,
                              Chapter 12). In atmospheric science the most frequently used mode is S-
                              mode (locations = variables; times = observations; meteorological variable
                              fixed), but T-mode (times = variables; locations = observations; mete-
                              orological variable fixed) is not uncommon (see, for example, Salles et al.
                              (2001)). Richman (1986) discusses the other four possibilities. Weare (1990)
                              describes a tensor-based variation of PCA for data having four ‘dimen-
                              sions,’ three in space together with time. He notes a similarity between his
                              technique and three-mode factor analysis.
                                Some types of multiway data convert naturally into other forms. In some
                              cases one of the modes corresponds to different groups of individuals mea-
                              sured on the same variables, so that the analyses of Section 13.5 may be
                              relevant. In other circumstances, different modes may correspond to dif-
                              ferent groups of variables. For two such groups, Section 9.3 describes a
                              number of techniques with some connection to PCA, and many of these
   428   429   430   431   432   433   434   435   436   437   438