Ipsative Data

The term ipsative (from the Latin ipsum, meaning self) was originally coined by Raymond Cattell in 1944, in the framework of factor analytic approaches for psychological assessment to describe measurements that are meaningful only relative to a person but that cannot be directly compared between persons. For example, if two persons S1 and S2 are asked to ranked three occupations A, B, and C, these two persons can give the same ranking [A, B, C] but S1 considers these occupations as favorite occupations whereas S2 considers these occupations as dreadful. Therefore, even though the preference order on the occupations can be compared, the participants cannot be compared, as they cannot be considered similar because, on a continuum describing their preference for these occupations, they would represent two extremes (i.e., one rater loves everything and the other hates everything). So, with ipsative data, variables (or stimuli) can be compared but participants cannot, and so, current consensus discourages using ipsative data for psychological testing and assessment, except for personal counseling (see, e.g., Cornwell & Dunlap, 1994; Kline, 2015).

When used to compare variables or stimuli, ipsative data need to be analyzed in a different way than the usual statistical approaches such as factor analysis or principal component analysis (PCA).

For these methods, as the measurements are considered quantitative and comparable across participants, [Page 737]variables are routinely centered and normalized (because the data of two different participants are assumed to be measured on the same scale). But for ipsative data, these assumptions are not met and therefore different scaling and centering schemes need to be considered. For example, Paul Horst in an early factor analysis text in 1965 suggested to center the rows (or also to center both rows and columns) and, in some cases, to normalize the rows instead of the columns (contrary to centering, which can be performed on both rows and columns, normalizing, such as, e.g., Z scores, can be performed on only one set). Some alternatives could be to rank order the rows and perform a noncentered multivariate analysis such as a noncentered and nonnormalized PCA. It is also worth noting that some normative measurements such as, for example, Likert-type scales, or other scoring systems, can be considered ipsative if there is reason to believe (as could often be the case in educational measurement practice when comparing different raters) that the raters roughly agree on ranking the stimuli but do not agree on the mean or the variability of the scoring system: In this case, the raters agree on the ranks (e.g., this is the best work, this is the worst work) but do not agree on the basic score (e.g., the best score for one rater is 10 out of 50 but another rater’s best scores is 50 out of 50).

Figure 1 Principal Component Analysis of the Data From Table 1 When the Data Are Considered Normative (the Data Are Centered and Normalized by Column)

Note: The eigenvalues are denoted by λ and the percentage of explained variance are denoted by τ.

...