LAUSR

1Freie Universität Berlin, Berlin, Germany. 2Ulm University, Ulm, Germany. 3Humboldt-Universität zu Berlin, Berlin, Germany. *e-mail: [email protected] It was recently suggested that at least four meaningful and robust types of personality exist1. As a notable improvement to previous approaches, the authors applied Gaussian mixture models to very large datasets with various questionnaires. While we concur that the authors’ analyses reveal four meaningful clusters in the Johnson-300 data, we also found that only 42% of the sample was associated with those four personality types. That is, the vast majority of individuals could not be classified within a distinct, meaningful cluster. Against the background of these findings, Gerlach et al.’s suggestion that this typology might potentially be useful in personality assessment seems premature1. It is widely accepted that the Big Five comprise five broad traits of personality2. Extending this multi-dimensional view on personality, typically three different personality types have been proposed3,4. In contrast, it was recently suggested that four meaningful types of personality exist1. As a notable improvement to previous approaches, these authors1 applied Gaussian mixture models to large datasets. Gaussian mixture models tend to yield preferable results for overlapping clusters of varying size and shape when compared to classical clustering methods5,6. As one implication of their results, the authors1 suggested that the development of reduced personality measures based on those four personality types could be used in clinical and organizational assessments. However, before following such recommendations, the value of the cluster solution for individuals needs to be considered. We identified the number of meaningful clusters in the Johnson-300 (n = 145,388) dataset based on the optimal number of clusters (nc) presented by the authors (nc = 13). In line with ref. 1, we also found four meaningful clusters. We additionally evaluated the final Gaussian mixture solution based on posterior probabilities. Posterior probabilities can be derived directly from the density distribution and reflect the probability of any individual belonging to a certain cluster. Surprisingly, in our analysis we found only 42% of the sample was associated with one of the four meaningful personality types. Personality models should be exhaustive, especially when put to practical purposes. Here, the vast majority of individuals could not be classified within a distinct, meaningful cluster. Furthermore, classification of individuals into specific personality types was associated with high uncertainty. The average posterior probability of individuals associated with one of the four types was only mean M = 0.51 (s.d. = 0.17), which may reflect large cluster overlaps and a rather unstable solution of the Gaussian mixture model. This uncertainty with regard to the generalization of four meaningful personality types is further demonstrated in ref. 1, where the authors found more than four meaningful clusters for all other analysed datasets (Johnson-120, MyPersonality-100, BBC-44). In fact, we re-analysed the Johnson-300, Johnson-120 (n = 410,376) and BBC-44 (n = 386,375) datasets (the MyPersonality-100 dataset is no longer publicly available), strictly following the code provided by the authors. We also found numbers of meaningful clusters that differed from the four proposed (see Supplementary Fig. 1). Against the background of these findings, the suggestion1 that the use of those four types in personality assessment might potentially be useful seems premature. In particular, the low posterior probabilities demonstrate a major difficulty regarding classification of an individual to a specific personality type, with misclassifications potentially leading to erroneous interventions. Notably, other studies that applied Gaussian mixture models to personality data typically report acceptable rates of posterior probability. However, the authors1 analysed several exceptionally large datasets using a slightly different approach by contrasting meaningful clusters with spurious ones. If placing confidence in this procedure, a more appropriate interpretation of the results might be that personality types are less robust and exhaustive than assumed. In fact, the authors1 themselves state that the four personality types show substantial overlap and that only three types could be found in the BBC-44 dataset. Overall, the shortcomings identified with this proposed typology are in line with general criticism of typologies7. Furthermore, the proposed typology is currently lacking evidence for criterion validity. However, previous head-to-head comparisons between personality types and domains demonstrated less explained variance for typologies than for dimensional personality models8. Thus, we would urge practitioners to rely on dimensional assessments rather than advocating the use of typologies, whenever possible. That said, we acknowledge that the measurement of dimensional personality traits also comprises problems9—in particular, uncertainty or unreliability of these measures influences their practical usefulness. However, there are remedies (for example, confidence intervals) that practitioners can use to deal with this. Similar approaches for dealing with uncertainty in type allocation are less widely used. Moreover, typologies generally build on dimensional measures and thus combine the shortcomings of such measures with their own problems. Research should now build on this method and the results presented in ref. 1 to determine whether typologies can be found that cluster individuals reliably and comprehensively and yield substantial test-criterion validity evidence.