The replicability and utility of three personality types

Abstract

Personality types are construed as constellations of features that uniquely define discrete groups of individuals. Types are conceptually convenient because they summarize many traits in a single label, but until recently most researchers agreed that there was little evidence for the existence of discrete personality types. Several groups of researchers have now proposed replicable, empirical person clusters based on measures of the Five‐Factor Model. We consider several methodological artifacts that might be responsible for these types, and conclude that these artifacts may contribute to the replicability of types, but cannot entirely account for it. The present research attempts to replicate these types in four large and diverse adult samples: the Baltimore Longitudinal Study of Aging (N = 1856); the East Baltimore Epidemiologic Catchment Area study (N = 486); the University of North Carolina Alumni Heart Study (N = 2420); and an HIV risk reduction intervention study (N = 274). A clear replication (kappa = 0.60) of the proposed types was found in only one sample by one standard of comparison. The failure of the three personality types to replicate in three of the four samples leads to the conclusion that they are not robust empirical entities. Type membership predicted psychosocial functioning and ego resiliency and control, but only because it summarized trait standing; dimensional trait measures were consistently better predictors. Nevertheless, while the types do not refer to distinct, homogeneous classes of persons, they do have utility as convenient labels summarizing combinations of traits that relate to important outcomes. Published in 2002 by John Wiley & Sons, Ltd.

Head‐to‐head comparison of the predictive validity of personality types and dimensions

“…This approach allows one to evaluate not only the overall predictive power of multiple types and multiple dimensions but also the incremental predictive validity of the dimensions over the types, and vice versa. 1 Costa et al (2002) applied this approach to two data sets, contrasting the predictive power of five Big Five scales (NEO-PI-R; Costa & McCrae, 1992) with two dummy-coded indicators for three personality types that were empirically derived from the multivariate data through a clustering procedure proposed by Asendorpf et al (2001). For numerous continuous cross-sectional personality correlates, they found substantial predictive power of the types and the dimensions, and significant incremental validity of the dimensions over the types but not vice versa.…”

“…To test hypothesis 1(i), these personality dimensions were predicted by both the three-and the six-type solution. The three-type solution provided a test as parallel as possible to the Costa et al (2002) analysis (two dummy-coded variables are needed to represent all type differences) whilst the six-type solution was fairer to the type approach with regard to the number of predictors (five dummy-coded variables are needed to represent all type differences).…”

“…Thus, studies such as that of Costa et al (2002) that play three types obtained by variablecentred assessments against five uncorrelated traits in the cross-sectional prediction of continuous outcome dimensions can be considered unfair for studies of the general predictive utility of type versus dimensional approaches because every factor that potentially moderates the relative advantage of dimensions over types was chosen in favour of dimensions. Studies that vary these moderating factors within or between samples are needed to draw a more balanced picture of the relative (dis)advantages of a type approach to prediction.…”

“…If k > 2 types are distinguished, k À 1 dummy-coded variables capture all information on differences between the types, and can be used for prediction through multiple regression (see e.g. Costa, Herbst, McCrae, Samuels, & Ozer, 2002).…”

“…Everything else being equal, 5 predictors yield stronger predictions than 2 predictors unless the 5 predictors show much higher intercorrelations than the 2 predictors. Costa et al (2002) used NEO-PI-R factor scores that maximized the predictive strength of the dimensions because they are uncorrelated by definition; in contrast, dummy-coded type indicators are correlated because membership for the non-type overlaps across types.…”