Differences between children appear very early in life. One such difference is children’s temperament. It has been an important clinical and research issue. However, the definition of temperament is debatable: many researchers have defined it differently (Goldsmith, Buss, Plomin, Rothbart, Thomas, Chess, Hinde, & McCall, 1987). Most of the research on child temperament has been focused on dimensions of temperament. Factor analysis of rating scales of temperament has yielded several factors. Whereas the majority of studies on temperament are variable-centered, temperament structure has rarely been challenged from a person-centered perspective (i.e., typology of temperament). A seminal report by Thomas & Chess (1977) identified three types: easy, difficult, and slow-to-warm-up. Easy babies are cheerful, easy to calm, and able to adjust to new situations without difficulty. Difficult babies are slow to adjust to a new experience and react negatively and intensely. Slow-to-warm-up babies are difficult at first but gradually become easier.

Although Thomas & Chess’s (1977) proposal has gained world-wide recognition, little empirical evidence has been demonstrated. In 1995, Caspi & Silva (1995) used the scores of 3 temperament factors (lack of control, approach, and sluggishness) to perform cluster analysis of over 800 3-year-old children and they identified five clusters, i.e., groups of children: under-controlled, inhibited, confident, reserved, and well-adjusted. A similar but different approach was conducted with Q-sort patterns (Asendorpf & van Aken, 1999) that identified three prototypic patters. Robins, John, Caspi, Moffitt, and Fisher (2001) analyzed the data of the California Child Q-Set (CCQ) among children aged 12 to 13 years old by Q-factor analysis. This yielded three types: overcontrollers, undercontrollers, and resilients. Aksan et al. (1999), in a multi-wave (1, 4, and 12 months; and 2, 3, and 4 years) study, used configural frequency analysis and yielded two types: controlled-nonexpressive and noncontrolled-expressive. In Sanson et al.’s (2009) study, 200 children were assessed on four occasions (4 - 8 months, 1 - 2 years, 2 - 3 years, and 3 - 4 years) by different scales (Revised Infant Temperament Questionnaire, Toddler Temperament Scale, and Childhood Temperament Questionnaire). A hierarchical cluster analysis with a dendrogram showed that a 4-cluster model was the best. Subsequently, k-means with 3- to 6-cluster models were used to measure distances between cluster centers. Again, a 4-cluster model (nonreactive/outgoing, high attention regulation, poor attention regulation, and reactive/inhibited) was found to be the best. Prokasky et al. (2017) examined three samples of children (n = 96, 187, and 757) aged around 4 years old with the Child Behavior Questionnaire (CBQ: Rothbart et al., 2001). Seven subscales (activity, anger, approach, fear, shyness, attention focusing, and inhibitory control) were entered into a hierarchical cluster analysis using Ward’s method with squared Euclidean distance as a means of distance between cases. The best model was identified by comparing k-mean cluster analyses and the best was defined as the one that showed the most similar patterns between samples. As a result, a 6-cluster model was identified as the best: unregulated, reactive, bold, subdued, regulated, and inhibited. These studies have not yet arrived at a consensus as to the best temperamental typology possibly because of, among other reasons, use of different temperament measures and different clustering methods.

One of the statistical tools used to identify types according to individual differences is cluster analysis (Borgen & Barnett, 1987). Widely used clustering algorithms, i.e., agglomerative hierarchical cluster analysis and k-means cluster analysis, however, suffer from methodological drawbacks. The former is characterized by ambiguity of determining the appropriate number of clusters whereas the latter demands that the researcher determine the number of clusters a priori. In the two-step cluster analysis, the number of clusters is automatically determined without the researcher’s idiosyncrasy. In the first step, the initial number of clusters is calculated by means of the Schwartz Bayesian Criterion or the Akaike Information Criterion. This is then followed by refinement by finding the largest increase in distance between the two closest clusters in each hierarchical clustering stage. The two-step cluster analysis has recently been used by social science researchers (e.g., Satish & Bharadhwaj, 2010).

We report here a study of temperamental typology of Japanese toddlers using the EASI survey and a two-step cluster analysis. We also examine the validity of the typology in terms of the children’s internalized and externalized behavior problems.

A two-step cluster analysis yielded 4 clusters. We performed a one-way analysis of variance (ANOVA) for the scores of the 4 EASI subscales. All 4 EASI subscale scores differed significantly (p < .001) between the clusters (Table 2). The first cluster consisted of 288 children. They were characterized by the highest S scores and mildly high A and I scores. The second cluster consisted of 179 children. They were characterized by extraordinarily low E scores, the lowest A and I scores, and mildly high S scores. The third cluster consisted of 288 children. While their I and E scores are almost the same level as the first cluster, they were characterized by mildly low A and S scores. The last cluster consisted of 145 children. They were characterized by the highest E, A, and I scores and the lowest S scores. We interpreted the first, second, third, and fourth clusters as Average-Active, Regulated, Average-Quiet, and Sensitive/Hyperreactive, respectively (Figure 1).

When the children of the four clusters were compared in terms of the CBCL scores, the internalizing and externalizing behavior scores as well as the total score all differed significantly (p < .001). Children in the Regulated cluster scored the lowest in all of the subscale and total scores followed by children in the Average-Active and Average-Quiet clusters. Children in the Sensitive/Hyperreactive cluster scored the highest in all of the subscales and total scores (Table 2).

Our two-step cluster analyses identified 4 clusters. A majority of the children belonged to the Average-Quiet (Cluster 3) and Average-Active (Cluster 1) clusters. Both of these groups of Average cluster children were in the middle of the four clusters in terms of E and I subscales. In addition, Average-Quiet children scored lower in A and S, whereas Average-Active children scored higher these subscales. That is, the children of these categories are ordinary, where the former are quieter and the latter more energetic. On the other hand, the other two clusters of children seem to have extreme traits. The children in the Regulated cluster (Cluster 2) scored lowest in E, A, and I but were very sociable. These children seem to be stable in their surroundings, and friendly with other people. Sensitive/Hyperreactive children (Cluster 4) were the highest in E, A, and I but the

NS, not significant.

least sociable. These children seem to be highly sensitive, emotionally unstable, and unsociable.

Construct validity was sought by associations of the temperament typology and the CBCL subscale scores. As expected, children in the Regulated cluster scored the lowest in terms of both internalized and externalized behavior problems while children in the Sensitive/Hyperreactive cluster scored the highest. Regulated cluster children may easily adapt to a change of surroundings so that they have few behavioral problems. In contrast, Sensitive/Hyperreactive cluster children were extremely sensitive to an environmental change so that their confusion may be expressed as various behavioral problems.

Previous studies on personality and temperament suggested 3 to 6 temperament types in childhood and adolescence (Thomas & Chess, 1977; Caspi & Silva, 1995; Robins, John, Caspi, Moffitt, & Stouthamer-Loeber, 1996; Aksan et al., 1999; Sanson et al., 2009; Prokasky, Rudasill, Molese, Putnam, Gartstein, & Rothbart, 2017). They used different nomenclature to describe temperament types (Table 3). A possible reason for this lack of consensus may be the use of different statistical methods. For example, a combination of hierarchical and k-mean cluster analysis was used by Sanson et al. (2009) and Prokasky et al. (2017), configural frequency analysis by Aksan et al. (1999), and Q-factor analysis by Robins et al. (1996). As noted, a drawback of cluster analyses is how to determine the number of clusters. This often depends on the researchers’ arbitrary impression. The two-step cluster analysis, however, leaves this to the predetermined statistical rules. This is a strength of our study. In the present study, we revealed 4 clusters in Japanese 3- to-4 year-old children using a two-step cluster analysis which excluded researchers’ arbitrariness.