Development of temperament and age differences

Manifestations of temperament transform over development, with rapid change during infancy [8]. Positive emotionality (e.g., smiling), rarely expressed during the newborn period, is observed more reliably between ages two and three months, and increases in expression throughout the first year of life [8,9]. Levels of activity, approach, distress to limitations, and fear increase throughout the first year of life as well [10–14]. Anger reactions across infancy appear to follow a U-shaped trajectory [12,15]. The decrease in anger responses occurring between 2 and 6 months of age has been linked to greater flexibility in attention shifting [16]. In the second half of the first year, infants are likely to respond with anger when unable to grasp an attractive stimulus that has been placed out of reach, or when a caregiver has removed a forbidden object. Fear generally increases throughout the second half of the first year of life [10,12–14], with inhibition of approach toward novel and/or intense stimuli “coming online” [14,17].

The developmental course of attentional orienting has been described as U-shaped in the first year of life [18]. Carranza and colleagues [12], for example, noted decreases in Duration of Orienting between 6 and 9 months, followed by an increase between 9 and 12 months. Toward the end of the first year, skills associated with the development of the executive attention system may contribute to the flexibility of orienting reactions [19–21]. Infants also gain communication skills rapidly during the first year of life [22,23], and thus exhibit greater vocal reactivity over time.

With respect to age/developmental differences discerned via the IBQ-R, older infants obtain higher scores on Approach, Vocal Reactivity, High Intensity Pleasure, Activity Level, Perceptual Sensitivity, Distress to Limitations, and Fear, whereas younger infants’ scores are higher for Low Intensity Pleasure, Cuddliness/Affiliation, and Duration of Orienting [24,25]. More recent longitudinal investigations provided further evidence of increases in Fear across the first year of life [5,26], also noting increases in Distress to Limitations and Sadness, albeit not always linear in nature. Falling Reactivity was associated with a quadratic trajectory, with increases followed by declining values later in infancy. Increasing trajectories were noted for attributes associated with Positive Affectivity/Surgency, with trends toward greater Activity Level, Smiling and Laughter, High Intensity Pleasure, Approach, Perceptual Sensitivity, and Vocal Reactivity later in infancy. Growth modeling provided evidence of nonlinear changes in Duration of Orienting, Soothability, Cuddliness, and Low Intensity pleasure, wherein initial growth in values was followed by decreases later in infancy [5]. These findings are largely consistent with prior research relying on different measurement approaches. Although the data examined in this study are cross-sectional in nature, earlier longitudinal evaluations are informative as their results speak to the importance of age in shaping temperament presentations, and vice versa–temperament features as predictors of infant age. It should be noted that no study to date has explored the latter, that is, used temperament dimensions to classify infants with respect to their age, likely due to sample size limitations and only recently available methodological advances in empirically based classification techniques.

Gender differences in temperament

Although a number of gender differences in temperament have been reported for older children and adults, fewer exist for children younger than one year of age [8,25,27,28]. Differences in infancy have been limited to activity level and fear/behavioral inhibition. Higher activity level and approach is evident in boys [29,30], with girls exhibiting greater hesitation in approaching novel objects [14,31]. Campbell and Eaton [29] applied meta-analytic procedures to summarize 46 studies addressing activity level in infancy, estimating the size of the gender difference at 0.2 standard deviations based on objective measures (parent-report measures estimated the difference to be smaller). Gender differences in approach-withdrawal have been reported for samples from different countries [30,32–34], with parents rating boys higher in approach. Martin et al. [31] reported a large and significant gender difference for distress to novelty, with 6-month-old girls receiving higher scores.

Gender differences also have been documented with the IBQ-R, as boys received higher scores on Activity and High Intensity Pleasure, and girls higher scores on Fear [24,25,35,36]. Infant gender also predicted intercept values of Fear trajectories, with girls demonstrating higher levels at 4 and 6 months [5,26]. Girls also started out at lower values (i.e., intercept estimates) for Activity Level, Approach, and High Intensity Pleasure. Similar to age/developmental differences research, gender-related temperament studies have only compared temperament for boys and girls, not considering gender classification based on temperament features. Importantly, age- and gender-based temperament distinctions have not been considered jointly, discerning whether age-related changes inform gender differences.

Present study

In this study, we leveraged IBQ-R data collected across multiple laboratories (N = 4,438) to further investigate age and gender differences in infancy, addressing yet unanswered questions. Specifically, algorithmic modeling techniques were used to discern the extent to which the 14 IBQ-R subscale scores (referred to as features) accurately classified participating children as boys (n = 2,298) or girls (n = 2,093; 47 children were missing gender data) and into three age groups: youngest (< 24 weeks; n = 1,102), mid-range (24 to 48 weeks; n = 2,557), and oldest (> 48 weeks; n = 779), because of previously noted gender-based variability [14,29–34] and significant developmental differences among these age groups (e.g., with respect to brain growth and maturation; [37,38]). This study addresses an important gap in research, being the first to consider temperament attributes as determinants of age and gender groupings, quantifying the extent to which early reactivity and regulation provide the features necessary for accurate prediction. Importantly, this work also allows for simultaneous classification of age and gender categories, providing an opportunity to consider the extent to which gender differences are informed by infant age, and to our knowledge, this is the first to study to do so. That is, despite prior demonstrations of reliable age and gender differences in temperament, the two classifications have not been considered jointly, examining whether gender differences were age dependent in a single investigation. Moreover, this effort provides a new direction for infancy and temperament research, serving as a methodological demonstration of machine learning applications, not yet utilized in these areas of scientific inquiry. This meta-analytic data driven effort is the first to rely on advanced machine learning techniques using temperament features to classify infants into age and gender groups, rather than compare temperament of children who vary in age and gender, considering these classifications simultaneously. This cross-laboratory effort also overcomes prior limitations associated with small samples that were not representative, producing results circumscribed in terms generalizability.

Measures

The Infant Behavior Questionnaire-Revised (IBQ-R; [24]. This parent-report measure of temperament was developed for infants between 3- and 12-months of age. The IBQ-R contains 191 items, which yield 14 scales: Activity Level, Smiling/Laughter, Approach, High Intensity Pleasure, Perceptual Sensitivity, Vocal Reactivity (loading onto Positive Affectivity/Surgency); Fear, Distress to Limitations, Sadness, Falling Reactivity (Negative Emotionality); Duration of Orienting, Soothability, Cuddliness/Affiliation, Low Intensity Pleasure (Regulatory Capacity/Orienting). Individual items are rated on a 7-point scale reflecting the frequency of occurrence of the behavior in the past week (two weeks for less frequent events, such as encounters with unfamiliar settings/adults). Reliability of the IBQ-R has been supported for mothers and fathers, as well as samples from different cultures, with Cronbach’s α values ranging from .77 to .96 [39–41]. Evidence supports the predictive and construct validity of IBQ-R scores [42–44]. Cronbach’s α values for the 14 subscales included in the current analysis, derived from 29 datasets, ranged from .74 to .89 (mean α = .82). These temperament features were used to classify children into gender and age categories via Machine Learning algorithms.

Procedure

Data sets (N = 29) were acquired by emailing researchers who requested the IBQ-R or published research using the instrument between 2006 and 2019. All of the researchers had received approval from their respective Human Research Protection Programs (HRPPs)/Institutional Review Boards (IRBs) prior to initiating data collection: Human Studies Committee at the Brigham and Women’s Hospital in Boston, MA and the Icahn School of Medicine at Mount Sinai in New York; IRB at Boston Children’s Hospital; Pennsylvania State University IRB; Rutgers-Newark IRB; Virginia Tech IRB; University of North Carolina at Greensboro IRB; Emory IRB; University of Washington IRB Committee D; Northern Illinois University IRB #1; Brown University HRPP/IRB; IRB of the University of Minnesota’s Human Research Protection Program; University of Michigan Health Sciences and Behavioral Sciences IRB; Health Sciences IRB at University of Wisconsin; Harvard University Committee on the Use of Human Subjects in Research; University of California, Irvine HRPP/IRB; University of Missouri IRB; IRB of Western Kentucky University; University of North Carolina at Chapel Hill IRB; Western Washington University IRB; University of Virginia IRB for the Social and Behavioral Sciences; Wayne State University IRB; Colorado Multiple IRB; obtaining written informed consent. Contributors were asked to provide item level data from the IBQ-R as well as infant age, gender, and race. For all participants, the IBQ-R was completed by the infant’s mother. See Table 1 for a brief description of the samples.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Sample descriptions.

https://doi.org/10.1371/journal.pone.0266026.t001

Analytic strategy

Descriptive statistics across gender and age groups were computed first (Table 2). We then constructed a model framework allowing us to assess the utility of fine-grained temperament dimensions with respect to gender and age classifications. This framework resulted in a total of five (5) model types, which included: 1) gender: boys vs. girls; 2) age groups: youngest (< 24 weeks) vs. mid-range (24 to 48 weeks) vs. oldest (> 48 weeks) infants; and gender by age group analyses: 3) boys vs. girls in the youngest age group; 4) boys vs. girls in the mid-range age group; 5) boys vs. girls in the oldest age group. Classification of infant gender within age groups allows us to determine if predictive strength of gender-based classification is more accurate for younger vs. older infants.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Descriptive statistics for the temperament subscales by gender and age group.

https://doi.org/10.1371/journal.pone.0266026.t002

Established machine learning techniques, methodologically rigorous and shown to provide reliable/reproducible results, were used in this study (e.g., [45,46]). Specifically, for all models, we used repeated 10-fold cross-validation partitioning with random assignment: a training dataset including 70% of the sample, and 30% reserved as a hold-out dataset (testing) to evaluate the predictive utility of the trained models. A total of 11 different algorithms were considered for each model type, including: (1) linear discriminant analysis; (2) generalized linear modeling; (3) support vector machines; (4) K-nearest neighbor; (5) naïve bayes; (6) classification and regression trees; (7) C5.0 classification; (8) bootstrapped aggregated trees; (9) ensembled decision trees (Random Forest; [47,48]); (10) gradient boosting; and (11) multi-class adaptive boosting (AdaBoost). These algorithms were chosen based on their applicability and widespread use in the classification modeling literature [45,46], and in order to achieve most robust and replicable results discernable across multiple modeling techniques. The aforementioned models were then compared to discern the most effective classification of infant gender and age with temperament features based on misclassification rates, Cohen’s kappa coefficients, and sensitivity and specificity via the area under the curve (AUC) from Receiver Operator Curves (ROC), considered as indicators of predictive accuracy.

Misclassification provides a simplistic posterior assessment of model classification based on contingency tables and is often used for initial classification and model accuracy evaluation. Accuracy indicators, reported herein, represent the inverse of misclassification rates. Cohen’s kappa coefficient assesses reliability of categorization, which incorporates chance agreement, is normalized, and can range from -1 to 1. Kappa values will typically be lower than overall misclassification indictors, as it represents a more conservative estimate given its assessment of accuracy compared to random assignment. The area under an ROC curve (area under the curve, or AUC) is a third metric used to evaluate the accuracy of binary classifiers, which encapsulates both Type I and Type II errors [49]. However, ROC-AUC is limited insofar as it does not take predicted probability values and goodness of fit of evaluated models into account. While all three indicators provide unique assessments of classification accuracy, overall misclassification rate (or, inversely, accuracy) is the most broadly used metric for classification evaluation [50]. For all of the model classification indices, higher values (i.e., closer to 1) can be considered superior, indicative of more optimal performance.