The Stability of Temperament from Early Childhood to Early Adolescence: A Multi–Method,Multi–Informant Examination

Models of the structure of temperament

The content of temperament measures necessarily changes across developmental periods such that indicators in early childhood will be different from those in early adolescence (Neppl et al., 2010). As such, it is necessary to measure developmentally appropriate manifestations of the same traits at different ages. Theoretical models that specify core aspects of temperament are necessary to guide such work. The three–factor model proposed by Rothbart and colleagues (Rothbart, 1981; Putnam, Ellis, & Rothbart, 2001) is arguably the most widely used contemporary model of childhood temperament (De Pauw & Mervielde, 2010; Zentner & Bates, 2008). It comprises the higher order factors of surgency, including high–intensity pleasure, activity, impulsivity, and shyness (reversed); negative emotionality, including discomfort, anger/frustration, sadness, fear, and soothability (reversed); and effortful control (or constraint), including inhibitory control, attentional focusing, perceptual sensitivity, and low–intensity pleasure (Rothbart, Ahadi, Hershey, & Fisher, 2001). This model maps fairly closely onto taxonomies of personality and temperament derived from studies of adults (Caspi et al., 2005; Rothbart, Ahadi, et al., 2000), the best well–known of which are the five–factor and three–factor models (John, Naumann, & Soto, 2008; Markon, Krueger, & Watson, 2005; Tackett, Krueger, Iacono, & McGue, 2008; Watson, Clark, & Harkness, 1994). Tellegen's (1985) influential three–factor model comprises positive emotionality, or a tendency towards positive affect and engagement with the environment; negative emotionality, or a propensity towards anxiety, anger, and fearfulness; and effortful control/constraint, or, at the opposite pole, disinhibition, or a tendency toward impulsivity, risk–taking, and unconventional behaviour. These higher order trait dimensions encompass and explain many of the narrower trait dimensions from the different models of child and adult temperament (Caspi et al., 2005; De Pauw & Mervielde, 2010; Markon et al., 2005; Zentner & Bates, 2008).

Temperamental stability in younger children

Although it is somewhat unclear how rank–order stability of temperament in early childhood will extend to adolescence, shorter term studies can inform expectations (Roberts & DelVecchio, 2000). A meta–analysis of longitudinal studies found average respective retest stabilities of .32, .52, and .45 across all traits in the periods from birth to 3 years, 3 to 6 years, and 6 to 12 years old (Roberts & DelVecchio, 2000). However, these estimates were not adjusted for measurement error, relied primarily on parent reports of childhood temperament, and did not examine stability from early childhood to early adolescence.

More recent studies have reached similar conclusions about the rank–order stability of temperament from early to middle childhood. Several studies have examined the stability of temperament via latent variables and reported modest (∼.20–.35) to moderate (∼.35–.50) stabilities from early to middle childhood for negative emotionality, positive emotionality, and effortful control (e.g. Carranza et al., 2013; Durbin et al., 2007; Dyson et al., 2015; Komsi et al., 2008; Komsi et al., 2006; Neppl et al., 2010; Rothbart et al., 2001). Most of these studies, too, have similar limitations as those in Roberts and DelVecchio's (2000) meta–analysis in that they rely exclusively on one method such as parent reports (Komsi et al., 2008; Komsi et al., 2006; Neppl et al., 2010; Rothbart et al., 2001) or lab reports (Dyson et al., 2015), do not adjust for measurement error by creating latent variables (Rothbart et al., 2001), and/or examine stability over relatively limited periods of development (Carranza et al., 2013; Durbin et al., 2007; Dyson et al., 2015; Komsi et al., 2008; Komsi et al., 2006; Rothbart et al., 2001).

We are aware of only two studies that have followed a sample from early childhood to early adolescence or beyond. Guerin and Gottfried (1994) found that parent reports of child temperament showed non–significant to moderate stability from age 2 to 12 years. Specifically, examining lower order facets of temperament, they reported test–retest correlations ranging from .00 for adaptability and .09 persistence to .23 (mood) to .30–.40 (activity, approach, intensity, and distractibility). However, they did not incorporate child reports or behavioural assessments or create latent variables to adjust stability estimates for measurement error. Similarly, in what is to our knowledge the longest term study to examine continuity of early childhood temperament, Caspi and Silva (1995) used observers’ ratings of temperament in three–year–old children completing cognitive and motor tasks to cluster children into groups and found significant differences between subgroups on self–reports of positive and negative emotionality and constraint (versus disinhibition) at age 18 years. However, as noted by Caspi and Silva (1995) and Caspi, Henry, McGee, Moffitt, and Silva (1995), it is unclear whether these motor and cognitive tasks adequately elicited the range of behaviours of interest. Moreover, they did not use constructs derived from contemporary models of temperament, making it difficult to compare their findings with the current literature. For instance, their ‘inhibited’ cluster of children showed a range of behaviours that are consistent with low positive emotionality, high negative emotionality, and high effortful control. At the age 18 follow–up, they relied solely on participant reports, whereas parent reports may still have provided useful information about their late adolescent children's temperament. Analyses in both studies also did not adjust for measurement error, and while the long–term follow–up in their study is a significant strength, it is unclear how results would extend from early childhood to preadolescence rather than late adolescence. Finally, neither of these studies integrated standardized lab–based assessments of early childhood temperament with parents’ reports. Thus, knowledge of the rank–order stability of temperament from early childhood to early adolescence is surprisingly limited.

The current study builds upon previous work by integrating, at age 3, standardized lab–based assessments of temperament as well as widely used questionnaires from both parents, both of which are grounded in well–established contemporary models of temperament. Further, we integrated child, mother, and father reports in early adolescence using a widely used, well–validated measure of temperament that was designed to map on to the most widely accepted three–factor models of child temperament.

Integrating multiple sources of information to study the stability of childhood temperament

Parents’ reports are immensely valuable because they draw on observations over extensive periods of time and in a variety of contexts. However, they may be confounded by a variety of reporting biases (Christensen, Margolin, & Sullaway, 1992; Durbin & Wilson, 2012; Jensen, Traylor, Xenakis, & Davis, 1988; Webster–Stratton, 1988; Youngstrom, Izard, & Ackerman, 1999), and stability may be inflated by shared method variance and stability of parental perceptions rather than child behaviour. Another approach to assessing temperament uses lab–based tasks designed to evoke the affects and behaviours characterizing different temperament traits (Goldsmith, Reilly, Lemery, Longley, & Prescott, 1995). However, this approach also has potential limitations, such as concerns about ecological validity, although parents consider their child's responses during laboratory temperament assessments as highly typical of their behaviour outside the laboratory (Lo, Vroman, & Durbin, 2015). There may also be transient influences such as mood states as well as restrictions in the range of affect and behaviour elicited in the child. As parent reports are only modestly associated with observational measures of child temperament, it is likely that both approaches provide unique perspectives (Durbin et al., 2007; Mangelsdorf et al., 2000).

In older youth and adults, temperament is often assessed via self–report. Self–reports can provide critical information about traits often unknown to other informants, such as parents and observers. Moreover, adolescent self–report instruments can map directly onto measures of adult personality and temperament, potentially providing a bridge between assessments of early childhood and adult temperament.

We propose that the most informative and comprehensive approach to understanding the stability of temperament is to integrate lab–based observations with both mother and father reports in early childhood and to examine their association with mother, father, and child reports in early adolescence. Each method provides important and informative but potentially incomplete information regarding children's temperament. This approach will also minimize shared method variance, thereby yielding more accurate estimates of the true rank–order stability of temperament over this span of development. Including both parents’ reports of their child's temperament also mitigates parent–specific biases. Lab–based measures also provide an objective measure, whereas parents often have no comparison or norms by which to answer questionnaires. By measuring temperament via latent constructs at each age, analyses also avoid attenuation of stability estimates because of measurement error. Finally, a latent variable modelling approach also permits the extraction of variance shared between each parent's report and lab–observations at age 3 and mother, father, and child reports at age 12, presumably providing a more valid measure of temperament.

There are, however, potential limitations to this approach. It is assumed in structural equation modelling (SEM) that the non–shared variance that is not included in the latent variable is measurement error, rather than variance in the true temperament construct. However, it is also possible that the non–shared variance provides valid and informative informant–specific or method–specific information. That is, while SEM is widely considered an appropriate way to assess the variance in an underlying construct that is imperfectly measured by several observed constructs, there may be valid and important variance that is lost by this approach. In addition, the present study examined only two time points and, because of the lengthy developmental span, was forced to use different measures at each occasion. Hence, it is limited to examining rank–order stability. Identical measures and three or more observations would be required for examining growth trajectories over time.

Using confirmatory factor analysis (CFA) to construct temperament and personality models with multiple latent factors can also be challenging (Hopwood & Donnellan, 2010). Exploratory factor analyses of multi–factorial personality measures typically yield a good fit to the data as they allow cross–loadings of items and scales across multiple latent factors. However, because of these cross–loadings, CFA approaches to modelling multi–factorial personality measures typically show a poor fit to the data (Hopwood & Donnellan, 2010; Marsh, Hau, & Wen, 2004; Marsh, Scalas, & Nagengast, 2010). More complex models that explicitly incorporate latent methods factors (e.g. Podsakoff, MacKenzie, Lee, & Podsakoff, 2003) often fail to identify or converge. As such, in the current study, we constructed latent variables in separate models for each of the Big Three temperament traits.

Measurement of temperament in early adolescence

Given that the same measures of particular temperament traits cannot be used across widely spaced developmental periods, it is necessary ensure that the different measures used validly assess those constructs. In the current study, at the age 12 assessment, the Schedule for Non–adaptive and Adaptive Personality for Youth (SNAP–Y; Linde, Stringer, Simms, & Clark, 2013) was administered. The adult SNAP and SNAP–Y temperament scales were derived from the General Temperament Survey (Watson & Clark, 1992), which was based on Tellegen's (1985) three–factor model of temperament, and contain the three factors of negative emotionality, positive emotionality, and disinhibition, with the latter mapping onto effortful control (Rothbart & Bates, 2006). Clark and Watson (2008) have shown that the Big Three factors meet all the criteria generally used to define temperament, in that they refer to patterns of affect and behaviour that are relatively stable over time and are largely heritable but also influenced by the environment. Moreover, Rothbart and Bates (2006) and Shiner and Caspi (2003) note the substantial continuity between three–factor models of temperament in children, as examined in the current study, and three–factor models of personality and temperament in adults. These theorists argue that these three temperament dimensions are the foundation for later personality and ultimately differentiate over development in a broader array of personality traits, such as the Big Five (Ready & Clark, 2002; Rothbart & Bates, 2006).

Overview and hypotheses

The goal of this study was to examine the stability of childhood temperamental negative emotionality, positive emotionality, and effortful control/constraint over a nine–year period from age 3 years to age 12. First, we examined zero–order correlations of each trait across different measurement methods and informants. We expected to find the strongest convergence when assessed both within–informant and within–construct. We also expected to observe significant but weaker convergence when the same construct was assessed via parent report but across mothers versus fathers. We expected the lowest, but still significant, convergence between lab–based measures at age 3 and parent reports at age 12. Given the lack of prior research on this topic, we had no a priori hypotheses for whether lab versus parent reports at age 3 would best predict age 12 child reports. Analyses also examined whether age 3 constructs correlated more strongly with their corresponding age 12 counterpart than with non–matching temperament traits (i.e. discriminant validity).

Second, we constructed latent temperament constructs at age 3 based on mother and father reports and lab–based observations. We also constructed latent age 12 temperament constructs based on mother, father, and child reports. We then examined rank–order stability between these latent constructs. We expected these stability coefficients to be moderate in strength but higher than the bivariate test–retest correlations. By incorporating multiple informants and multiple methods and following a large cohort of three–year–old children over a nine–year period, the current study represents the most comprehensive examination of the stability of early childhood temperament of which we are aware.

Open materials and open data statement

Although the data have not been deposited as they are being used in ongoing projects, relevant data, procedures, and syntax for models are available upon request from Dr. Klein.

Participants

Participants were 559 three–year–old children and their parents living in Long Island, New York, who were recruited as a part of a longitudinal study of children's temperament (see Olino, Klein, Dyson, Rose, & Durbin, 2010 for details). The mean age of the children at baseline was 43.5 months (SD = 2.8).

In 2004–2007, participants were recruited through a commercial mailing list and screened by phone. Eligible children had no significant medical problems or developmental disabilities and had at least one English–speaking biological parent who could participate. Most children were European American and non–Hispanic (86.9%), lived with both biological parents (94.6%), and came from middle–class families, as measured by the Four Factor Index of Social Status (M = 45.33, SD = 10.99; Hollingshead, 1975). Families in this study had an average annual household income of approximately $100 000, which is highly comparable with the average household income in the geographic area from which this sample was drawn.

The effective sample size was based on the number of children who completed the Laboratory Temperament Assessment Battery (Lab–TAB) at baseline. Of the 559 families whose children completed the Lab–TAB at baseline, 41 mothers and 158 fathers did not complete parent–report temperament measures. Experimenters also rated participants’ traits across the entire lab visit (‘Experimenter–Impressions’); these scores were unavailable for 22 children. Participants with missing data at baseline did not differ from those with complete data for mothers, fathers, and lab–based observations in terms of sex, socio–economic status, living in a two–parent home, race, ethnicity, or any measure of temperament at age 3 (all ps > .05). Participants were assessed again at age 12 (2014–2016, M age = 151.95 months, SD = 5.53). Of the 559 original participants, 423 children, 426 mothers, and 359 fathers completed questionnaires at the age 12 follow–up. Participants with missing data at the age 12 assessment did not differ from those with complete data at both time points on any variable assessed in this study (all ps > .05). Little's missing completely at random test also confirmed that missingness was not significantly related to any variable in our study: χ² (436) = 482.103, p = .06. Data can thus be viewed as missing at random for analyses. Full information maximum likelihood (FIML) procedures in AMOS 22.0 were used to estimate the means and intercepts in the presence of missing data. This approach is generally acknowledged to be preferable to listwise deletion or mean imputation, which are more likely to yield biased estimates (e.g. Schafer & Graham, 2002). Our effective sample size in latent models was therefore 559, although sample sizes for zero–order correlations varied.

Procedure

At age 3, children participated in the Lab–TAB (Goldsmith et al., 1995). Mothers and fathers completed the Child Behaviour Questionnaire (CBQ; Rothbart et al., 2001). At age 12, mothers, fathers, and children completed the adaptive temperament subscales of the SNAP–Y (Linde et al., 2013). See Table 1 for a summary of the variables, measures, and constructs assessed at each time point.

OPEN IN VIEWER

Table 1

Summary of variables used to measure theoretical constructs

Variable and age	Theoretical construct measured	Sample item
Lab–TAB negative emotionality (age 3)	Negative emotionality
Experimenter Impressions negative emotionality (age 3)	Negative emotionality
Lab–TAB positive emotionality	Positive emotionality
Experimenter Impressions positive emotionality	Positive emotionality
Lab–TAB Impulsivity (age 3, reversed)	Effortful control
Experimenter Impressions effortful control (age 3)	Effortful control
CBQ negative affectivity (age 3)	Negative emotionality	‘Cries sadly when a favourite toy gets lost or broken’
CBQ approach–anticipation (age 3)	Positive emotionality	‘Gets so worked up before an exciting event (s)he has trouble sitting still’
CBQ smiling–laughter (age 3)	Positive emotionality	‘Laughs a lot at jokes and silly happenings’
CBQ effortful control (age 3)	Effortful control	‘Usually rushes into an activity without thinking about it’ (reversed)
SNAP–Y positive temperament (age 12)	Positive emotionality	‘I get excited when I think about the future’
SNAP–Y negative temperament (age 12)	Negative emotionality	‘I often feel nervous and tense’
SNAP–Y disinhibition (age 12)	Effortful control	‘I never buy things on a whim or an impulse’

Note: Lab–TAB, Laboratory Temperament Assessment Battery; SNAP–Y, Schedule for Non–adaptive and Adaptive Personality for Youth; CBQ, Child Behaviour Questionnaire.

Materials

Data analyses

Analyses initially consisted of zero–order correlations between ages 3 and 12 temperament variables that were corrected for attenuation due to unreliability (Muchinsky, 1996), as measured by Cronbach alpha values of internal consistency. As noted by Ferguson (2010) in his meta–analysis, zero–order correlations as measures of the stability of temperament are attenuated by 19–26% because of unreliability. The formula for correcting for attenuation is r_xy/ √ (r_xx^*r_yy) where r_xy is the raw correlation between x and y, r_xx is the Cronbach alpha of x, and r_yy is the Cronbach alpha of y. Uncorrected correlations and their two–tailed p–values are shown in Table 2, with the corrected correlations in Table 3. Adjusting a correlation for unreliability does not alter its p–value. These correlations examined the convergence and divergence of measures of temperament across methods, informants, and variables from ages 3 to 12. Where an age 3 trait correlated with multiple informants’ reports at age 12 on the same trait, or where multiple informants’ reports of a specific age 3 traits correlated with a particular age 12 score, correlations were z–scored and compared in order to test the significance of the difference in the strength of the associations. Significant differences between correlations are shown in Table 3, while results of all comparisons between correlations are shown in Tables S3–S5.

OPEN IN VIEWER

Table 2

Raw correlations between ages 3 and 12 temperament variables

	Age 3	LT–NE	EI–NE	LT–PE	EI–PE	LT–Imp	EI–EC	Mom CBQ NA	Dad CBQ NA	Mom CBQ approach–anticipation	Dad CBQ approach–anticipation	Mom CBQ smiling–laughter	Dad CBQ smiling–laughter	Mom CBQ EC	Dad CBQ EC
Age 12
N		559	537	559	537	559	537	518	401	518	401	518	401	518	401
Mother SNAP NT N = 426		0.07 (.15)	0.16 (.003)	−0.07 (.18)	−0.02 (.72)	0.07 (.20)	−0.09 (.07)	0.22 (<.001)	0.18 (.001)	0.11 (.04)	0.03 (.64)	−0.10 (.06)	−0.02 (.74)	−0.23 (<.001)	−0.16 (.04)
Father SNAP NT N = 359		0.01 (.90)	0.05 (.43)	−0.03 (.53)	−0.01 (.85)	0.07 (.43)	−0.04 (.46)	0.17 (.001)	0.34 (<.001)	0.07 (.23)	0.13 (.03)	−0.06 (.24)	−0.05 (.43)	−0.18 (.001)	−0.25 (<.001)
Child SNAP NT N = 423		0.10 (.04)	0.10 (.08)	0.01 (.83)	−0.01 (.95)	−0.01 (.28)	0.02 (.73)	0.08 (.12)	0.11 (.05)	−0.06 (.27)	0.01 (.94)	0.01 (.90)	−0.06 (.29)	−0.07 (.18)	−0.06 (.26)
Mother SNAP PT N = 426		−0.11 (.02)	−0.07 (.14)	0.11 (.02)	0.1 (.04)	0.01 (.95)	0.04 (.44)	−0.04 (.51)	−0.09 (.10)	0.18 (<.001)	0.06 (.31)	0.12 (.02)	0.08 (.13)	0.24 (<.001)	0.17 (.001)
Father SNAP PT N = 359		−0.03 (.52)	0.02 (.60)	0.12 (.02)	0.05 (.37)	0.07 (.18)	−0.02 (.69)	0.01 (.88)	−0.07 (.20)	0.10 (.08)	0.2 (<.001)	0.01 (.89)	0.19 (.001)	0.07 (.19)	0.24 (<.001)
Child SNAP PT N = 423		−0.02 (.65)	−0.08 (.22)	0.09 (.08)	0.08 (.13)	0.05 (.45)	0.02 (.63)	−0.01 (.85)	−0.09 (.13)	0.09 (.08)	0.05 (.33)	0.04 (.49)	0.13 (.02)	0.11 (.02)	0.15 (.006)
Mother SNAP DIS N = 426		0.07 (.19)	0.01 (.74)	−0.06 (.20)	−0.06 (.13)	0.18 (<.001)	−0.23 (<.001)	0.12 (.02)	0.07 (.21)	0.06 (.19)	0.03 (.52)	−0.14 (.004)	0.01 (.92)	−0.40 (<.001)	−0.25 (<.001)
Father SNAP DIS N = 359		0.05 (.38)	0.08 (.41)	−0.08 (.12)	−0.1 (.06)	0.19 (.002)	−0.2 (<.001)	−0.04 (.61)	0.1 (.07)	−0.01 (.98)	0.01 (.90)	−0.10 (.06)	−0.02 (.71)	−0.33 (<.001)	−0.39 (<.001)
Child SNAP DIS N = 423		0.08 (.10)	0.01 (.82)	−0.09 (.07)	0.08 (.10)	0.16 (<.001)	−0.18 (<.001)	0.02 (.55)	0.03 (.59)	−0.03 (.62)	0.03 (.47)	−0.03 (.56)	−0.02 (.65)	−0.2 (<.001)	−0.18 (.001)

Note: Sample size (N) is included as this varied for each correlation. Values in parentheses under each correlation are p–values. Significant correlations (p < .05) are bolded. LT, Laboratory Temperament Assessment Battery; EI, Experimenter Impressions during the Laboratory Temperament Assessment Battery; NE, negative emotionality; PE, positive emotionality; EC, effortful control; CBQ, Child Behaviour Questionnaire; NA, negative emotionality; SNAP, Schedule for Non–adaptive and Adaptive Personality for Youth; NT, negative temperament; PT, positive temperament; DIS, disinhibition; Mom, mother–rated; Dad, father–rated; Child, child–rated.

OPEN IN VIEWER

Table 3

Unreliability–adjusted correlations between ages 3 and 12 temperament variables

	Age 3	LT–NE	EI–NE	LT–PE	EI–PE	LT–Imp	EI–EC	Mom CBQ NA	Dad CBQ NA	Mom CBQ approach–anticipation	Dad CBQ approach–anticipation	Mom CBQ smiling–laughter	Dad CBQ smiling–laughter	Mom CBQ EC	Dad CBQ EC
Age 12
N		559	537	559	537	559	537	518	401	518	401	518	401	518	401
Mother SNAP NT N = 426		.08	.19 **	−.08	−.02	−.09	−.11 +	.25 **	.21 **	.14 *	.04	−.12 +	−.02	−.26 **	−.18 **
Father SNAP NT N = 359		.01	.06	−.04	−.01	−.09	−.05	.19 **	.39 **	.09	.16 *	−.07	−.06	−.20 **	−.29 **
Child SNAP NT N = 423		.12 *	.12 +	.01	−.01	.01	.02	.09	.12 *	−.08	.00	.01	−.07	−.08	−.07
Mother SNAP PT N = 426		−.13 *	−.09	.13 *	.11 *	.01	.05	−.05	−.10	.23 **	.08	.15 **	.10	.28 *	.20 **
Father SNAP PT N = 359		−.04	.02	.14 *	.06	−.09	−.02	.01	−.08	.12 +	.26 **	.01	.24 *	.08	.28 **
Child SNAP PT N = 423		−.02	−.10	.10 +	.09	−.07	.03	−.01	−.11	.11 +	.07	.05	.17 *	.13 *	.18 **
Mother SNAP DIS N = 426		.08	.01	−.07	−.07	−.23 **	−.29 **	.13 *	.08	.08	.04	−.18 **	.01	−.46 **	−.30 **
Father SNAP DIS N = 359		.06	.10	−.10	−.12 +	−.25 **	−.25 **	−.05	.12 +	−.01	.01	−.14 +	−.03	−.38 **	−.46 **
Child SNAP DIS N = 423		.10	.01	−.11 +	.09	−.21 **	−.23 **	.02	.04	−.04	.04	−.04	−.03	−.24 **	−.22 **

Note: Sample size (N) is included as this varied for each correlation. See Table 2 for exact p–values. LT, Laboratory Temperament Assessment Battery; EI, Experimenter Impressions during the Laboratory Temperament Assessment Battery; NE, negative emotionality; PE, positive emotionality; EC, effortful control; CBQ, Child Behaviour Questionnaire; NA, negative emotionality; SNAP, Schedule for Non–adaptive and Adaptive Personality for Youth; NT, negative temperament; PT, positive temperament; DIS, disinhibition; Mom, mother–rated; Dad, father–rated; Child, child–rated.

**

p < .01.

*

p < .05.

+

p < .10.

Structural Equation Modelling was carried out using AMOS 22.0. We examined the rank–order stability between temperament at age 3 measured via a latent variable indicated by mother and father CBQ reports, and Lab–TAB and experimenter impression scores, and at age 12 via a latent variable indicated by mother, father, and child reports on the SNAP–Y. Thus, ages 3 and 12 temperament were modelled simultaneously. At age 3, latent negative emotionality was indicated by mother and father reports of CBQ negative affectivity as well as negative emotionality from the Lab–TAB and Experimenter Impressions. Latent positive emotionality was indicated by mother and father reports of smiling–laughter and approach–anticipation on the CBQ, and Lab–TAB and Experimenter Impressions positive emotionality. Effortful control was indicated by mother and father reports of effortful control on the CBQ with Lab–TAB impulsivity (reversed) and Experimenter Impressions effortful control/constraint. At age 12, latent variables of negative emotionality, positive emotionality, and disinhibition were each composed of mother, father, and child reports on the relevant manifest indicator. Given that ratings from Experimenter Impressions and the Lab–TAB were based on overlapping samples of behaviour, error terms on Lab–TAB and Tester–Impressions variables were covaried within each construct a priori in each model.

Ideally, the interrelationships between all three latent factors at both ages 3 and 12 should be examined in the same model. However, consistent with prior findings that CFA approaches to modelling the structure of personality or temperament typically show a poor fit to the data (Hopwood & Donnellan, 2010; Marsh et al., 2004; Marsh et al., 2010), a three–factor model of temperament at age 3 including covariances on error terms within each construct of lab–based variables showed a poor fit to the data, χ² (71) = 694.70, p < .001, comparative fit index (CFI) = .70, root mean square error of approximation (RMSEA) = .13, 90% confidence interval (CI) [0.12, 0.13].

We also attempted to model all three age 3 temperament traits in one model following Podsakoff's proposed multi–trait multi–method CFA (see tables 4 and 5 in Podsakoff et al., 2003). Latent variables were indicated by separate mother, father, and lab–based scores, thereby portioning variance attributable to the informant or method into separate latent factors, while retaining separate latent factors for negative emotionality, positive emotionality, and effortful control (Figure S1). We also computed a similar model but with only two latent methods factors, one for lab–based variables and another for all parent reports (Figure S2). However, consistent with the Podsakoff et al. (2003) observation that the most frequent problem with these models is that they do not identify, neither model identified. Others have also noted frequent convergence problems with these models (Kenny & Kashy, 1992).

We also computed Podsakoff's correlated uniqueness model with all three traits at age 3 (table 4 in Podsakoff et al., 2003). This model accounts for methods effects by allowing error terms of specific measurement methods to be correlated. Thus, covariances were included between all parent–report variable error terms and between all lab–measured variable error terms (Figure S2). Again consistent with the Podsakoff et al. (2003) point that these models often have identification problems, this age 3 model failed to identify. A three–factor model of SNAP–Y scales at age 12 similarly showed a poor fit to the data, χ² (24) = 176.23, p < .001, CFI = .84, RMSEA = .11, 95% CI [0.09, 0.12]. Given that we were unable to model all three latent factors in one model at age 3, we did not compute similar models with age 12 variables. Therefore, we examined the stability of each trait in separate models.

In order to examine heterotypic continuity, we imputed factor score estimates from the model for each of the latent variables described earlier based on the factor loadings from our models that contained only one age 3 and one age 12 latent variable. That is, negative emotionality at ages 3 and 12, positive emotionality at ages 3 and 12, and effortful control at ages 3 and 12 were imputed separately, and each was used as a manifest (observed) variable in our models examining heterotypic continuity. AMOS uses regression imputation with FIML for estimating missing data. This is identical to applying the factor score weights to each individual's scores on the observed indicators. Estabrook and Neale (2013) note that when data are missing at random, FIML more accurately estimates individual factor scores with missing data compared with sum scores, mean imputation, or regression estimators that do not use FIML. We then tested a path model in which each age 12 latent temperament variable (i.e. negative and positive emotionality and disinhibition) was simultaneously regressed on all three age 3 latent temperament variables (i.e. negative emotionality and positive emotionality and effortful control). Thus, this analysis examined whether each age 3 trait predicted each age 12 trait, adjusting for that trait's respective age 3 levels.

We are not aware of any studies that have used multiple methods and informants and created factor scores estimates of the latent variables based on these observed indicators in separate models. However, it is routine for constructs to be created based on averages of either parent reports or behavioural episodes, and then use those to examine heterotypic stabilities of over time (e.g. Majdandžić & Van Den Boom, 2007; Pesonen, Räikkönen, Keskivaara, & Keltikangas–Järvinen, 2003; Putnam, Rothbart, & Gartstein, 2008). Komsi et al. (2006) examined heterotypic continuity of latent temperament constructs, but these constructs did not include observational measures as indicators, and so they were able to fit multi–factorial temperament models. As such, we are not aware of any studies that have taken an approach directly comparable with ours.

As measures of goodness of fit, we present chi–square, CFI, and RMSEA as well as the 90% CI around the RMSEA. Generally, CFI values greater than .90 (Hoyle & Panter, 1995) and an RMSEA less than .08 (Kline, 1998) indicate acceptable fit. Given that we expect a medium effect size in our latent stability models both in terms of loadings of observed indicators on latent variables as well as the relationships between ages 3 and 12 temperament traits, power analyses revealed that with two latent variables and nine indicators, which is the upper number of indicators in our models, a minimum sample of 90 would be required (Soper, 2017; Westland, 2010). Regression weights in our latent models are fully standardized and therefore represent effect sizes.

Zero–order convergent and discriminant correlations

Descriptive statistics and within–time correlations are available in Tables S1 and S2. Zero–order correlations between age 3 lab–based and parent–reported temperament traits and age 12 mother, father, and child–reported traits are shown in Table 2, and associations unadjusted for unreliability are presented in Table 3.

Negative emotionality correlations

Consistent with hypotheses, correlations were generally strongest between parent reports of negative emotionality at ages 3 and 12 compared with correlations of age 3 lab–based negative emotionality with age 12 negative emotionality, with some evidence of stronger correlations within compared with across informants (Tables 4, and S3). Also consistent with expectations, most correlations were modest in strength. Compared with parent reports of negative emotionality at age 3, Lab–TAB and Experimenter Impressions negative emotionality converged as well or better with child–reported negative emotionality at age 12. Lab–TAB negative emotionality also correlated negatively with mother–reported age 12 positive emotionality, while mother–reported age 3 negative emotionality correlated positively with mother–rated age 12 disinhibition.

OPEN IN VIEWER

Table 4

Significant r–to–z zero–order correlation comparisons

Correlation 1	Correlation 2	z	p
Negative emotionality
Father CBQ NA—Father SNAP–Y NT	Father CBQ NA—Mother SNAP–Y NT	2.45	.01
Father CBQ NA—Father SNAP–Y NT	Father CBQ NA—Child SNAP–Y NT	3.59	<.001
Effortful control
Mother CBQ EC—Mother SNAP–Y Dis	Mother CBQ EC—Child SNAP–Y Dis	−3.55	<.001
Mother CBQ EC—Father SNAP–Y Dis	Mother CBQ EC—Child SNAP–Y Dis	−2.09	.04
Father CBQ EC—Father SNAP–Y Dis	Father CBQ EC—Mother SNAP–Y Dis	−2.32	.02
Father CBQ EC—Father SNAP–Y Dis	Father CBQ EC—Child SNAP–Y Dis	−3.37	<.001
Mother CBQ EC—Mother SNAP–Y Dis	LT–Imp—Mother SNAP–Y Dis	−.3.73	<.001
Mother CBQ EC—Mother SNAP–Y Dis	LT–Imp—Father SNAP–Y Dis	−3.28	.001
Father CBQ EC—Father SNAP–Y Dis	LT–Imp—Father SNAP–Y Dis	−3.03	.002
Mother CBQ EC—Mother SNAP–Y Dis	EI EC—Mother SNAP–Y Dis	−2.82	.005
Father CBQ EC—Father SNAP–Y Dis	EI–EC—Father SNAP–Y Dis	−3.03	.002

Note: Each column indicates which correlation was used. Bolded correlation indicates which correlation was significantly larger. Positive emotionality correlation comparisons not listed here because none were significant. See Table 3 for correlations used to compute values in this table. LT, Laboratory Temperament Assessment Battery; EC, effortful control; Imp, impulsivity; EI, Experimenter Impressions; CBQ, Child Behaviour Questionnaire; NA, negative affectivity; SNAP–Y, Schedule for Non–adaptive and Adaptive Personality for Youth; NT, negative temperament; DIS, disinhibition; Mother, mother–rated; Father, father–rated; Child, child–rated.

Stability of latent constructs

Negative emotionality

Our model of latent age 3 negative emotionality predicting latent age 12 negative emotionality (Figure 1, top panel) showed a good fit to the data: χ² (12) = 33.97, p = .001, CFI = .96, RMSEA = .054, 90% CI [0.034, 0.079]. All loadings of indicators on age 3 and age 12 negative emotionality were at least .20 and were significant at p < .001, although CBQ variables showed higher loadings than Lab–TAB or Experimenter Impressions negative emotionality. The standardized stability of age 3 to age 12 negative emotionality was .41 (p = .003), indicating moderate stability.

OPEN IN VIEWER

Figure 1

^**p < .01. Factor loadings of observed indicators and stability of temperament from ages 3 to 12 years. All parameters are standardized estimates. Covariances on observed variable are covariances on the error terms of those variables. Error term on endogenous variables not depicted. Lab–TAB, Laboratory Temperament Assessment Battery; EC, effortful control; Imp, impulsivity; EI, Experimenter Impressions; CBQ, Child Behaviour Questionnaire; NA, negative affectivity; EC, effortful control; SNAP, Schedule for Non–adaptive and Adaptive Personality for Youth; NT, negative temperament; DIS, disinhibition; PT, positive temperament; Mother, mother–rated; Father, father–rated; Child, child–rated.

Heterotypic continuity

Using the imputed ages 3 and 12 temperament scores from our latent models, we examined whether each age 3 temperament trait predicted each discriminant age 12 temperament trait after adjusting for each age 12 trait's respective level at age 3 (e.g. age 3 negative emotionality predicting age 12 disinhibition adjusting for age 3 effortful control). Given that our model examining heterotypic continuity was fully saturated, fit indices are perfect and are therefore not reported. In this model (Figure 2), in addition to significant homotypic stability coefficients, greater age 3 effortful control predicted increases in positive emotionality at age 12 and decreases in negative emotionality at age 12. Including heterotypic paths substantially attenuated the homotypic paths.

OPEN IN VIEWER

Figure 2

Relationships between imputed latent ages 3 and 12 temperament traits. ^**p < .01, ^*p < .05. Note: Variables are rectangles as they were used as observed variables in the model but were derived by extracting latent scores from the latent models in Figure 1. All parameters are standardized estimates. Parameters on curved arrows are correlations. Curved double–headed arrows on endogenous variables are correlations on the error terms of those variables.

Zero–order stability and specificity of early to late childhood temperament

Mothers’ and fathers’ reports of child negative emotionality, positive emotionality, and effortful control at age 3 typically showed stronger convergence with their own report of the same trait at age 12 relative to the other parent's report. This is consistent with Neppl et al. (2010) who found stronger test–retest correlations within informants (mothers versus fathers) than across. Overall, results are consistent with the possibility that shared method variance, both in terms of method (questionnaires) and informant, increases estimates of stability. This is significant because most studies of the stability of temperament use the same measures and informants at each assessment.

Associations between age 3 parent–reported negative and positive emotionality with their respective age 12 child–reported counterparts showed some unexpected patterns. Father, but not mother, reports of negative emotionality at age 3 correlated with child–reported negative emotionality at age 12. Across mother and father reports of smiling–laughter and approach–anticipation, only father–reported smiling–laughter correlated with child reports of age 12 positive emotionality. Many would likely assume that a mother might have more detailed and accurate information regarding their young child's negative and positive emotionality. However, if child reports at age 12 are taken as an important marker of their age 12 temperament, this result suggests that fathers’ reports in early childhood provide better information in terms of their correlation with child reports in early adolescence.

The possibility that fathers have more detailed or accurate information compared with mothers regarding their young child's temperament, at least in terms of their convergence with children's reports in early adolescence, requires replication in future research. If verified, this may suggest that mothers should not be relied upon as the sole source of information regarding their child's temperament. These results may also suggest that fathers’ reports of temperament show greater utility compared with mothers’ reports in terms of their convergent validity with children's later reports. The reasons that fathers’ reports showed greater convergent validity than mothers’ with children's later self–reports cannot be gleaned from the current study. We speculate that fathers are more observant of, and attentive to, their young children's emotional expressions, which would be contrary to the oft–expressed position that mothers provide more accurate reports of child behaviour. Alternatively, fathers may be more likely to notice or attend to more extremes in children's emotional behaviour, which could lead to better discrimination between more and less emotional children. If so, this could be driving the greater convergence between early childhood father reports and early adolescent children's reports. Regardless, consistent with prior research highlighting that different informants provide unique sources of information (e.g. Achenbach, McConaughy, & Howell, 1987), these results underscore the importance of incorporating fathers’ reports in the comprehensive measurement children's temperament, whether for research or practical/clinical purposes.

Mother–reported and father–reported age 3 effortful control correlated with age 12 child reports of disinhibition, although more weakly than with parents’ reports of disinhibition at age 12. Experimenter Impressions and Lab–TAB positive emotionality as well as Experimenter Impressions negative emotionality were not significantly associated with child reports of their corresponding constructs at age 12, although Lab–TAB negative emotionality and both lab–based observations of effortful control were. These results suggest that lab–based measures, in particular those measuring effortful control and to some extent negative emotionality, tap the early antecedents of child reports of temperament in early adolescence.

It is possible that the behaviours and affects associated with negative emotionality (e.g. sadness and anxiety) and effortful control (e.g. actively restraining an impulsive behaviour) are experienced internally to a greater degree than they are expressed externally by the child. Lab–based tasks may elicit anxiety (e.g. a stranger approaching), sadness (e.g. not receiving an expected gift), or effortful control (e.g. being forced to wait while building a tower) that may be indicative of these specific behaviours in naturalistic settings and thus better converge with child reports early adolescence. As such, results further suggest the importance of including lab–based measures of child temperament in addition to parental reports.

Parent–rated age 3 temperament constructs often showed poor discriminant validity with age 12 temperament. This is consistent with some prior research that has used parent reports (e.g. Neppl et al., 2010). In the current study, father–reported age 3 negative emotionality showed the best convergent and discriminant validity, in that it correlated with mother, father, and child reports of negative emotionality at age 12 but not with any other age 12 variable. Mother reports of age 3 negative emotionality, however, correlated not only with mother and father age 12 negative emotionality but also with mother reports of disinhibition at age 12. While parent–reported CBQ smiling–laughter and approach–anticipation generally showed some specificity with age 12 positive emotionality, mother reports of smiling–laughter also correlated with mother–reported disinhibition at age 12. Parental reports of effortful control showed the poorest discriminant validity, in that they both correlated with parents’ reports of negative emotionality and positive emotionality at age 12, in addition to correlating with disinhibition.

Lab–TAB and Experimenter Impressions temperament variables, on the whole, showed good discriminant validity. Lab–TAB positive emotionality was uniquely related to father and mother reports of age 12 positive emotionality but was not significantly related to negative emotionality or disinhibition as rated by any informant. Similarly, Experimenter Impressions positive emotionality was uniquely related to age 12 mother–rated positive emotionality but was unrelated to any other age 12 variable. Lab–TAB impulsivity and Experimenter Impressions effortful control were uniquely related to mother, father, and child reports of disinhibition and were unrelated to any informant's report of positive or negative emotionality. Finally, Lab–TAB negative emotionality predicted child–reported age 12 negative emotionality, and Experimenter Impressions negative emotionality was specifically related to mother–reported age 12 negative emotionality. The one exception to the pattern of high discriminant validity for lab–based variables was that Lab–TAB negative emotionality at age 3 predicted lower levels of mother–rated positive emotionality at age 12.

Taken together, lab–based measures may show better discriminant validity in terms of their associations with age 12 temperament relative to parent reports of temperament in early childhood. One possibility is that parents may have somewhat misinterpreted the intent of the questions or their child's behaviour in everyday life. Similarly, the divergent correlations observed for parent reports may reflect shared method variance. Finally, we cannot rule out that what may appear to be a lack of discriminant validity is really greater sensitivity to true heterotypic continuity relative to lab–based measures. However, this would require systematic evidence across a variety of measures that particular traits predict other traits later in development. Regardless, these results further highlight the utility of incorporating lab–based measurements into the assessment of temperament in young children.

Stability of latent temperament constructs

Results from SEMs indicated that lab–based observations and mother and father reports of temperament in early childhood were viable manifest indicators of latent temperament constructs, although Lab–TAB negative emotionality and impulsivity and Experimenter Impressions negative emotionality and effortful control contributed relatively lower amounts of variance to their respective latent factors than the CBQ variables. Results suggest that assessing temperament via both parents and lab–based observations may yield a more comprehensive and nuanced measure than using only one of these methods.

Across all three temperament traits, SEMs indicated that temperament is moderately stable from ages 3 to 12 years, although effortful control showed somewhat greater stability than did negative or positive emotionality. Our results are consistent with Caspi and Silva (1995) who reported significant associations between temperament at age 3 and personality at age 18, as well as Guerin and Gottfried (1994) who reported moderate stability of maternal reports of children's mood (r = .23), distractibility (r = .37), and approach (r = .42) from ages 3 to 12.

Our stability estimates, based on latent variable modelling over a period of almost a decade, were comparable with or higher than those reported by Roberts and DelVecchio (2000), despite it being reasonable to expect higher stability estimates over shorter time periods. That suggests that if one uses multiple sources of information to reduce measurement error, there is a moderate to substantial degree of continuity in temperament from early childhood to early adolescence, even though this period is believed to be characterized by greater flux in individual traits than subsequent developmental periods (Posner & Rothbart, 2007; Rothbart & Bates, 2006). Interestingly, latent stability coefficients were higher than the highest within–informant correlations. This suggests that random error may have reduced the zero–order correlations more than informant/method variance inflated it. We are not aware of studies that have attempted to quantify the relative effects of informant/method variance versus random error variance on the associations between variables.

Taken together, these results provide strong evidence for homotypic continuity as well as moderate rank–order stability in temperament. These findings speak to the fundamental nature of temperament in children. Temperament is arguably at the core of human psychological identity; our results suggest that this identity begins to solidify over the course of childhood while also remaining malleable. These findings support the compromise perspective of temperamental stability and suggest that temperament should be conceptualized as only moderate stable, at least across childhood. This is consistent with the possibility that there are important environmental influences, such as parenting, culture (Ferguson, 2010), and non–shared environmental factors (Saudino, 2005), that affect the development of temperament. At the same time, the results suggest that early child temperament, at least when measured via multiple methods and informants, provides a marker with modest predictive utility of future temperament styles up to 9–10 years later.

Finally, the results from our path model also showed some evidence of heterotypic continuity. Specifically, greater effortful control predicted lower negative emotionality and higher positive emotionality at age 12. This is consistent with some previous research that found evidence of heterotypic continuity of temperament in early childhood. For instance, Putnam et al. (2008) infants high in surgency showed greater effortful control in toddlerhood, high levels of surgency in toddlerhood predicted lower effortful control in preschool, and toddler negative affect predicted lower effortful control in preschool (Putnam et al., 2008). Evidence also suggests that effortful control predicts lower levels of negative emotionality in later childhood (Kochanska & Knaack, 2003). Similarly, Dyson et al. (2015) also found that age 3 constraint, similar to effortful control, predicted lower age 6 fearfulness. The reasons that early childhood effortful control influences later negative and positive emotionality in early adolescence cannot be gleaned from the current study. However, this is consistent with prior evidence that effortful control is associated with more effective regulation of emotional reactivity and more generally with positive outcomes in later life (e.g. Eisenberg et al., 2009; Moffitt et al., 2011). Results suggest the need to search for potential mechanisms underlying homotypic and heterotypic continuity of temperament, such as specific genetic and environmental influences, gene–environment interactions, and gene–environment correlations (Caspi & Roberts, 2001; Plomin, Caspi, Pervin, & John, 1999).

Our results also highlight the importance of considering multiple sources of information regarding children's temperament in addition to parent reports. More generally, the results suggest that researchers should carefully consider the methods they use to assess temperament, as well as how they interpret their results based on those methods.

Limitations and future directions

This study had several notable strengths. It comprised a large sample of children who were followed for almost a decade and used fine–grained, standardized, lab–based measures of temperament along with both mother and father reports in early childhood, and mother, father, and child reports of temperament in early adolescence. However, some limitations should be noted. First, scores on the Lab–TAB and Experimenter Impressions variables may be influenced by situation–specific behaviours or affects. However, prior research suggests that the stability of lab–based observations of negative and positive emotionality is almost comparable with parent reports (Durbin et al., 2007).

Second, we were unable to assess the interrater reliability of the Experimenter Impressions measure. However, the integration of these variables with Lab–TAB ratings and parent reports in our latent models attenuates this concerns. Experimenter Impressions ratings were also derived from observations that overlapped with the Lab–TAB, which showed good interrater reliability and were moderately highly correlated with the Lab–TAB (Table S1).

Third, we also could not use lab–based measures at age 12, given that there are no validated observational measures of temperament in older youth. This raises the concern of whether stability estimates, even those derived from latent variables, may underestimate stability.

Fourth, we only examined rank–order stability. Future research should also examine mean level stability and individual trajectories over time in temperament. However, this will be challenging, as these questions require the same measures to be administered at each time point, which is difficult given considerations of developmental appropriateness. The lack of similarity of indicators in our latent models also precludes examining measurement invariance over time. This is a fundamental challenge to the field, as measures need to be developmentally appropriate while also still measuring the appropriate construct. This typically precludes using identical measures when developmental samples are followed across different development periods, as in the current study.

Fifth, although not a limitation per se, we did not examine predictors of change in temperament over time or moderators of the stability of temperament. While we hope to examine these issues in future research, prior to examining influences on stability, it is important to first understand normative stability. Sixth, future research would also benefit from examining stability of lower order temperament traits.

Seventh, the participants in our sample were predominantly White/European American and middle class and recruited through a commercial mailing list. Although the sample was demographically representative of the population in our geographic region (Suffolk County, Long Island, New York), this may constrain the generalizability of our findings.

Finally, were unable to test heterotypic continuity in our full latent models and instead had to circumvent this by imputing latent scores and using them as observed variables. It is unclear whether results would change using a multi–factorial confirmatory model. For instance, by estimating factors scores based on separate models for each construct, this may bias results in favour of finding homotypic paths as estimated scores were based on homotypic latent models that examined only one trait each.