Longitudinal Development of Executive Function from Infancy to Late Childhood

Abstract

Executive function (EF) is difficult to measure in young children because of its heterotypic profile and the rapid development that occurs from infancy through late childhood. To examine EF development, we created age-specific latent factors of EF at ages 5 months, 10 months, 24 months, 36 months, 48 months, and 9 years and used structural equation modeling to create an autoregressive model from infancy through late childhood. Although EF remains difficult to measure due to task constraints, toddler EF as measured by these behavioral tasks is a relatively stable predictor of later EF. Understanding this construct and its developmental trajectory is necessary for creating appropriate experimental designs and interventions to address EF in early childhood.

Executive Functions (EF) are a set of effortful cognitive processes that are utilized for top-down control of mental or physical actions ( Diamond, 2013). EF is necessary for decision making and problem solving and allows for higher order thinking to plan and execute complex actions and cognition. Childhood EF is associated with academic achievement (Zelazo & Carlson, 2020), as well as other important life outcomes such as health, financial stability, and wellbeing (e.g., Brown & Landgraf, 2010; Drever et al., 2015; Gathercole et al., 2004; Moffitt et al., 2011). Understanding the development and structure of EF in childhood is vital to allow for better measurement and potential intervention on this critical skill. The purpose of this study is to explore the structure and stability of EF using behavioral tasks from infancy to late childhood.

EF is well studied in adulthood and consists of three main core cognitive processes: inhibitory control (IC), updating or working memory (WM), and set-shifting or attentional flexibility (Lehto, 1996). Confirmatory factor analysis of nine commonly used EF tasks in typical adults has provided evidence that the three subconstructs, though correlated, are discrete in adulthood and can be measured separately by different tasks (Miyake et al., 2000). These core processes interact and work with each other to allow for complex planning and reasoning abilities.

IC is the ability to suppress a prepotent or dominant attentional, cognitive, behavioral, or emotional response (Diamond, 2013). The inhibition of these responses allows an individual to consciously and effortfully refrain from habitual or automatic responses and instead choose a less dominant response. IC is associated with self-control and behaving in a manner counter to temptation or impulse (Hofmann et al., 2012). WM is the maintenance and manipulation of information (Baddeley, 1992; Baddeley & Hitch, 1974). Without WM, it is impossible to relate concepts or experiences with current events or reorganize thoughts for goal-directed behavior (Baddeley, 1996, 2000). WM is vital for the perception of both written and verbal language (Daneman & Merikle, 1996). WM also contains the buffer between long term memory and EF, which allows for previous experiences to be brought into consideration when making decisions or planning (Baddeley, 2000). Importantly, WM allows for creativity by manipulating information into novel ideas or goals and allows an individual to maintain information even when their attention has shifted (Takeuchi et al., 2011; Vandervert et al., 2007). Shifting is the capacity to switch between mental tasks and rules (Monsell et al., 2000). Cognitive flexibility is a broader term that includes shifting perspectives as well as shifting tasks or rules (Stemme et al., 2007). Poor shifting leads to perseverating on certain concepts or previous ways of completing a task and not being able to consider alternative perspectives (H.L. Miller et al., 2015; Stuss et al., 2000).

The core EF measures build upon one another (Garon et al., 2008). Complex IC tasks require greater WM relative to simple IC tasks. Inhibition of previous rules and maintenance of the current goal in WM are essential for appropriate shifting and switching to a new set of rules. This demonstrates the interdependency of these constructs, they can be theoretically separated but show unitary function (Miyake & Friedman, 2012).

Although these subtypes of EF are statistically correlated and conceptually interrelated, their discrete existence is well-validated in adult studies using confirmatory factor analysis (e.g., Miyake et al., 2000). In childhood, however, it is more difficult to quantify IC, WM, and shifting as separate constructs, leading some to propose that there is a single unitary EF factor in childhood that separates into discrete factors later in development. Specifically, EF during early childhood may not be best explained by a three-factor model as Miyake and colleagues (2000) found with adults and other researchers have found with older children and adolescents (Lee et al., 2013; Lehto et al., 2003). Instead, work with children from 3 to 6 years old suggests that a one factor model can adequately explain the structure of EF in this age group (Wiebe et al., 2008; Wiebe et al., 2011). On the contrary, there is also empirical evidence that a two factor model with WM and IC can best fit the data, if more tasks are included and age ranges are restricted (Howard et al., 2015; M.R. Miller et al., 2012). This uncertainty may reflect hierarchical development of EF, wherein basic abilities are required before children can complete complex tasks and changes in the structure of EF over development (Devine et al., 2019). Theoretically, IC tasks contain characteristics that are conceptually different from those that measure WM as well as those that measure shifting, although there is no pure measure of any EF (Garon et al., 2008; Miyake & Friedman, 2012). Therefore, when comparing development across ages it may be prudent to take a single factor approach to EF and use structural equation modeling to investigate tasks that load onto a latent factor representing EF at each age.

Although there is uncertainty over the structure of EF in early childhood, the development of EF is believed to begin in the first year and improve through adolescence (Ahmed et al., 2019; Blankenship et al., 2019; Wiebe et al., 2008, Wiebe et al., 2011). Researchers label EF tasks for infants and young children as targeting specific EF constructs (Garon et al., 2008; Petersen et al., 2016). For example, infants as young as 5 to 8 months maintain mental representations over a short delay, an early measure of working memory (Bell, 2012; Pelphrey et al., 2004), and early ability to inhibit a dominant response is observed to develop between 8 and 12 months (Cuevas et al., 2012; Kochanska et al., 1998). This rudimentary EF improves from the first year through early childhood, with great development occurring between 3 and 5 years (Carlson, 2005; Diamond, 2013; Garon et al., 2008). EF continues to improve throughout childhood and adult-like performance is reached on EF tasks by adolescence (e.g., Luciana et al., 2005; Luna et al., 2004; Williams et al., 1999).

The behavioral changes in EF correspond with neural maturation of the prefrontal cortex (Roberts & Pennington, 1996). The prefrontal cortex develops slowly compared with other brain areas and does not fully mature until early adulthood (Huttenlocher, 1979). Developmental changes of the prefrontal cortex coincide with improvements in EF (Diamond, 2002) as grey matter increases volumetrically in early development until around age 4 and then begins to undergo selective pruning, resulting in a reduction of gray matter (Huttenlocher, 1984). These developmental changes in the structure and connectivity of the prefrontal cortex are important for the development of higher order processing skills, and they are associated with improvements in EF (Welsh & Pennington, 1988). Thus, the neural development that occurs in early childhood may constrain the age at which EF becomes stable, with evidence suggesting changes in PFC development that appear around age 4 (Fiske & Holmboe, 2019). Individual differences in prefrontal cortex maturity have been proposed to underlie observed differences in behavioral performance on EF tasks (McKenna et al., 2017; Tamm et al., 2002). As the structure and foundation of EF develops, so too do the behavioral tasks used to measure the construct.

Although tasks used to evaluate EF in young children are more limited than those used in adults, there are well-validated and reliable tasks available (e.g., Carlson, 2005; Garon et al., 2008; Petersen et al., 2016). Many of these tasks take adult measures and manipulate them to be appropriate for young children. For example, in children, Stroop-like tasks induce the Stroop error in children that cannot read by asking them to inhibit a prepotent response to an image, such as the day/night task, or gesture, such as the yes/no task. Other tasks that measure IC in young children include delay-of-gratification tasks, similar to the classic marshmallow test (Mischel et al., 1989). However, there are very few infant tasks that measure EF. One such task that is appropriate is the A-not-B task, which employs the classic Piagetian A-not-B error to measure how well infants can maintain a goal in WM and inhibit a previously reinforced response (Bell, 2001). As children develop, tasks become more similar to those used with adults and by late childhood children are able to complete more complex Stroop tasks. WM and shifting tasks show a similar trajectory with early childhood tasks such as the Dimensional Change Card Sort (DCCS) employing a sorting requirement similar to the Wisconsin Card Sort Task, but with more explicit rules and scaffolded directions (Bell & Garcia Meza, 2020). It is important to note, however, that all of these tasks in early childhood rely on verbal comprehension to understand the expectations of the tasks. Thus, individual differences in verbal ability conflate these measures of EF and language and EF are intertwined throughout development (Kuhn et al., 2014).

Despite infants engaging in EF tasks and showing individual differences in ability within the first year of life, there is limited evidence showing stability in these measures of EF from infancy into childhood. There is evidence that measures of attention and information processing in early infancy are associated with EF in later infancy and childhood (Blankenship et al., 2018; Cuevas & Bell, 2014; Devine et al., 2019; Kraybill et al., 2019), but it is unclear if infant EF tasks that measure aspects of WM, shifting, and IC predict individual differences in EF beyond infancy. Even in toddlerhood and early childhood, there is confusion about the longitudinal stability of EF. For example, when examining separate tasks S.E. Miller and Marcovitch (2015) report no correlation in A-not-B performance between 14 and 18 months. Similarly, Johansson and colleagues (2016) report no correlation between 12 month structured EF tasks and 24 month or 36 month EF, with the caveat that 12-month hide and seek was negatively correlated with a Stroop-like task at 36-months. By 24-months, Carlson and colleagues report that an EF composite at 24-month is significantly correlated with a composite at 39 months, suggesting that EF may become self-predictive and demonstrate some degree of stability by the end of the second year. However, longitudinal data is necessary to answer questions about the foundations of EF in infancy and how it emerges across childhood (Hendry et al., 2016).

We aimed to elucidate how the structure, or types of tasks, used to measure EF changes from early infancy to late childhood. By using many behavioral measures of EF, we demonstrate which tasks load together to create a latent factor of EF at multiple time points and how these age-based factors are stable across development. Previous work with this sample showed continuity of EF composites (not latent factors) from 10 months to 48 months in the context of predicting academic achievement at age 6. This previous work did not include the 24-month or 9-year time points because of the focus on infant attention, early EF, and age 6 outcomes, as well as a funding schedule that only allowed for data collection of 75% of the participants at the age 6 timepoint (Blankenship et al., 2019). We suggest that EF begins to show stability in early childhood, but were interested to see at what age stability in EF could be observed in a longitudinal sample using multiple behavioral tasks. Based on previous EF work in early childhood suggesting a single EF factor, we took an exploratory approach to create a single latent factor of EF at each age using behavioral measures and examine the stability of EF across childhood.

Method

Participants

Four hundred and ten infants (201 boys) were recruited to participate in a longitudinal study investigating cognition and emotion at two locations in the mid-Atlantic region of the United States. The cohorts were broadly recruited by two research labs (Virginia Tech-A, University of North Carolina Greensboro-B) when the children were infants using mailing lists, media advertisements, flyers, and word of mouth. The A lab and the B lab each recruited half of the participants in the longitudinal study and data were collected between 2002 and 2017.

The original sample was recruited as three cohorts, with children and parents visiting the laboratory at 5 months, 10 months, 24 months, 36 months, 48 months, and 9 years of age. The sample size for the larger longitudinal study was planned with sufficient power to detect medium to large (f² = .15 to .35) effect sizes in a structural equation model (Cohen, 1988; Gignac & Szodorai, 2016). Two of the three cohorts of children also completed an age 6 visit, but because that time point included only 75% of the sample due to the funding schedule, the age 6 time point is not included in the current analyses. Participants were recruited as healthy infants from the community. Eighteen children weighed less than 5.5 lbs. at birth or were born at less than 36 weeks gestation. We examined the task performance of these 18 children relative to the rest of the sample. The healthy preterm/low birth weight infants performed lower on three tasks: tongue task at 24 months (t = 3.88, p = .002), Simon says at 36 months (t = 4.37, p = .009) and Backwards digit at 48 months (t = 2.79, p = .007). All children with available data were included in the current analyses.

Because of the longitudinal nature of the study, there was some attrition in laboratory visits over the nine years. The most common reason for attrition was families moving out of the area. At the 10-month visit 365 children returned, at the 24 month visit 305 children returned, at the 36-month visit 293 children returned, at the 48-month visit 254 children returned, and at the 9-year visit 224 children returned. Examination of missingness showed that there was no difference in sex (χ² = 2.03 p = .154), race (χ² = 1.00 p = .910), or maternal education (χ² = 8.61 p = .072), between all children who returned for the 9-year visit and those that did not from the original 5-month sample. All children with available data were used in the current study using Full Information Maximum Likelihood Estimation, which is robust to data missing at random. The sample is primarily White (77.6%) with 13.7% of children identified by the parent as Black/African American, .5% as Asian, and 8.3% as multiracial/other. When asked about ethnicity, 6.6% identified their child as Hispanic. For maternal education level at the 5-month visit, 2.4% of mothers had not completed high school, 14.4% completed high school, 20.0% completed an Associates or two-year degree, 37.8% finished a four-year degree, and 23.9% had a postgraduate degree. Six mothers, 1.5%, did not report their education level. Mothers averaged 29 (SD = 6) years of age at the time of their child’s birth. These demographics aligned with the populations from which the combined samples were recruited and represent infants in the And B geographic locations in the mid-Atlantic region of the United States.

Procedures

Data were collected at both research locations using identical protocols. Research assistants from each location were trained together by the project’s Principal Investigator (second author) on protocol administration and behavioral coding. To ensure that identical protocol administration was maintained between the labs, the Virginia Tech-A lab periodically viewed video recordings of lab visits by the University of North Carolina Greensboro-B lab. To ensure that identical behavioral coding maintained between labs, the Virginia Tech-A lab provided the reliability coding of all University of North Carolina Greensboro-B behavioral coding.

EF Tasks

EF tasks were selected to be both developmentally appropriate for age at each time point and reflective of the overarching construct. All tasks are commonly used in the EF literature and reflect combinations of IC, WM, and shifting demands and include tasks that required behavioral regulation in addition to cognitive demands (Garon et al., 2008; Petersen et al., 2016; Zelazo & Carlson, 2012). At each age, reliability of behavioral coding was measured using intraclass coefficient (ICC) for at least 20% of researcher-coded variables. ICCs ranged from .75 to 1.00.

5 months.

The 5-month task was the looking version of the A-not-B task. There are very limited EF tasks that are appropriate for use with infants this young and the A-not-B task is the most commonly used task that requires aspects of IC and WM, rather than measuring attention or emotionality. The use of a single task in infancy is a limitation of the current study and we revisit this in the discussion. For the looking A-not-B, infants were seated on the lap of their parent, approximately 1.1 meters away from a table with the experimenter seated on the other side following the procedure detailed by Bell and Adams (1999). The experimenter first showed the infant an attractive toy then covered the toy by one of two orange and blue buckets while the infant watched. Then the experimenter distracted the infant by calling their name and directing their attention away from the bucket and to midline. The experimenter then asked, “where’s the toy?” and evaluated if the infant’s first eye movement looked towards the bucket. If the infant looked at the correct bucket, the trial was counted as correct and the experimenter verbally praised the infant. If the first eye movement was towards the incorrect buckets, the researcher would verbalize disappointment and reveal the toy underneath the correct bucket. After two correct trials on the same side, the experimenter switched the placement of the toy to underneath the second bucket, to test the A-not-B error. Starting side was counterbalanced across participants. The infant had to correctly complete two out of three trials to receive credit. Video recordings were used to score participant’s success on the following scale: 1. Object partially covered with one tub. 2. Object completely covered with one tub. 3. Object hidden under one of two identical tubs. 4. A-not-B with 0 delay. 5. A-not-B with 2-sec delay. 6. A-not-B with 4-sec delay (Bell & Adams, 1999).

10 months.

EF task at this age was identical to the 5-month procedure.

24 months.

Five EF tasks were administered at the 24-month visit and included A-not-B, Simon says, tongue task, DCCS, and crayon delay. The A-not-B task involved invisible displacement (Diamond et al., 1997; Morasch & Bell, 2011). The child sat in a highchair across from the experimenter who hid a ball underneath one cup on either the left or right side of the table. A screen was then placed in front of the bucket and a second identical cup was added on the other side. The screen remained up for a 5 second delay. The screen was then lifted and the RA asked, “where’s the ball?” The child’s first look, gesture, or vocalization towards a cup was coded based on video recordings by trained research assistants as correct or incorrect. After two successful trials on one side, the experimenter switched the hiding to the other side. Beginning side was counterbalanced across participants. Performance was scored offline from video recordings and the proportion of correct responses was the variable of interest.

The Simon says task was based on the Bear/Dragon task (Kochanska et al., 1996; Reed et al., 1984). For this task, the child was instructed to do what the “nice pig tells us” and not to do what the “mean bull tells us.” Children demonstrated understanding of the rules during two practice trials and then ten test trials were administered with pig and bull trials interspersed in a pseudorandom order. The proportion of correct “bull” trials was the variable of interest.

For DCCS (Zelazo, 2006) the experimenter first instructed the child to sort six cards with either a red or blue car or flower according to either shape or color. If the child demonstrated that they had learned the rule, the experimenter told them that the rules had changed and now they should sort by the opposite dimension, so if they had been sorting by shape, not they would sort by color. Starting dimension was counterbalanced across participants and presented in a pseudo random order (ABABBA). Proportion of correctly sorted cards in the “pre-switch” condition was the variable of interest at this age. Despite no rule switch, children still had to maintain the rule in WM and engage in the task while inhibiting other responses.

During the tongue task (Kochanska et al., 2000) the experimenter sat on the floor with the child and instructed them to hold a goldfish cracker on their tongue without eating it. The first trial lasted 10 seconds, followed by a 20 second trial, and finally a 30 second trial. Proportion of trials during which the child successfully did not eat the goldfish before time was up was the variable of interest.

During the crayon delay task an experimenter told the child they were going to color and placed a brand-new box of crayons and fresh white paper on the table in front of the child. The experimenter then told the child that she had to go to the other room and not to touch the crayons until she returned. Parents remained in the room, but were instructed to ignore the child and the experimenter was gone for 60 seconds. Behavior coded on the following scale was the variable of interest: 1 (colors with crayons), 2 (takes crayons out of box), 3 (picks up box), 4 (touches box), 5 (touches paper), or 6 (does not touch). Latency to touch was not used as it was highly correlated with the behavioral coding (r = .82, p < .001) and would have required transformation to be included in a factor with the other EF measures.

36 months.

Five tasks at this visit were similar to the ones administered at 24-months and included the tongue task, Simon Says, day/night, DCCS, and crayon delay. For the day/night task (Gerstadt et al., 1994) the experimenter told children they were going to play a silly game. The experimenter showed them a picture of a moon and stars on a black background and a yellow sun on a white background and instructed them to say “day” when they saw the picture of the moon and “night” when they saw the picture of the sun. Two practice trials preceded the task to ensure the child understood. Sixteen cards were presented in a pseudorandom order and no feedback was given during the task. Proportion of correct trials was the variable of interest.

Other tasks were identical to those administered at the 24-month with the following exceptions. The Simon says procedure used a horse and cow puppet instead of a pig and a bull. DCCS post-switch trials were used. The crayon delay and tongue task were the same procedure as the 24-month appointment.

48 months.

Seven tasks were administered at the 48-month visit and were similar to the 24- and 36-month tasks, including yes/no, backwards digit span, DCCS, crayon delay, tongue task, day/night, and Simon says. The yes/no task (Wolfe & Bell, 2004, 2007) asked the child to shake their head no when the experimenter said “yes” and to nod their head yes when the experimenter said “no.” During this task there were two practice rounds and 16 trials were presented in a pseudorandom order. Proportion of correct trials was the variable of interest.

The backward digit span task asked children to repeat a series of digits backwards (Hilbert et al., 2015). For example, the experimenter would say “4, 3” and the child would respond “3, 4.” Trials began with a two-digit series. If children were able to correctly list both digits in the opposite order during one of two trials they moved onto a three-digit series. If they successfully complete one of two three-digit trials they moved onto a four-digit series. Highest correct trial achieved was the variable of interest. Scores ranged from 1 to 4 digits successfully achieved.

During DCCS children were given the borders condition if they successfully completed 5 of the 6 post-switch trials. For the borders condition of DCCS, the child was instructed to sort 12 cards according to one dimension (i.e. color) if the card had a border and according to the other dimension (i.e. shape) if the card had no border (Zelazo, 2006). Dimensions were counterbalanced across participants. Proportion of correct trials was the variable of interest.

Instead of crayon delay, children completed a gift delay (Kochanska et al., 1998). For the first part of the task children were instructed to stand facing the wall while the experimenter wrapped a gift for them. The experimenter made loud wrapping noises and reminded the child not to peek for 60 seconds. Then the child was shown a wrapped gift that was for them. The experimenter then said they had forgotten a bow and were going to get it, but that the child should not touch the gift until the experimenter returned (180 seconds later). Extent of peeking (1 meaning turns fully around to peek to 5 meaning no peeking) was coded for the wrapping portion and qualitative coding of degree to which the child interacted with the gift (1 meaning opens gift to 4 meaning no touching) were the variables of interest. No child in the current study completely opened the gift (range 2-4).

Tongue task, day/night, and Simon says were administered identically to the 36-month visit, except children were instructed to listen to the “nice horse” and not to listen to the “grumpy cow” during Simon says.

9 years.

The EF tasks at 9 years were chosen to show continuity with the early childhood tasks, but be developmentally appropriate for older children and similar to those used in the adult literature. Tasks included number Stroop, Wisconsin Card Sort Task (WCST), N-back and backward digit span. All but the backward digit span were administered on the computer and not coded by research assistants.

The Number Stroop (Ruffman et al., 2001), similar to the Stroop-like tasks previously described, began with instructions for the child to count the number of letters that appeared on the screen and indicate that number of letters by pressing the number key on a keyboard. There were two practice trials and the child was instructed to, “go as fast as you can, but get as many right as you can.” The experimenter then told the child they were going to play the exact same way, but now there would be numbers on the screen instead of letters. Again, there were two practice trials. Finally, the experimenter introduced the third, mixed condition. During this condition, the experimenter told the child to respond the same way, but there would be both letters and numbers appearing on the screen. The goal in all three conditions was to count the number of digits on the screen. RT during the mixed condition was the variable of interest. RT was rescaled using the proportion of maximum scaling (POMS) method (Moeller, 2015) to preserve the distribution and covariance matrix but put it on a similar scale as other variables and higher scores are indicative of worse performance.

The WCST (Greve et al., 2005) consisted of 64 cards the child was instructed to sort according to one of 3 dimensions: color, shape, or number and in the current study was conceptualized as similar to the DCCS in that it required sorting cards by varying dimensions and required additional set-shifting between sets of rules. The child was told that the computer would tell them if they were sorting correctly or incorrectly and to try to sort as many cards correctly as possible. Undisclosed to the participant, the sorting rule changed several times throughout the duration of the task. Age standardized t-scores of number of perseverative errors was the variable of interest. T-scores were calculated based on independent sample norms and rescaled using the POMS methods

The N-back task was developed through combining strategies used by a number of studies (Ciesielski et al., 2006; Richards et al. 2009). For the N-back task children were shown a series of images and instructed to press the space bar when they saw an image that matched the image they had seen one image earlier (1-back). For the second block, children were instructed to hit the space bar when they saw an image that matched the image they had seen 2 images earlier (2-back). Each block consisted of 16 images of sea creatures, with 25% of the images requiring a response. Each image was shown for 2 seconds with an intertrial interval of 1 second. Children were instructed to identify a pattern and to press the spacebar whenever this pattern was identified. To elaborate, the children were told to press the spacebar if an image matched an image they saw two images earlier, a concrete example was shown before practice to aid in understanding. Number of overall false alarms in both the 1-back and 2-back condition was the variable of interest (Pelegrina et al., 2015) and scores ranged from 0 to 20.

The backward digit span was administered identically to the 48-month visit. Variable of interest was total span, measured as the highest number of digits that were correctly repeated back. Scores ranged from 1 to 7 at this age.

Analyses

Confirmatory factor analysis (CFA) and structural equation modeling (SEM) were used using the lavaan package in R version 3.6.1 (Rosseel, 2012). SEM allows latent factors that cannot be measured directly to be included in a model; this is particularly important for EF which by definition is multifaceted and difficult to measure using a single task. We did not use a composite measure of EF because although composites are a more simple method of including multiple measures, they are limited as they assume that each score included in the composite equally contributes to the construct. We used SEM which allows the manifest variables loading onto a single latent factor to account for the same variance and previous work has found better longitudinal continuity of EF when using latent factors compared to composite scores (MacCallum & Austin, 2000; Willoughby & Blair, 2016). Standard errors were calculated using robust MLR, which is robust to nonnormality and nonindependence of observations and missing data was accounted for using full information maximum likelihood estimation. Fit statistics were calculated and evaluated holistically with better fit associated with a nonsignificant chi square value, RMSEA of less than .08, and CFI of over .90 (Hu & Bentler, 1999; Schermelleh-Engel et al., 2003). We were particularly interested in which individual tasks significantly loaded together at each age to help inform task choice in future studies. A data driven approach was taken to understand which tasks loaded onto a factor of EF at each age and is described below.

Results

Descriptive statistics and data transformations

Means, standard deviations, and correlations of all study variables after transformation can be found in Table 1. Skew and kurtosis were calculated and three variables had a skew or kurtosis value of greater than two: A-not-B at 5 months, tongue task at 48 months, and n-back at 9 years. We took a conservative approach as suggested by Kline (2016) and transformed those variables. A value of one was added, as zero was a possible score, and the natural log was taken to improve normality. This improved the normality to a value of less than 2 for A-not-B and n-back, but not tongue task. Thus, tongue task at 48 months was not included in any further analyses.

Table 1.

Means, standard deviations, and correlations for all study variables

Variable	M	SD	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26
1. A-not-B 5m	1.27	0.23
2. A-not-B 10m	5.79	1.56	.09
3. A-not-B 24m	0.66	0.21	.00	.09
4. DCCS pre-switch 24m	0.63	0.24	.05	.03	.15*
5. Simon says 24m	0.53	0.42	−.05	.01	.00	−.18*
6. Tongue task 24m	0.43	0.41	−.02	−.04	.09	.17**	−.29**
7. Crayon delay 24m	2.90	1.89	−.03	−.04	.10	.16**	−.14	.13*
8. DCCS post-switch 36m	0.45	0.45	.04	.16*	.09	.09	−.20*	.06	.15*
9. Simon says 36m	0.85	0.28	−.02	.05	−.07	.04	−.22**	.06	.06	.05
10. Tongue task 36m	0.70	0.40	−.12*	.02	.06	.06	−.08	.31**	.09	.01	.11
11. Day/night 36m	0.42	0.30	−.09	.11	.04	.11	−.12	.19**	.20**	.22**	−.07	.16*
12. Crayon delay 36m	4.58	1.41	−.07	.03	.02	.05	−.09	.12	.16**	.07	−.04	.22**	.12
13. DCCS border 48m	0.56	0.15	−.01	−.07	.05	.03	−.14	−.03	.21*	.13	.11	.04	.05	.03
14. Simon says 48m	0.84	0.33	−.01	.07	.00	.02	−.01	.04	.14*	.05	.17*	.16*	.05	.19**	.07
15. Tongue task 48m	2.93	0.21	−.19**	.01	.00	.11	−.10	.05	.14*	.00	−.02	.14*	.08	.17**	.06	.12
16. Day/night 48m	0.73	0.21	−.04	−.08	.08	.05	−.12	.14	.05	.01	.03	−.02	.01	.01	.21*	.09	−.01
17. Yes/no 48m	0.76	0.27	.06	−.01	.14*	−.02	−.01	.01	.07	−.06	.13	.20**	−.05	.10	.21*	.15*	.05	.23**
18. Gift peek 48m	3.65	1.50	.01	.01	.07	.04	−.10	.14*	.10	−.03	.04	.16*	.17*	.22**	.16	.10	.13*	.11	.07
19. Gift touch 48m	3.27	0.66	−.04	−.09	−.06	.12	−.08	.06	.13*	−.06	−.12	.07	.05	.14*	.15	.10	.24**	.02	.02	.32**
20. Backwards digit 48m	1.86	0.67	−.03	.09	.06	−.26*	.11	.05	.09	−.07	.00	.01	−.08	.18	−.02	.18	.17	.03	.04	−.10	−.01
21. WCST 9y	0.56	0.27	.15*	.12	.00	.13	−.12	.20**	.03	−.04	.07	.05	−.07	.09	.03	.03	.10	.10	.15	−.03	.10	−.08
22. Stroop 9y	0.33	0.19	−.10	−.08	−.10	−.12	10	−.02	.06	−.10	−.21**	.08	−.07	−.03	.18	−.11	.00	.06	−.10	−.07	−.01	−.01	−.17*
23. Backwards digit 9y	3.56	0.85	.14*	.04	.09	.24**	−.31**	.07	.09	.19**	.13	−.05	.04	.12	.08	.16*	−.02	.00	.11	.06	.12	.11	.20**	−.25**
24. N-back 9y	1.34	0.60	.00	−.02	.02	−.18*	.12	−.06	−.18**	−.15*	−.04	−.02	−.17*	−.02	.01	−.09	−.10	−.06	.00	−.20**	−.21**	−.02	−.04	.10	−.21**
25. Sex	1.51	0.50	.06	.06	−.08	.05	−.08	.19**	−.02	.12	.11	.07	.07	.18**	.01	.12	.21**	.05	−.06	.03	−.05	.08	.12	.00	.07	−0.1
26. Maternal education	2.58	1.16	.03	.08	.10	.16**	−.12	.07	.14*	.18**	.10	.19**	.12	.19**	.10	.15*	.07	.04	.12	.14*	.15*	.09	.12	−.15*	.29**	−.20**	−.08
27. Race	4.95	0.57	.02	.00	.01	.07	−.29**	.14*	.04	.10	.01	−.05	.04	.07	.03	−.06	−.06	.05	.01	.11	.12	.01	.04	.00	.18**	−.07	−.04	.09

Note. M and SD are used to represent mean and standard deviation, respectively.

^*indicates p < .05.

^**indicates p < .01.

Measurement model

CFA 1/Initial CFA.

An initial CFA with all possible EF variables at each age point converged normally after 306 iterations. The fixed variable at each age was decided by examining the correlations between variables at each age point (Steiger, 2002). This model fit the data well (χ² = 240.79, df = 217, p = .128, RMSEA = 0.02, CFI = .90), but several factor loadings were not significant and contrary to expectations, the coefficient for 24-month Simon says was negative (see Table 2).

Table 2.

Factor loadings for CFAs

Initial CFA						Final CFA
		B	SE	Standardized coeffecient	p			B	SE	Standardized coeffecient	p
EF 5m						EF 5m
	A-not-B	1					A-not-B	1
EF 10m						EF 10m
	A-not-B	1					A-not-B	1
EF 24m						E F 24m
	DCCS pre-switch	1		0.36			DCCS pre-switch	1		0.37
	Simon says	−2.39	0.67	−0.49	0.000		Tongue task	1.80	0.67	0.40	0.000
	Tongue task	2.27	0.74	0.49	0.000		A-not-B	0.50	0.21	0.21	0.014
	A-not-B	0.40	0.18	0.17	0.023		Crayon delay	8.92	3.90	0.43	0.000
	Crayon delay	7.87	2.26	0.37	0.001
EF 36m						EF 36m
	Day/night	1		0.44			Day/night	1		0.46
	Simon says	0.87	0.28	0.28	0.003		Simon says	0.94	0.31	0.31	0.006
	Tongue task	1.31	0.51	0.44	0.000		Tongue task	1.11	0.58	0.39	0.003
	Crayon delay	4.04	1.56	0.39	0.000		Crayon delay	3.88	1.95	0.38	0.001
	DCCS post-switch	0.87	0.35	0.27	0.020		DCCS post-switch	0.97	0.38	0.3	0.027
EF 48m						EF 48m
	DCCS-borders	1		0.47			DCCS-borders	1		0.51
	Simon says	1.45	0.72	0.30	0.013		Simon says	1.53	0.79	0.36	0.002
	Yes/no	1.06	0.51	0.31	0.015		Yes/no	1.40	0.80	0.40	0.000
	Day/night	0.83	0.34	0.27	0.017		Day/night	1.42	0.54	0.33	0.012
	Gift touching	4.17	1.92	0.44	0.000		Gift touching	2.57	1.16	0.30	0.003
	Gift peeking	11.06	5.02	0.51	0.000		Gift peeking	7.37	3.14	0.38	0.000
	Backwards digit span	−0.20	1.85	−0.02	0.914
EF 9y						EF 9y
	Backwards digit span	1		0.6			WCST	1		0.6
	Stroop	−0.13	0.04	−0.35	0.000		Stroop	−0.13	0.04	−0.35	0.000
	N-back	−0.56	0.19	−0.35	0.001		N-back	−0.55	0.19	−0.35	0.001
	WCST	0.19	0.07	0.36	0.000		Backwards digit span	0.19	0.08	0.36	0.001
Fit	X2 = 240.79, df = 217, p = .12B, RMSEA = 0.02, CFI = .90					Fit	X2 = 186.25, df = 174, p = .249, RMSEA = 0.01, CFI = .94

CFA 2.

Based on the non-significant factor loadings from CFA 1, backwards digit span at 48 months was removed from the second CFA model. Additionally, Simon says at 24 months was removed because of negative coefficient. This model was a moderate fit for the data (χ² = 182.46, df = 157, p = .08, RMSEA = 0.02, CFI = .88) and all factor loadings were significant.

CFA 3/Final CFA.

Finally, based on suggested modification indices, a third CFA model correlated gift delay peek and touch scores at 48 months and correlated tongue task scores at 24 and 36 months,. This model was a good fit for the data (χ² = 186.25, df = 174, p = .249, RMSEA = 0.01, CFI = .94; see Table 2) and was used in the stability analysis (see Figure 1).

Figure 1. Final CFA model with standardized factor loadings.

Note. EF: Executive Function; AB: A-not-B; PreDC: DCCS Pre-switch; PostDC: DCCS Post-switch; BorDC: DCCS Borders; TT: Tongue task; CD: Crayon delay; DN: Day/night; SS: Simon Says; YN: Yes/no; GD peek: Gift delay peeking; GD touch: Gift delay touching; WCST: Wisconsin Card Sort Task; ST: Stroop; BD: Backwards digit span. Bolded values are significant at p < .05.

SEM Stability Analysis

Using the latent factors from the final CFA model, we tested an autoregressive longitudinal model with EF at each time point predicting the following age using SEM. This model fit the data moderately well (χ² = 206.78, df = 184, p = .120, RMSEA = 0.02, CFI = .90). Regression coefficients from 5-month EF to 10-month EF and from 10-month EF to 24-month EF were not significant. However, 24-month EF predicted 36-month EF, which in turn predicted 48-month EF and 48-month EF predicted 9-year EF (see Table 3).

Table 3.

Stability model

Path	B	SE	Standardized coeffecient	p	R2
EF 5m → EF 10m	0.30	0.18	0.09	0.096	0.01
EF 10m → EF 24m	0.00	0.00	0.07	0.413	0.01
EF 24m → EF 36m	1.51	0.74	0.90	0.043	0.81
EF 36m → EF 48m	0.35	0.16	0.63	0.031	0.39
EF 48m → EF 9y	4.47	1.75	0.62	0.011	0.38

χ² = 206.78, df = 184, p = .120, RMSEA = 0.02, CFI = .90

Discussion

We created latent factors of EF at six different ages from infancy through late childhood using a variety of lab based behavioral EF measures. Using these factors, we showed that by 24-months EF is relatively stable and predictive of the next age-based factor, but that infant EF does not show the same stability.

Simon Says at 24 months was removed from the model because the factor loading was in a different direction than expected, meaning that proportion of correct trials was negatively loading onto the EF factor. After confirming that the data were coded correctly and the score was indeed the proportion of trials that that child correctly did not do what the “mean bull” said, it was removed from the 24-month factor. Kochanska and colleagues (1996) used the task with children beginning at 26 months and Garon and colleagues (2008) report that the task is reliable at 25 months. It may be that the dual demands of that task were too difficult at 24 months and that children with higher EF were better overall at following directions at this age, even the directions of the “mean bull.” Further work is needed to examine this task with 24-month old children and understand how very young children understand the demands of the task.

Interestingly, backwards digit span was also not supported as part of our EF factor at 48 months, although it loaded well onto the 9-year EF factor. This might be due to the low overall performance of 48-month-old children on this task and lack of variability, despite others findings that backwards digit span is useful in children as young as 3 and correlates with other EF tasks such as Simon says (Carlson et al., 2002; Garon et al., 2008). However, Carlson and colleagues (2004) report that backwards digit span did not correlate with delay tasks, such as the gift delay used as part of the EF battery in the current study. It is an interesting finding that backwards digit span did load onto the EF construct at 9 years, meaning that the structure of intercorrelations between EF tasks that tap different aspects of IC and WM may change from early to late childhood. Similarly, the profile of EF changes from early to late childhood with more differentiation of factors emerging later on (Xu et al., 2013), so perhaps the WM demands of the backwards digit span load better with more complex EF tasks that are appropriate for older children. This finding highlights the importance of using a battery of behavioral EF tasks in preschool children and a data driven approach in creating latent factors (Carlson et al., 2004). A single task most likely has too much measurement error to be a useful indicator of EF and even a popular task may not be a useful indicator in a given sample (Carlson, 2005; Rushton et al., 1983). This must be balanced with study design, as using multiple, often repetitive tasks, may make it difficult for young children to stay engaged or cause fatigue (Dekkers et al., 2017; Peverill et al., 2017).

Both the CFA and the SEM model showed high associations between the 24-month and 36-month factors. The tasks used during these ages were almost identical and the extracted shared variance between them was highly correlated between the two visits. The SEM model showed that EF stabilized at 24 months, suggesting that by toddlerhood, EF performance is measurable using the tasks included in the CFA: DCCS pre-switch, A-not-B, tongue task, and crayon delay. This shows that although EF remains difficult to measure due to task constraints, toddler EF as measured by these tasks is a relatively stable predictor of later EF. The use of multiple age appropriate indicators of toddler EF is a strength of the current study and suggests that using multiple measures of EF at 24 months shows continuity across early childhood EF. The autoregressive paths between the EF factors supported the hypothesis that 5- and 10-month EF would not be a good predictor of later EF. This fits well with other work showing low stability from 10-month to early toddler EF (Friedman et al., 2011; Hendry et al., 2021). However, 24-month EF did predict the 36-month factor, which then predicted 48-month factor. These coefficients were all in the expected direction with higher EF at each time point predicting higher EF at the following time point.

Although statistically significant and in the hypothesized direction, the path from 48-month EF to 9-year EF was less robust that we expected, given previous work suggesting that EF stabilizes in late preschool until developing further in adolescence and adulthood (Best & Miller, 2010). This association may be due to the changes in tasks used at the 9-year visit compared to the 48-month visit. Although the tasks were chosen to be conceptually similar, tasks at the 9-year visit were computer based, rather than interaction tasks administered by the experimenter. This change in modality may have effected individual’s performance based on their interpersonal abilities or social motivation. Additionally, the 9-year tasks involved more WM tasks (Backwards Digit and N-Back) compared to the 48-month battery, which may have introduced more heterogeneity into the factor compared to the 48-month tasks.

This longitudinal path of EF development from toddlerhood to late childhood is an indication of the relative stability of EF over childhood, but we were not able to show continuity from the infant measures of EF we reported previously in Blankenship et al (2019). This may be due multiple reasons, including our use of 9-year EF as the outcome of interest rather than reading ability at age 6, which was the focus of Blankenship et al (2019). In addition, our previous work used a different measure of infant EF from the A-not-B task than we used in our current analyses (i.e., percentage correct on reversal trials versus scale score, respectively) and the current study utilized more tasks at 36 and 48 months because of the SEM approach we took and we added a 24-month time point. Finally, the Blankenship et al (2019) study also included infant attention as the foundation for infant EF as the foundation for later developing EF; we focused solely on infant EF as the foundation for later developing EF. Infant attention is a precursor to EF but does not require the same cognitive complexity These differences point to the importance in task selection when measuring EF and suggest that the stability of EF over the course of early childhood may rely on task choice and whether a reflective (latent factors) or formative (composite) approach is used (Willoughby & Blair, 2016). We posit that using more tasks and a reflective latent variable approach allows for increased clarity of the stability of a common factor of EF across early childhood.

Measures EF in toddlerhood, evidence of stability of EF by 24 months, and 24-month EF as an early indicator of later EF are important contributions to the literature (S.E. Miller & Marcovitch, 2015). Our findings are likely due to the wide range of tasks we used and the similarity of the tasks at each age. For example, DCCS was used as an indicator in the 24-, 36-, and 48-month factors, but different parts of the task were used at each age. The pre-switch, post-switch, and borders conditions of DCCS become increasingly difficult, meaning that there was enough variability at each time point to capture a range of skills. WCST used at age 9 has many similarities to the DCCS, but at a more difficult level. Our findings suggest that in addition to using multiple tasks, it may be important to use comparable tasks with similarities in rules and application as well as constructs across development to measure EF. This leads us to echo a call to create and validate EF tasks that can be used across the lifespan (e.g., Anderson, 2002; Carlson et al., 2016; Holmboe et al., 2008; S.E. Miller & Marcovitch, 2015; Willoughby et al., 2018), with the hope that by using similar tasks at different periods in development we will be able to capture continuity in EF development.

Our current findings are not able to parse the possibility that general intelligence or verbal ability underlie the EF tasks. There is a high correlation between IQ measures and EF in childhood (Arffa, 2007) and it may be that children who are better able to understand the expectations of the tasks are then able to perform better, not because of their EF abilities, but because of their verbal abilities. The stability we show here in EF from 24 months to 9 years may be the result of continuity in verbal ability. There is differential longitudinal prediction of EF compared to verbal abilities in preschool children, however, so that while correlated, verbal ability does not explain all the variance in EF (Gooch et al., 2016). Future work may use verbal ability as a longitudinal predictor or concurrent covariate in models measuring EF to determine the profile of stability in relation to verbal ability or IQ.

The existence and measurement of EF in toddlerhood is important for early intervention to target EF improvement that might be valuable for language and social development (Diamond & Ling, 2016; S.E. Miller & Marcovitch, 2015). There are many EF interventions, but most are targeted to children preschool aged and older (e.g., Barnett et al., 2008; Sasser et al., 2017). Our findings on the stability and measurement of EF at 24 months illuminate the possibility of measuring EF during toddlerhood and creating interventions to improve EF before children enter preschool. These interventions could target parenting behavior such as parental affect and engagement, which have been demonstrated to be related to EF development (Cuevas et al., 2014; Fay-Stammbach et al., 2014; Valcan et al., 2018). Using multiple behavioral tasks to measure toddler EF may be important to understanding if and how early interventions are effective.

The limitations of our study include the retrospective nature of the analysis and inability to add additional EF tasks that have been developed since the inception of the current study (Broomell & Bell, 2017; Holmboe et al., 2008), as well as only one EF task in infancy. Including additional developmentally appropriate EF tasks at the 5- and 10-month time points may have improved the predictive stability of infant factors and our results do not rule out the possibility that with better measurement infant EF may predict later EF (Hendry et al., 2016). The current results are based on a large, low risk, community sample that is relatively highly educated (62% of mothers had a university degree), which may not generalize to other populations.

In summary, we demonstrated that EF is measurable using multiple, age appropriate behavioral tasks and that, beginning at 24 months, EF is stable to late childhood. We transparently demonstrated our data driven task selection process to show why a data driven approach to using multiple tasks is critical for examining EF. Our longitudinal findings contribute to the understanding of the structure and development of EF.

Data availability statement:

Data included in this manuscript will be made available upon request from the corresponding author (A.P.R. Broomell).