Transitivity Types Predict Communicative Abilities in Infants at Risk of Autism^*

Learning can be defined as an understanding of action-reaction contingencies, and being able to measure such contingencies may aid in investigating developmental processes leading to atypical outcomes. Despite the promising potential of such measurements, methods to incorporate multivariate factors, especially those beyond dyadic interactions, are few. As part of an exploratory investigation, we have developed a protocol to capture all actions and reactions as they occurred temporally between three actors without regard for the information they held. All behaviors, whether social, directed, or seemingly random, were included in this protocol to create a real-time stream of action-reaction sequences.

Analysis of Multirelational Valued Temporal Triads

The Analysis of Multirelational Valued Temporal Triads (AMVTT) was designed to enable the generalized coding of all species across all contexts, developmental stages, behaviors, and environments (White et al., 2014). This method captures all behaviors as interactive links between organisms and their environment. AMVTT measures all interactions, whether they are communicative, social, or attentional, and classifies these interactions as repeatable patterns over time.

Triadic sequencing was useful in studying the brown-headed cowbird, a highly social species (White et al., 2014). By constructing triadic interaction patterns of male-male and male-female song, for example, we could organize a potentially random string of events into meaningful units of interaction that allowed us to predict the birds’ reproductive success. Applying this method to infancy research was an exploratory undertaking. Parallels between cowbird and human social and communicative development demonstrated by Goldstein et al. (2003) led to the current hypothesis that a method predictive of effective communication and reproductive outcomes in birds may be able to predict social and communicative outcomes in humans. To ensure substantial variation in social and communicative outcomes in humans, we explored human infants at risk for developing autism. Studying this population allowed the investigation of non-autistic infants (those who did not receive a diagnosis), infants with communicative delays (those diagnosed with language delays), and infants with social and communicative atypicalities (those diagnosed with autism).

The interactions analyzed occurred in home settings, in which infant and caregiver were told to interact and play normally with their own set of toys. The observed behaviors were recorded in real time in an ecologically valid setting, potentially producing more representative results. This social network analysis was designed to incorporate both social and non-social stimuli by considering all behaviors between caregivers, infants, and surrounding objects as equal, rather than weighting specific behaviors (White et al., 2014). Each directed behavior occurring between the three was counted as a behavioral link. In a mutually exclusive manner, links were combined based on sequence and timing to form interaction patterns, or triads. The resulting triads were then categorized by the completion of transitivity, and assessed by directionality and reciprocity (Butts, 2008; Wasserman & Faust, 1994).

For analysis, triadic interactions are laid out in 15 simple patterns, or types. For a full list of potential triad types, see Figure 1. Each type represents a set of possible links between all three nodes. One triad type can have several different states based on which node first linked to another. The interactive links between each node are assembled into triad types.

Figure 1:

For the current research, the three nodes within a triad always represented an infant, a caregiver and an object. To apply this method, a thorough protocol defining behavioral links between the three nodes was implemented. Specific parameters for what constitutes a behavioral link were set for each of the three potential nodes. For example, while touches from either caregiver or infant were always considered links, infants’ subtle directed behaviors (e.g., eye gaze, head turns) were also coded as links, whereas caregivers’ behaviors had to be more pronounced (e.g., vocalizations) to qualify. For objects, only those with state-changing properties could form outbound links, while static ones could not. For further descriptions of these parameters, see Appendix A.

Triadic Census

For this exploratory project, social networks of triads were measured representing interaction patterns between an infant, a caregiver and surrounding objects (Holland & Leinhardt, 1971; White et al., 2014). This type of interactive triad has been shown to demonstrate the most consistency across ages 6 through 18 months, when infants display social and object-oriented engagement (Bakeman & Adamson, 1984).

Focusing on triads, interaction patterns were assessed for whether transitivity was present as dictated by the presence or absence of structural balance, or transitivity. The theory of transitivity is used in social network analysis, as well as in the current project, to categorize these patterns into three types of transitivity as defined by Wasserman & Faust (1994): transitive, intransitive, and vacuously transitive. Interaction patterns are defined by whether transitivity was absent (vacuously transitive), violated or incomplete (intransitive), or present (transitive) (Wasserman & Faust, 1994).

In a balanced structure, when element A is related to B and C, B and C are also related creating a transitive property (Heider, 1946; Holland & Leinhardt, 1971). Transitivity has been used, by applying balance theory, to assess the structural integrity and balance of social networks pertaining to transmission of communication, interactions, friendships, etc. (Moore, 2013; Schimada, 2013; van den Oord et al., 2000). According to balance theory, relationships between elements can exhibit or lack balance based on their structural makeup. Across domains, balanced relationships remain while unbalanced ones either disintegrate or are transformed into balanced configurations.

The most well-known and well-studied instances of transitive interactions are those of joint attention (Figure 2). Transitive interactions include instances in which two participants direct their attention toward a third entity, and either both or only one direct their attention toward the other participant as well. Such transitive interactions have been documented to develop around 9 and solidify at 18 months of age (Carpenter et al., 2002).

Figure 2:

Spiro et al. (2013) have described processes of brokerage, in which one of the three entities within a triad can facilitate balance through bridging behaviors. Matchmaking brokerage, for example, occurs in joint attention instances when caregivers name objects for infants, encourage their infants to play with an object, or simply shift their attention between infant and object. The current research analyzes instances of joint attention and bridging behaviors as they occur dynamically in real time and across months with respect to communicative development.

Butts (2008) studied temporally and sequentially constructed triads among human adults and found that triadic interactions contributed little to communication dynamics when compared to reciprocal dyads. In contrast, the current research investigated temporally and sequentially constructed triads among three entities that represent immense stability across a child’s first 1.5 years of life and may consequently constitute important building blocks of communicative development (Bakeman & Adamson, 1984). Triadic interactions were thus predicted to play a significant role in communication dynamics in human infants and their social and physical surroundings.

Intransitivity and transitivity were hypothesized to affect infants differently based on the infants’ propensities toward social versus physical stimuli. While intransitive interactions most commonly represent a lack of contingent responses to the social bids for joint attention between caregiver and infant, transitive interactions contain contingent responses enabling the identification of action-reaction relationships. These contingent responses can occur with social (caregiver and infant) or non-social (infant and object) stimuli. There is evidence that females develop and maintain a higher sensitivity to social stimuli early in life in comparison to males (Connellan et al., 2001). Less frequent social input was thus hypothesized to have a greater negative impact on females, and less socially sensitive males were hypothesized to benefit more from transitivity available via both social and non-social stimuli.

Because triadic interactions move toward transitive patterns that enable effective exchange (van den Oord et al., 2000), we hypothesized those interactions that were more conducive to communicative development would exhibit transitivity. Interaction patterns, in which transitivity was present, were thus predicted to represent instances in which communicative learning is more readily achieved.

For this pilot study, 37 infants, for whom we had developmental outcome information at the time, were studied, and their average proportions of transitivity types obtained between 5 and 14 months were correlated with their developmental outcome measures at 36 months.

As male and female infants have been shown to develop communicative skills differently, correlations for each gender were run separately (McGillicuddy-De Lisi & De Lisi, 2002). Despite the discrepancy in ASD prevalence between males and females, gender differences in symptomology do not seem to appear until later in development (Reinhardt et al., 2015). Some scholars posit that biases in ASD diagnosis toward males may be related to more readily identified hyperactivity or aggression in males in contrast to a higher social and communicative sensitivity cultivated in females (Connellan et al., 2001; Halladay et al., 2015). The male prevalence of ASD may thus have to do with typically occurring differences between genders rather than anything specific to ASD itself (Messinger et al., 2015). Specifically, it may be due to a divergence in sensitivity to social and physical stimuli affecting male and female language development (Connellan et al., 2001). In addition, where females may acquire language more holistically across syntax, semantics, and pragmatics, males may learn language in a more compartmentalized manner (McGillicuddy-De Lisi & De Lisi, 2002).

This study tested the predictive effectiveness of AMVTT by measuring the relationship between transitivity types and developmental outcomes. Hypotheses of this study posited that transitivity and intransitivity proportions would correlate with developmental measures positively and negatively, respectively.

No significant differences were found between males and females and their transitivity type proportion averages or their developmental outcome measures. The way in which transitivity types and developmental outcome measures associated, however, differed by gender.

For males, neither VacT (cor=-0.36, p=0.16) nor IntT (cor=0.00, p=0.99) proportions between 5-14 were predictive of CDI-III subcategory scores at 36 months. TraT proportions between 5-14 months were, however, strongly positively correlated with CDI-III raw word count (cor=0.53, p=0.03).

For females, neither VacT (cor=0.16, p=0.56) nor TraT (cor=-0.20, p=0.48) proportions between 5-14 were predictive of CDI-III raw word count. IntT proportions between 5-14 months were, however, strongly negatively correlated with CDI-III raw word count (cor=-0.51, p=0.04). This strong negative correlation existed across all other subcategories of CDI-III, including number of complex language used (cor=-0.62, p=0.03), number of times language was used (cor=-0.55, p=0.03), and longest sentence means (cor=-0.55, p=0.03).

These results do not simply reflect inter-correlations among CDI-III subcategories: VacT interactions among females correlated significantly with only one of the many subcategories, indicating that while they are all related, they are not necessarily all predicted by the same factors (times language was used: cor=0.55, p=0.03).

There were no significant correlations between male interactions and MSEL scores. Female IntT proportion averages correlated negatively with MSEL standard scores (cor=-0.56, p=0.03). Subcategories of the MSEL also correlated with average intransitivity proportions in females (Fine Motor: cor=-0.55, p=0.03; Receptive Language: cor=-0.55, p=0.02; Cognitive Sum: cor=-0.53, p=0.03). ADOS severity scores ranged from 1 to 7, and all infants but one, who scored 5 or higher (n=4), had ASD. There were no significant correlations between male or female proportions and their ADOS severity scores or their demographic data, as measured by years of maternal education and paternal occupational prestige. See Table 4 for a list of all significant correlations split by gender.

The proportions of interactions in three categories of transitivity that were observed in infants at 5-14 were correlated with their CDI-III, MSEL and ADOS scores. Although males and females did not differ in their averages for transitivity type proportions or developmental outcome measures, TraT correlated positively with a CDI-III measure in males, and IntT correlated negatively with communicative outcome in females across CDI-III and MSEL measures.

The transitive theorem appeared to manifest itself in a complementary fashion in both genders. Where females seemed to be more strongly affected by the negative impacts of IntT, the positive effects of TraT impacted males more. IntT interactions most commonly represent a lack of contingent responses to the social bids for joint attention between caregiver and infant, as well as primarily describing play with a highly interactive (often electronic) object. TraT interactions contain contingent responses enabling the identification of action-reaction relationships. These contingent responses can occur with social (caregiver and infant) or non-social (object) stimuli. There is evidence that females develop a higher sensitivity to social stimuli earlier in life than males (Connellan et al., 2001). Interactions mainly generated through object play may thus have a greater negative impact on socially sensitive females, and less socially sensitive males may benefit more from transitivity available via both social and non-social stimuli. This gender difference, considering that it is present when correlating gender-normed outcome measures (namely, CDI-III) with similar transitivity proportions for males and females, speaks for a substantial finding representing differing learning and interactive sensitivities in males and females. It is possible that similar findings would have become apparent in the MSEL correlations, had those measures been gender-normed as well.

In this exploratory project, we set out to measure patterns so basic and intrinsic to development, that they could be found in any learning organism at any developmental stage. The hope was to develop a method predictive of behaviors without the presence of their precursors or related behaviors. The current study found that intransitive patterns in 5-14-month-old females predicted not only language use (CDI-III), but also communicative and social understanding (MSEL), both assessed at 36 months. The relationship between these early interaction patterns and later, more sophisticated outcome measures shows promise that our method may indeed capture an underlying strategy fundamental to the general development of learning itself. Engaging in certain interaction patterns may set the stage for infants to apply early relational understanding to other, more specific types of understanding, such as those required for communicative and social development.

In the current study, the potential for a broader connection between different types of learning, however, only applies to the female infants. Male effects were found only on raw word count, whereas for females, effects were found across all CDI-III and some MSEL subcategories. These gender differences may not necessarily be specific to the high-risk infants studied as much as they may reflect typically occurring divergence in male and female language development (Messinger et al., 2015). Where females may acquire language more holistically across syntax, semantics, and pragmatics, males may learn language in a more compartmentalized manner (McGillicuddy-De Lisi & De Lisi, 2002). Nevertheless, patterns categorized by transitivity allowed us to measure early patterns predictive of later outcomes in both genders.

Usually, transitivity analyses are applied to static structures. Interested in developmental processes, we applied a technique to take time into account while using the concept of transitivity (White et al., 2014). Additionally, in studying learning mechanisms, our technique was sensitive to contingencies by using action-reaction sequences to form interaction patterns. We thus used sequentially formed triads to analyze real-time interactions as action-reaction units. Instead of modeling singular actions in sequence, as has been done using a Relational Event Model when examining social interactions in jackdaws, for example, (Butts, 2008; Tranmer et al., 2015), we wanted to capture possibly transitive interactions as a whole as they occurred in real time between infants and their social and physical surroundings. Just as cowbird social behaviors could predict reproductive outcome when organized into units, categorizing infant-caregiver-object interactions by transitivity allowed us to predict communicative outcomes (White et al., 2014). In addition, our results provide a noteworthy addition to Butts’ (2008) findings in which transitivity had little effect on communicatively skilled adult humans. By the time communicational abilities have been established, transitivity may not be as important as reciprocity, for example, as demonstrated by Butts (2008). Our results, on the other hand, indicate that transitivity may represent a fundamental process when first learning how to communicate. By applying the theory of transitivity, we may have developed a method of identifying early interaction patterns indicative of communicative development in infants at risk for ASD.

Our method allowed us to categorize interactions patterns based on whether they represented mutuality (and/or symmetry) in dyads or transitivity in triads (Heider, 1946; Wasserman & Faust, 1994). Across several studies, results have shown that social structures move toward mutuality in dyads (A and B interact mutually) and transitivity in triads (e.g., A interacts with B and C, and B interacts with C) as children develop more complex communication and play systems (McGrew, 1972; van den Oord et al., 2000).

As we have mentioned before, joint attention represents a key example of transitivity, and thus typifies the kind of social pattern infants would ideally move toward across development. It follows then that the absence of joint attention has been used as an early indicator of potential deficits later in life. In literature, however, distinctions have been made between joint attention instances in which the infant engages with only the third entity versus with both participant and third entity. While the former has been deemed non-social with respect to the infant’s involvement, the latter has been considered social due to the synchronized social interaction between both participants while attending to a third entity (Moore & Dunham, 1995). It is debated whether this distinction is useful (Yu & Smith, 2013), especially as joint attention, in its various manifestations, may represent such a potent indicator of later development.

AMVTT was designed to either consider this distinction (as tested in further studies) or override it by examining overall transitivity types and consequently weighing both joint attention instances equally (current study). In doing so, we may have created a new method of assessing and identifying various patterns conducive to communicative development in young infants at risk of ASD. Using this method may provide possibilities to not only identify early intransitive patterns that may be indicative of communicative deficits, but also train parents to implement the types of transitive patterns associated with communicative abilities. The findings of the current study emphasize the wide variety with which joint attention and the like can occur, and that when taken as a whole, these instances may predict and possibly further communicative development in infants.

It is important to note that some of the significant relationships measured occurred between two parent-generated variables, namely CDI-III scores and triadic interactions. The possible co-variance between the two variables was partially addressed by adding another trainer-assessed variable, MSEL. Because MSEL scores only correlated with triadic interactions involving female infants, CDI-III correlations in males and females must be viewed with a level of caution. Nevertheless, possibly telling relationships were identified between early interaction patterns and two separate assessments of development two years later.

The audio/video recordings used for behavioral coding of links represent a limitation of our study. We placed great value on obtaining the most ecologically valid recordings by looking at caregiver-infant interactions as they occurred in the home setting. To preserve the validity of such a setting, it was important to keep research personnel at a minimum. This restriction meant using only one camera person per caregiver-infant interaction. The resulting recordings thus likely failed to capture certain gazes, touches, and other pragmatic subtleties important for effective social communication. Ideally, caregiver and infant would have been recorded and coded separately to ensure the most accurate depiction of their behavioral links during interactions.

Another limitation is represented by the small sample size of only high-risk infants. We recognize that our results pertain only to a highly specific population of humans and thus can only speculate on the generalizability of our results to typically developing infants and children. Further studies are needed to assess the applicability of our current results to larger populations and typically developing children. Nevertheless, the findings provide first insights into how triadic analysis and transitivity can shed light on early interactive factors essential to communicative development.

Using the comprehensive approach of triadic analysis to incorporate all behaviors occurring between infant, caregiver, and objects, we aimed to investigate a broader concept of how learning occurs within interactions that include both verbal and non-verbal social and non-social exchange. In the current analysis, categorizing interactions by transitivity allowed us to predict words before words were present, possibly revealing a mechanism representative of something broader than word learning and/or language delays. Rather than measuring how infants specifically learn words, we may have found a way to measure how infants, by engaging in different types of behavioral patterns, learn about learning.