Measures

MacArthur-Bates Communicative Development Inventory (CDI)

Primary caregivers completed the CDI Words and Sentences (CDI-WS; Fenson et al., Reference Fenson, Marchman, Thal, Dale, Reznick and Bates2007), which is normed for children from ages 16 to 30 months to evaluate expressive language skills. Although toddlers from ages 16 to 18 months can also be administered the CDI Words and Gestures form, the CDI-WS is normed for the younger range of this demographic. The CDI-WS reports excellent metrics of reliability, internal consistency, and validity (Fenson et al., Reference Fenson, Marchman, Thal, Dale, Reznick and Bates2007). Previous research supports the use of parent-report questionnaires for the measurement of ISL in this age range (Kristen et al., Reference Kristen, Sodian, Licata, Thoermer and Poulin-Dubois2012) and suggests parent-report measures better capture ISL production compared to naturalistic observations or semi-structured interaction tasks (Pascual et al., Reference Pascual, Aguado, Sotillo and Masdeu2008), as naturalistic or elicitation tasks may represent a skewed perspective of the child’s ISL knowledge (Drummond et al., Reference Drummond, Paul, Waugh, Hammond and Brownell2014).

The CDI-WS comprises three sections; the first section evaluates the words children use with a 680-word vocabulary checklist. Parents select ‘yes’ or ‘no’ for whether their child produces words on the checklist. From this section, 43 ISL terms were extracted (Figure 2). The 43 terms were selected due to their overlap with other measures of ISL (i.e., the Internal State Language Questionnaire; Olineck & Poulin-Dubois, Reference Olineck and Poulin-Dubois2005) and due to their use in previous ISL research using the CDI-WS (e.g., Lamb, Reference Lamb1991; Longobardi et al., Reference Longobardi, Spataro and Colonnesi2018; Raines, Reference Raines2014). Children received a score of “1” for every word the caregiver endorsed, otherwise they received a “0” for that item.

Figure 2. Median number of ISL terms produced over time.

Analytic plan

The analytic plan for aims one and two were pre-registered on Open Science Framework: https://doi.org/10.17605/OSF.IO/TGB2Y.

Aim 1: the factor structure of ISL

To evaluate the latent factor structure of ISL, exploratory (EFA) and confirmatory factor analyses (CFA) were conducted in Mplus (Version 8; Muthén & Muthén, Reference Muthén and Muthén2017). EFA and CFA can be used in succession to provide a more rigorous and comprehensive examination of the factor structure in an observed dataset. When there is limited knowledge of the underlying factor structure of a construct, as is the case for ISL, EFA first identifies the number of latent factors that best explain the pattern of correlations in the data. Once a tentative factor structure has been identified, CFA is then used to confirm, or check, that the exploratory structure still holds. Splitting the larger study sample in half prior to conducting EFA and CFA enhances the robustness of the factor structure exploration and model confirmation by independently validating the findings in one subset and testing their replicability in the second subset. Thus, the WordBank sample of 6,373 children was split into two subsamples using simple random sampling in SPSS (Version 27). One sample was used to conduct an exploratory factor analysis (EFA) to identify the best fitting factor model (Sample A: N = 3,121; M_age = 23.49 months, SD_age = 4.39; 47.2% female). The second sample was then used to perform confirmatory factor analyses (CFA) to confirm the factorial structure obtained from the EFA (Sample B: N = 3,252; M_age = 23.52 months, SD_age = 4.35; 49% female). As anticipated, the subsamples did not differ in terms of child sex, χ²(1) = 1.77, p =. 19, child race, χ²(4) = 4.72, p =. 32, or maternal education, χ²(1) = 1.23, p =. 27.

For the EFA, given that our data are dichotomous (yes/no), a tetrachoric correlation matrix was generated and we utilized the robust weighted least squares (WLSMV) estimator with an oblique geomin factor rotation (Yates, Reference Yates1987). The scree plot, eigenvalues, and factor loadings were examined to identify the optimal number of factors. Factors with an eigenvalue ≥ 1 were retained (Kaiser, Reference Kaiser1960) and the factor loadings were inspected and considered acceptable if ≥. 45 (Hair et al., Reference Hair, Black, Babin and Anderson2010). CFA using the WLSMV estimation method was then performed to validate the identified one-factor model, where all 43 items were loaded onto a single latent ISL factor. The factor loadings were once again inspected and model fit was assessed using the following descriptive and inferential indices of model fit: Chi-square (χ²) test, root mean square of approximation (RMSEA, values ≤. 06 reflect acceptable fit; MacCallum et al., Reference MacCallum, Browne and Sugawara1996), standardized root mean squared residual (SRMR, values ≤. 08 reflect acceptable fit; Hu & Bentler, Reference Hu and Bentler1999), and confirmatory fit index (CFI, values ≥. 90 reflect acceptable fit; Hu & Bentler, Reference Hu and Bentler1999).

Aim 2: The early development of ISL

To evaluate ISL growth across early childhood, multilevel modeling was used to account for the nested structure of the data (i.e., repeated observations within children over time). Multilevel modeling can be used to examine whether and how children’s ISL production grows over time (i.e., within-person changes). By introducing potential predictors of ISL development into the multilevel model, we can further evaluate whether ISL growth patterns or children’s starting points are influenced by child- or maternal-factors such as sex and maternal education (i.e., between-person differences). A balanced dataset is not a requirement of multilevel modeling, and it is therefore well-suited to analyzing data with varying numbers of measurement waves that are unevenly spaced across individual participants (Singer & Willett, Reference Singer and Willett2003). Furthermore, partially missing data (e.g., when some children have just one or two observations) can be handled using full information maximum likelihood, which estimates population parameters that would most likely produce the estimates from the sample data (Curran et al., Reference Curran, Obeidat and Losardo2010). This advantage of multilevel modeling is particularly important as both cross-sectional and longitudinal data are available within the WordBank dataset.

Analyses were conducted in HLM Version 7.0 (Raudenbush et al., Reference Raudenbush, Bryk and Congdon2011) using a taxonomy of two-level statistical models, which provides a baseline from which to compare model fit between nested models using model deviance (Bryk & Raudenbush, Reference Bryk and Raudenbush1992), the Akaike information criterion (AIC), and the Bayesian information criterion (BIC). In other words, increasingly complex models can be generated by introducing new variables into the original multilevel model. This approach allows us to evaluate whether the additional variance accounted for by a new variable is significant, above and beyond the previous model. In the two-level hierarchical linear model, the Level 1 (L1) equation represents the within-person model, the Level 2 (L2) equation represents the between-person model, t represents repeated observations, and i represents individual children. To quantify how much of the total variation in ISL is within-person vs. between-person variance using the intraclass correlation coefficient (ICC; Bryk & Raudenbush, Reference Bryk and Raudenbush1992), we first estimated an unconditional means model (Model A) as follows:

$$ \mathrm{Level}\;1:\hskip1.2em {\mathrm{ISL}}_{\mathrm{ti}}={\pi}_{0\mathrm{i}}+{e}_{\mathrm{ti}} $$

$$ \mathrm{Level}\;2:\hskip1.32em {\pi}_{0i}={\beta}_{00}+{r}_{0\mathrm{i}} $$

In the unconditional means model, the total variance in ISL aggregated across timepoints is partitioned into its within, σ², and between, $ {\tau}_{00} $ , components, which can then be used to calculate ICC. Next, to test our research question regarding the rate of ISL change across early childhood, we estimated an unconditional linear growth model (Model B) as follows:

$$ \mathrm{Level}\;1:\hskip1.2em {\mathrm{ISL}}_{\mathrm{ti}}={\pi}_{0i}+{\pi}_{1i}\ast \left({\mathbf{AGE}}_{\mathrm{ti}}\right)+{e}_{\mathrm{ti}} $$

$$ {\displaystyle \begin{array}{l}\mathrm{Level}\;2:\hskip1.32em {\pi}_{0i}={\beta}_{00}+{r}_{0\mathrm{i}}\\ {}\hskip2.879999em {\pi}_{1i}={\beta}_{10}+{r}_{1\mathrm{i}}\end{array}} $$

In the unconditional linear growth model, AGE (in months) was the only variable entered at L1 in order to generate a baseline model depicting children’s rate of ISL growth (slope) without any L2 predictors entered. To set a meaningful and variable intercept (initial status), AGE was centered at the grand mean so that the initial status value corresponds to children’s average ISL at the end of the second postnatal year (i.e., at 23.5-months). Finally, to address whether variation in ISL initial status or slope is accounted for by either child sex or maternal education status, these time-invariant covariates were dummy-coded and entered into the model. That is, we parameterized a conditional linear growth model (Model C) as follows:

$$ \mathrm{Level}\;1:\hskip1.2em {\mathrm{ISL}}_{\mathrm{ti}}={\pi}_{0i}+{\pi}_{1i}\ast \left({\mathbf{AGE}}_{\mathrm{ti}}\right)+{e}_{\mathrm{ti}} $$

$$ {\displaystyle \begin{array}{l}\mathrm{Level}\;2:\hskip1.32em {\pi}_{0i}={\beta}_{00}+{\beta}_{01}\ast \left({\mathrm{SEX}}_{\mathrm{i}}\right)+{\beta}_{02}\ast \left({\mathrm{EDUCATION}}_{\mathrm{i}}\right)+{r}_{0\mathrm{i}}\\ {}\hskip2.759999em {\pi}_{1i}={\beta}_{10}+{\beta}_{11}\ast \left({\mathrm{SEX}}_{\mathrm{i}}\right)+{\beta}_{12}\ast \left({\mathrm{EDUCATION}}_{\mathrm{i}}\right)+{r}_{1\mathrm{i}}\end{array}} $$

In the conditional linear growth model, ISL for child i at age t is modeled as a function of the intercept parameter (β ₀₀; i.e., ISL at age 2), the slope parameter (β ₁₀; i.e., ISL rate of change), the individual-specific, between-person predictors (β _{_1} and β _{_2}; i.e., child sex and maternal education), and the residual error (e _ti). As mentioned above, missing observations at L1 (within-person language data) were handled using full information maximum likelihood as this estimation method enables all children with at least one wave of ISL data to be included in the models. Children with missing observations at L2 (n = 1,447; between-person predictor data) were excluded only from the conditional linear growth model, hence the change in sample size between Models A and B vs. Model C in Table 2.

Table 2. Multilevel models: fixed effects, variance components, and goodness-of-fit

Results: Aim 1

Results: Aim 2

Multilevel models

Unconditional means model

As a first step toward building the final HLM growth model, we estimated an unconditional means model (Model A) to confirm that the variability in ISL at the individual (L1) and between-person (L2) levels were significantly different from zero. Table 2 shows the variance components, which resulted in an ICC of 0.323. Thus, 32% of the variance in ISL occurs at L2 between children, which is visually represented in the individual growth trajectories depicted in Figure 3.

Figure 3. ISL growth trajectories for a randomly selected subset of children with longitudinal data.

Unconditional linear growth model

Next, we estimated an unconditional linear growth model (Model B) to evaluate the rate of change in children’s ISL over time, and the results are presented in Table 2. Our results suggest that on average by age 2, children produced approximately 17 ISL terms. Additionally, the average rate of growth per month was 2.58, which suggests that children are producing approximately 5 new ISL terms every two months across this age range. We also examined the association between initial status (𝜋₀) and slope (𝜋₁), r = 0.69, p <. 01. This value indicates that children with a larger ISL vocabulary by the end of their second year exhibited a faster rate of ISL growth over time across toddlerhood. As expected, the deviance test verified that the linear model better fit the data in comparison to the means model, χ²(3) = 1494.28, p <. 001.

Conditional linear growth model

Finally, we ran a conditional linear growth model (Model C) that included two L2 predictors to examine whether children’s initial status or rate of change differed as a function of child sex and/or maternal education status. In addition to comparing model AIC and BIC, the deviance test was once again used to verify that model fit improved in moving from Model B to Model C, χ²(4) = 16801.74, p <. 001. As can be seen from Table 2, sex, β₀₁ = 3.08, p <. 001, and maternal education, β₀₂ = 1.53, p <. 001, predicted the number of ISL terms children produced on average at 23.5-months of age. This finding implies that both girls as well as children of mothers with a college degree produced more ISL terms on average near age 2. Together, these variables accounted for 27.96% of the variance in ISL initial status as determined by the pseudo R² value. By contrast, although maternal education predicted ISL slope, β₁₂ = 0.48, p <. 001, child sex did not, β₀₁ = 0.09, p =. 54. Thus, children of more highly educated mothers exhibited a faster rate of ISL growth in comparison to children of mothers who do not hold a college degree. In total, our results suggest that 65.07% of the variance in ISL growth was accounted for by the variables in Model C.

Using the fixed effects generated by Model C, we constructed a model-based growth trajectory starting at the intercept to visualize ISL development between the subgroups in the current study (Figure 4). Four groups were created by pairing child sex with maternal education status. Therefore, Figure 4 depicts how the ISL starting point and trajectories differ between boys and girls as a function of whether their mother holds a college degree.

Figure 4. Model-based ISL growth trajectories as a function of child sex and maternal education.

Aim 1. Factor structure of ISL

Aim 1 evaluated the factor structure of ISL between the ages of 16- and 30-months to investigate the validity of the wide range of ISL categorization practices (Hashmi et al., Reference Hashmi, Vanderwert, Paine and Gerson2021; Lemche et al., Reference Lemche, Kreppner, Joraschky and Klann-Delius2007; Meins et al., Reference Meins, Fernyhough, Fradley and Tuckey2001, Reference Meins, Fernyhough, Johnson and Lidstone2006; Roger et al., Reference Roger, Rinaldi and Howe2012; Siller et al., Reference Siller, Swanson, Serlin and Teachworth2014; Tager-Flusberg & Sullivan, Reference Tager-Flusberg and Sullivan1995). Results from EFA and CFA suggested a one-factor structure best fit the ISL data with all statistical indices of model fit meeting cutoff criteria, aside from the chi-square statistic which was expected given this statistic’s sensitivity to sample size (Babyak & Green, Reference Babyak and Green2010). In other words, collapsing across this narrow age range, results suggest that ISL can best be considered a unitary factor.

Our results can lend a response to an ongoing debate in the ISL literature about whether physiological terms should be considered as internal state language (Kristen et al., Reference Kristen, Sodian, Licata, Thoermer and Poulin-Dubois2012; Meins et al., Reference Meins, Fernyhough, Fradley and Tuckey2001, Reference Meins, Fernyhough, Johnson and Lidstone2006). Although they may emerge in a distinct temporal window with an onset prior to other ISL terms (Kristen et al., Reference Kristen, Sodian, Licata, Thoermer and Poulin-Dubois2012), our results suggest physiological terms warrant consideration as ISL. Further, our results suggest a categorical consideration of ISL in toddlerhood may not be rooted in the underlying factor structure of ISL at this age. Continuing to parse by linguistic category may serve some theoretical relevance for studies interested in certain terms or groups of terms based on a priori research questions but is not empirically supported in this demographic for these ISL terms in this age range.

Our results could be seen as presenting findings that are contradictory to previous works suggesting a developmental order of acquisition of internal state words (and maternal/paternal use of ISL words differentially by category and by age). However, it is important to note that our findings do not discount previous research showing that children begin producing some ISL terms before others. Rather, our results suggest that there appears to be a common, underlying mechanism for ISL production in toddlerhood. That is, children’s production of cognitive ISL terms may not be entirely dissociable from perception ISL terms, for example. These findings may have meaningful implications for targeting the core, underlying mechanism of ISL development in clinical groups such as autism and ADHD. An additional explanation to the developmental timing of certain words could be the general growth of vocabulary. In other words, perhaps as children gain greater vocabulary knowledge with age, this provides access to new ISL terms. Future research may wish to empirically evaluate the onset of ISL terms as a function of vocabulary size (e.g., 50 words, 200 words, 500 words) rather than simply as a function of age in toddlerhood.

Furthermore, although a single factor of ISL was observed in the current study, it is possible that the factor structure of ISL changes over time, which in turn could yield different trajectories during later toddlerhood or the preschool years. Indeed, this pattern would be similar to that which has been uncovered by early childhood executive function researchers. For example, despite evidence that inhibitory control and working memory emerge before cognitive flexibility, a unitary factor model has been identified in children ages 3 years and under, whereas a two- or three-factor model has been identified in later stages of development (Lehto et al., Reference Lehto, Juujärvi, Kooistra and Pulkkinen2003; Wiebe et al., Reference Wiebe, Espy and Charak2008, Reference Wiebe, Sheffield, Nelson, Clark, Chevalier and Espy2011).

One limitation of the current approach was the use of a parent-report form to capture ISL knowledge. Although the MCDI has been used in previous ISL research with this specific word list (e.g., Lamb, Reference Lamb1991; Longobardi et al., Reference Longobardi, Spataro and Colonnesi2018; Raines, Reference Raines2014), it does have noted limitations. First, a few words have multiple meanings, such as “can” (“Can I have this?” is an ISL term, but a “tinned can” is not), and these varied meanings are not captured on the parent-report form. However, due to the observed results of a unitary factor structure, there is greater confidence that these few, polysemous words did not change or impact our results. Second, the MCDI comprises a varied number of words from different proposed ISL categories. The large majority of words are dispositional or perceptual; there are only a few volition and cognitive words such as “need” and “want”. Interestingly, these items had the weakest factor loading (although these values were still quite high and well-above threshold,. 86-.87). Future research may wish to use other vocabulary measures with a greater variety of volition and cognitive terms to replicate the current results.

Aim 2. Longitudinal development of ISL

The second aim of the current study sought to evaluate the individual predictors of ISL growth given the unitary factor structure of ISL observed via Aim 1. The results from our best-fitting model (Model C) suggested that right around the second postnatal year (23.5-months, intercept), children are producing an average of 13 ISL terms of the available 43 coded from the MCDI. However, the number of ISL terms children produced at this age varied considerably as a function of both sex and maternal education. Female toddlers had higher ISL production than males at age 2 by approximately three ISL terms, and toddlers with college-educated mothers had higher ISL production than toddlers of a mother without a college education by approximately two ISL terms. Consequently, the greatest difference emerged at age 2 between male toddlers of mothers without a college-education (~13 ISL terms) and female toddlers with college-educated mothers (~17 ISL terms). Furthermore, we illustrated that children are producing approximately 5 new ISL terms every two months across early childhood, and that this growth occurs at a linear rate in early childhood. It is worth noting that we inspected the data for non-linear patterns of growth; however, the linear model emerged as the best-fitting model as evidenced by non-significant quadratic terms and decreased model fit. Future research may wish to evaluate more complex patterns of growth via piecewise linear modeling or the use of linear splines.

Maternal education also emerged as a significant predictor of ISL growth (slope). These results suggested that college-educated mothers reported their toddlers as having faster rates of growth in ISL terms compared to less than college-educated mothers. That is, by 30-months of age, female toddlers of mothers without a college degree produced an average of 30 ISL terms whereas female toddlers of mothers with a college degree produced an average of 35 ISL terms. Interestingly, sex did not significantly predict the rate of growth in ISL, meaning males and females experienced similar growth trajectories of ISL over time. Given the significant difference in intercept, however, females in this sample maintained their advantage in ISL production across this developmental period compared to males. Of note, when comparing the model fit indices, adding these two individual-level predictors significantly improved the overall fit of the model, suggesting that child sex and maternal education are significant factors to consider when evaluating between-person differences in ISL development.

Related to the significant sex findings, our results replicate previous literature suggesting a female advantage in ISL and overall language during this developmental period (Hughes & Dunn, Reference Hughes and Dunn2002; Kristen et al., Reference Kristen, Chiarella, Sodian, Aureli, Genco and Poulin-Dubois2014). However, it is important to keep in mind that this is simply a snapshot of a 14-month period, wherein females show this ISL advantage. Indeed, the use of a narrow age range may be artificially highlighting a sex difference that ebbs overtime. In a systematic review and meta-analysis (Etchell et al., Reference Etchell, Adhikari, Weinberg, Choo, Garnett, Chow and Chang2018), sex differences were most apparent in language when using tight age ranges, suggesting sex differences may be most evident at certain developmental stages and negligible in others, likely due to differential rates of maturation. Certainly, these sex differences may be transient or specific to this developmental period, as previous research suggests no significant sex differences in ISL production from 28-months (Bretherton & Beeghly, Reference Bretherton and Beeghly1982) through age 5 (Jenkins et al., Reference Jenkins, Turrell, Kogushi, Lollis and Ross2003).

Taking a developmental approach, one should consider the role of equifinality, wherein with different starting points, groups may emerge at the same endpoint. Previous research has underscored the equifinality of ISL development across males and females when using extended age ranges. Cervantes and Callanan (Reference Cervantes and Callanan1998) found that 2-year-old girls talked more frequently about emotions than boys with their mothers, but by age 4, boys had increased their frequency of emotion talk and no sex differences were evident. More recent research has supported this finding, suggesting that sex differences in language competence decreased from ages 3-6 years, with males showing greater variance in language outcomes than females (Lange et al., Reference Lange, Euler and Zaretsky2016). Despite previous research arguing for equifinality in ISL development as age increases, we did not see any trend to suggest this, as males and females did not have significantly different rates of growth. The lack of different growth trajectories suggests that during this developmental period, females are showing a clear advantage at each age point. Future research could continue to extend these growth trajectories beyond 30-months, as children were not yet at ceiling on ISL, to evaluate if males and females reach equifinality in their development. Extending these growth trajectories may reveal greater insights into the sex differences of longitudinal ISL development.

Although sex emerged as a significant predictor of intercept, maternal education was the only factor that significantly predicted intercept and slope, or growth. These findings replicate previous literature highlighting the connection between maternal education and child ISL production across developmental periods (Dunn et al., Reference Dunn, Brown, Slomkowski, Tesla and Youngblade1991; LaBounty et al., Reference LaBounty, Wellman, Olson, Lagattuta and Liu2008; Longobardi et al., Reference Longobardi, Spataro and Colonnesi2018; Roger et al., Reference Roger, Rinaldi and Howe2012), which has been postulated to be mediated by maternal ISL production (Adrian et al., Reference Adrian, Clemente, Villanueva and Rieffe2005). Maternal education, coded here dichotomously around a college education, is a coarse proxy for access to resources or maternal language production, with ample room for future research to investigate the mechanism behind this relationship, including investigating maternal ISL production, book-reading practices, or the gendered relationship between parent and child ISL discourse. Finally, in our longitudinal analyses, maternal education was entered as a fixed characteristic (L2 predictor) based on the data that were available. As many mothers will pursue continued education after the transition to motherhood (Augustine, Reference Augustine2016), future research should take a more dynamic approach by treating maternal education as a time-varying covariate to better capture changes in mothers’ education status as they may relate to early ISL development.

Integrative summary

Combining the results of Study 1 and Study 2, our findings suggest a unitary structure of ISL between the ages of 16- to 30-months in terms of expressive vocabulary. This unitary structure develops linearly during this developmental period, with variations in start point and growth due to infant factors (sex) and maternal factors (education). Results of our hierarchical linear modeling suggests that females produce more ISL terms than males, with linear growth that continues to be significantly higher in females through the second postnatal year. Toddlers of college-educated mothers also present with a higher ISL vocabulary at the end of the second postnatal year (23mos), with faster rates of growth over time compared to mothers without a college education. These results are specific to the demographics of the current sample, including monolingual, English-speaking toddlers who are typically-developing between the ages of 16- and 30-months. Despite this specificity, these results have practical and theoretical implications. Practically, these results provide age-normed insights for expected ISL production that can be assessed using a common parent-reported measure, the MCDI, with anticipated averages around 13 ISL terms at 23-months of age. These results also provide data to suggest that certain demographics may show differential rates of growth and development in this domain. Theoretically, these results are the first to lend empirical support to the idea that ISL should be considered a unitary construct at this age and does not require further binning into different term types, such as physiological, cognitive, or emotional. Future research avenues are numerous and include multiple ideas for generalizing these findings further, including to novel groups of non-English speakers, multilingual households, and neurodiverse toddlers. Future work could also extend these findings into later toddlerhood and early preschool years to evaluate the multi- or equi-finality of ISL development by sex, evaluate other predictors of ISL development such as birth order, or relate these findings to other domains of interest such as theory of mind and social skills in childhood.