Longitudinal Study of Language of Twins at Ages 9 and 14 Years: Twinning, Zygosity, and Heritability

Abstract

Purpose:

This unprecedented longitudinal twin study focused on the arc of language acquisition from first words to adolescence, with data collection at 2, 4, and 6 years of age, reported in four previous studies and now new data at ages 9 and 14 years.

Method:

The method is a classic twin study design for estimating inherited versus environmental effects. The sample consisted of 1,188 children from 594 twin pairs, ascertained from birth records and followed for studying language and cognitive development. Here, we report language and nonverbal cognitive phenotypes at ages 9 and 14 years. Previous research reports from the sample at younger ages, as well as data from cross-sectional samples in other studies, supported the following generalizations: (a) Twinning effects persisted into adolescence for young children, such that twin children's language milestones were delayed relative to norms for singleton children; (b) twin type, that is, zygosity effects persisted although perhaps reduced with age; (c) gender also played a role for young children with advantages for girls relative to boys; and (d) heritability increased with age.

Results:

As in earlier studies, twins' standard scores for language were below the expected mean scores of 100 provided by test developers, 97 (SD = 14) at age 9 years and 95 (SD = 15) at age 14 years, although the gap relative to age peers is closing. Twin type effects, in which there was a slight disadvantage for monozygotic twins at age 4 years, disappeared at ages 9 and 14 years. Gender was significant at ages 9 and 14 years for vocabulary only but in the opposite direction than expected—with boys outperforming girls; however, this is consistent with singleton longitudinal findings from our lab. Models of heritability included adjustments for gender effects and other covariates. Heritability estimates at ages 9 and 14 years did not significantly differ. The final heritability estimates at age 14 years were .49, .71, and .85 for vocabulary, omnibus language, and nonverbal intelligence quotient (IQ), respectively, suggesting greater environmental effects for vocabulary relative to language and nonverbal IQ.

Conclusion:

The findings are consistent with previous reports, offering further empirical precision in estimation of the role of genetics in the transition from childhood to adolescence and further support for the need to consider genetics as well as the home environment in the lives of twin children.

Genetic studies of child language have increased significantly in recent years as new clinical methods were developed. New molecular technologies and analytic methods are available for genetic studies of typically developing populations as well as clinical conditions including speech and language disorders (Andres et al., 2023; Fisher, 2025; Yousaf et al., 2024). Large-scale genomic studies require very large samples for inference of potential associations with target genetic loci. Such methods are dependent on reliable phenotypic measures across a wide range of children, such as questionnaires or behavioral measures or sometimes single children with interesting clinical profiles (Soblet et al., 2018). For example, in a recent genome-wide association study that examined reading- and language-related outcomes from school-age children from 12 different countries, there was significant single-nucleotide polymorphism heritability in the range of 13% for nonword repetition to 26% for nonword reading (Eising et al., 2022). Such outcomes are encouraging and at the same time incomplete.

Alternatively, twin studies provide a natural experimental study, involving comparisons of identical twins with mostly shared DNA and fraternal twins with sibling-level DNA similarity. It is then possible to estimate shared genetic effects as well as shared versus unshared environmental effects. Studies of twin children can provide estimates of inherited mechanisms at play as children age, from preschool into adolescence. Children's language is a fascinating challenge for models of the human language capacity and how it unfolds over time, with or without direct instruction from adults, across the various languages and communities around the world. A core question is the extent to which this language capacity is “hard wired” in some way in human inheritance to facilitate effortless acquisition in young humans across child-rearing environments.

Twin studies have provided a powerful natural research design in the DNA comparison of two kinds of twin pairs: monozygotic (MZ; “identical”) versus dizygotic (DZ; “fraternal”) twins. The home environment and the DNA are shared for MZ twins although the pattern differs for DZ twins who also share home environments, but the DZ twins do not share the same DNA. Comparisons of the heritability of a given trait in MZ versus DZ twins are a widely accepted research design for inferring possible genetic effects across the full age span.

Although most of the literature on twin language development (Stromswold, 2001) reports on cross-sectional studies comparing MZ versus DZ twin pairs, longitudinal studies have revealed that such twinning effects seem to vary across age levels and across phenotypes, that is, the trait assumed to be under genetic guidance. Children's language acquisition moves from a very limited spoken language repertoire early on, followed by a relatively rapid emergence of more complex language as children change in other ways. The linguistic details shift as children progress toward their adult grammar and vocabulary levels, leaving open the need to document multiple measures of language, benchmarked to age-level expectations, for possible genetic effects. The features of language studied vary from one twin study to the other. Furthermore, heritability effects for young children on early language phenotypes are weakest relative to subsequent measurements as children age. Thus, phenotypes must be age appropriate to sort out age-related heritability effects independently from zygosity and gender effects over time.

The following sections review a series of four interrelated longitudinal language studies of twins recruited from shared health systems in Perth, Australia. The overall study was designed to capture shifts in the strength of language heritability in twins over time, at target ages 24 months, 4 years, 6 years, 9 years, and 14 years of age for the full sample and at 4 and 6 years of age focusing on children with specific language impairment (SLI) and nonspecific language impairment (NLI; Rice, 2020). Figure 1 displays the sample size per study wave, retention percentages from 24-month twins onward, and publication dates for papers published on the sample at 2, 4, and 6 years. Twins were recruited from birth notifications in Western Australia during the time periods of 1997–1998 and 2000–2003 and were assessed via parent questionnaire when the twins were 24 months of age. At 4–6 years of age, additional twins were recruited from the same birth cohorts (i.e., birth years) as the 24-month twins. Longitudinal (in-person) data collection was then completed on the full sample of twins at ages 4, 6, 9, and 14 years of age. Retention of the participants across rounds of measurement was relatively high (87%). Also included in this background section of the article is a study carried out in a sample of twins in England at age 16 years with an experimental test of finiteness marking adapted for the study reported below (Dale et al., 2018).

Figure 1.

Australia twins study participants from 24 months to 14 years of age.

Summary of Related Studies

Longitudinal Language Heritability Estimates at 24 Months, 4 Years, and 6 Years of Age

A sample of nine hundred forty-six 24-month-old monolingual English-speaking twin children (473 twin pairs) participated in the first round of data collection via postal questionnaire (Rice et al., 2014). Subsequent rounds of data collection were collected in person. For the 2-year-olds, mothers provided reports of vocabulary and grammar phenotypes, which were obtained from the MacArthur Communicative Development Inventories: Words and Sentences (Fenson et al., 2006). Phenotypes were (a) number of words produced, (b) percentile words produced, (c) proportion of children with late language emergence (LLE), and (d) early use of word combinations. Here, we highlight key outcomes. Each of the first three measures showed significant zygosity effects (i.e., differences in twin type, MZ vs. DZ), such that scores were statistically lower for the MZ twins when compared to DZ twins. Furthermore, a consistent trend toward disadvantage for MZ twins was also apparent for word combinations, a level of language that is beginning to appear around 24 months. The “mini finite” early verb tense marking measure (derived from the MacArthur Communicative Development Inventories) showed floor effects. Thus, early in language acquisition, multiple language measures reveal delays in language acquisition for twins, with some areas of language just beginning to be measurable and therefore may not yet be apparent. Significant heritability was detected for vocabulary and grammar phenotypes and for a subset sample of children selected for LLE. Significant heritability estimates ranged from .22 (word combinations) to .52 (grammar phenotypes; Rice et al., 2014). The group differences were consistent with Dale et al. (1998, 2000).

The next rounds of assessment were at age 4 years and again at age 6 years, using the same language tests (phenotypes) designed for this age period. The sample included 96% of the 24-month participant base, continuing into the early school–age sample (Rice et al., 2018). In addition to the children who were measured at 24 months, additional twin pairs were recruited into the 4- and 6-year study. The total number of participants was 1,354, comprising 677 twin pairs (see Figure 1). At each of the two times of measurement, nine phenotypes from the same comprehensive in-person direct behavioral assessment protocol were investigated. The phenotypes were benchmarked to singleton age group expectations. Of interest here are the following language phenotypes: Peabody Picture Vocabulary Test–Third Edition (PPVT-III; Dunn & Dunn, 1997), Test of Oral Language Development–Primary: Third Edition (TOLD-P:3) Spoken Language, TOLD-P:3 Semantics, and TOLD-P:3 Syntax (Newcomer & Hammill, 1997). We also included the summative Composite and Screener scores from the Rice–Wexler Test of Early Grammatical Impairment (TEGI; Rice & Wexler, 2001). The major question of interest was whether the twinning effect, that is, the delays in language acquisition for young twins relative to singleton children, diminishes with age. We evaluated whether twins, regardless of twin type (MZ vs. DZ), showed a twinning effect that leads to lower-than-age expectations of language performance across both types of twins in the 4- to 6-year age level.

The outcomes of interest at ages 4 and 6 years for the full sample are as follows: Children had higher standard scores at age 6 years than at age 4 years for all phenotypes, indicating that, as a group, twins were closing the gap with singleton children's age-expected levels of performance. Most phenotypes with standard scores had confidence intervals around 90–100, somewhat lower than age expectations (i.e., the gap did not fully close). From ages 4 to 6 years, heritability estimates increased across all phenotypes and were higher than reported for the children at age 2 years, in the range of .54–.92 at age 6 years. The clinical grammar marker of finiteness yielded the highest heritability score (.92).

The next study of the twin samples at 4 and 6 years of age focused on subgroups of children with SLI or NLI (Rice et al., 2020) to examine heritability of lower levels of performance. Groupings of the children were defined by levels of performance on standardized language and nonverbal intelligence quotient (IQ) tasks relative to age peers. The SLI group had low language levels relative to age expectations and nonverbal IQ in normal range or above. The NLI group (a smaller group) had low levels on language assessments as well as on nonverbal IQ. There was increased momentum in twin acquisition of language during this 2-year period, although 4-year-old twins did not catch up to their age peers by 6 years. Low levels of nonverbal IQ did not show the expected patterns of heritability across ages and levels of severity, as evident in many language measures. In short, in this age range, heritability differs by affectedness severity levels per phenotype (across dimensions of language and comparing language to nonverbal IQ) and age differences at 4 and 6 years. Such patterns of outcomes are not consistent with a shared causal pathway for language impairments and low levels of nonverbal cognition.

Cross-Sectional Study of a Sample of 16-Year-Old Twins Participating in the Twins Early Development Study in England

The Rice and Wexler Test of Finiteness Marking (Rice & Wexler, 2001) that was used as a phenotype in the studies of 4- and 6-year-old twins above was adapted for use as a phenotype in the sample of 16-year-old twins in the Twins Early Development Study (TEDS) in England (Dale et al., 2018; https://www.kcl.ac.uk/research/teds-study). Liability threshold estimates of genetic influence on low performance on the task were substantial, as we can see by considering heritability defined by below-normal levels of performance: The lowest 10% of the sample yielded a liability threshold estimate of 36%, whereas the lowest 5% of the sample yielded a liability threshold of 74%. That is, more distance from normative expectations yields higher estimates of heritability.

The group differences were consistent with Dale et al. (1998, 2000) and Harlaar et al. (2007) across different analytic methods. Differences from the studies reported here and the TEDS studies are that all assessments in the studies we have reported from the Australian twin project were conducted in person, with language measures standardized for age levels across the dimensions of vocabulary and morphosyntax, as well as nonverbal cognitive assessments.

Previous Studies, External to Our Lab, in the Age Range of the Present Study

Recent findings from other labs are consistent with the findings from our lab. For example, a 2016 study (Harlaar et al., 2016) used a factor analytic approach on a sample of same-sex twins (aged 10–12 years) to calculate heritability for a “formal language” factor using three subtests from the Clinical Evaluation of Language Fundamentals–Fourth Edition (CELF-4; Semel et al., 2003) plus one test of narrative language. Heritability for the formal language factor ranged from .82 to .86 across three time points. In contrast, a 2017 study (Tosto et al., 2017) on the TEDS sample at ages 7, 12, and 16 years reported heritability of oral language using different measures across age. At age 7 years, oral language was measured via telephone using the Expressive Vocabulary subtest of the WISC-III with an estimated heritability of .27. At ages 12 and 16 years, web-based language measures yielded heritability estimates of .47 and .55, respectively. Such data from adolescent participants are limited in our research literature.

Overall Working Conclusions for Persistently Heritable Effects on Language Acquisition Across Childhood

Encompassing ages 2–16 years, there is consistent support for increased heritability of language acquisition (including morphosyntax) with age, whereas there is less robust evidence for heritability of nonverbal cognition. Over the course of the studies described above, we found evidence suggesting the need to examine the effects of covariates such as maternal education and age, as well as child sex. These conclusions from our earlier studies on our twin sample at earlier ages were further evaluated by the study reported here, focusing on investigation of key language phenotypes for the twin sample reported above, at 9 and 14 years of age.

Goals of the Present Study

As summarized above, a focus on language phenotypes for young children has yielded new insights about how children's language acquisition could be supported by some as-yet-unidentified genetic elements of language change over time. With the support of R01 DC00526, we continued to follow the participants in the study for documentation of longitudinal trends in language acquisition at 9 and 14 years of age. This would bring new data for documenting the transition from early language acquisition with or without language impairments for the full arc of language acquisition, from 2-year-old twins to adolescents. The focus on language brought new measurements for the older children for a better understanding of the continuity of language abilities as well as language delays relative to age peers. The overarching theme was to add to the general knowledge base relevant to a better understanding of individual differences in language outcomes in adolescence as children transition to adulthood.

The research questions (RQs) focused on the period of transition from childhood into adolescence, a period for which there is no continuous longitudinal language information tied to late childhood outcomes. Precise linguistic and nonverbal cognitive/IQ measures from toddlers through early childhood into adolescence and the transition to adulthood are needed. Thus, our RQs targeted gaps in the literature for adolescent twins by following up on a twin sample previously tested at younger ages (i.e., 2, 4, and 6 years of age). Would the patterns be similar across three areas of assessment: receptive vocabulary, spoken language, and nonverbal IQ? Thus, our RQs were as follows:

1. Twinning effects: Do low language scores, relative to age peers, persist to 9 and 14 years of age? How do twins' scores on standardized, age-normed measures of receptive vocabulary, omnibus language, and nonverbal IQ (as phenotypes) compare to those of the general population (i.e., is there evidence of twinning effects at 9 and 14 years of age)?

2. Zygosity effects: Do MZ twins score lower than DZ twins (i.e., a zygosity effect) at 9 and 14 years of age?

3. Predictor/covariate effects: In conditional predictive models, do maternal education, maternal age, and child sex predict the extent of lower-than-age expectations scores for MZ and DZ twins on the phenotypes at 9 and 14 years of age?

4. Heritability: Is there evidence of heritability at 9 and 14 years of age across the three phenotypes? Does heritability increase with age?

Method

Ethics

This study was approved by the University of Kansas Institutional Review Board (12582) and two institutions in Perth, Western Australia: Curtin University of Technology Human Research Ethics Committee (HR3/2001) and the Department of Health Western Australia Human Research Ethnics Committee (2010/6). Participant reimbursement for effort included small toys for children and movie vouchers for adolescents and adults.

Participants

The sample was composed of twins participating in a population-based longitudinal study of child language with assessments at target ages 2, 4, 6, 9, and 14 years. The overall study design was a prospective cohort study of twins born in Western Australia between 1997–1998 and 2000–2003 (see Rice et al., 2014, 2018). The present study reports on twin outcomes at the final two time points of 9 and 14 years of age at the end of the National Institutes of Health funding period. The average age in “years;months” at the target age of 9 years was 9;1 (SD = 10 months, range: 8;0–11;11), and at the target age of 14 years, it was 13;7 (SD = 9 months, range: 12;11–17;6). Figure 2 displays the range of assessment ages within the two target age points; as shown, most children (> 94%) were tested within 1 year of the target ages of 9 or 14 years. The sample included 1,188 children (578 boys and 610 girls) from 594 twin pairs: 93 MZ girls, 93 MZ boys, 104 DZ girls, 89 DZ boys, and 215 DZ opposite sex twin pairs. In terms of pair-level data across age, 536 of 594 pairs (90.2%) had complete data at ages 9 and 14 years, 50 pairs (8.4%) had data at age 9 years only, five pairs (0.8%) had data at age 14 years only, and three pairs (0.5%) had data for both twins at age 9 years but for only one twin at age 14 years. We examined the potential role of missing data on the phenotype outcomes by inspecting the raw (unmodeled) means at age 9 years for subsamples of twins with complete data across age and twins with data at only 9 years of age. Across outcomes, the means at age 9 years were similar across subsamples, and any differences were small (i.e., 1–2 standard score points or 1 percentile point for Raven Nonverbal IQ) and within the mean standard error of measurement per outcome. Thus, we feel that missing data did not compromise the analytic results and conclusions of this article. In total, the final analysis sample included 1,178 and 1,085 individuals at ages 9 and 14 years, respectively—this is the maximum number of children available at each age per phenotype. Given the wide age range, we statistically controlled the exact timing of the measurement occasions in the conditional models, as described below.

Figure 2.

Twin assessment age within target ages of 9 and 14 years.

Details about the participants, study design, sampling, and exclusionary criteria are described extensively in earlier reports on this sample at 24 months (Rice et al., 2014) and 4 and 6 years of age (Rice et al., 2018, 2020). We report here the participant details relevant to this article. The racial composition is nearly identical to those reported previously (i.e., Rice et al., 2014) at 90.24% White, 0% Black, 0.17% Asian, and 9.60% all other/mixed (includes Aboriginal and Torres Strait Islanders). Information on Hispanic ethnicity is not provided here, as it is not included in Australia census calculations and the participants do not define themselves in that way. Regarding exclusionary criteria, the intent was to limit the sample to children without concurrent conditions likely to affect language acquisition. As described previously (Rice et al., 2014, 2018, 2020), twins with exposure to languages other than English were excluded (based on a parent report questionnaire), and pairs in which at least one twin had a hearing impairment, neurological disorders, or developmental disorders were excluded. Exclusionary conditions included Down syndrome, Angelman syndrome, cerebral palsy, cleft lip and/or palate, agenesis of the corpus callosum, and global developmental delay. Children's hearing was assessed at 4, 6, 9, and 14 years of age via pure-tone screenings (500, 1000, 2000, and 4000 Hz) using headphones in everyday ambient noise in field testing. A pass was defined as a participant responding to each frequency in either the right or left ear at 25 or 30 dB. At age 9 years, four participants (0.34%) failed the hearing screen; at age 14 years, five failed (0.46%). All children who failed at age 9 or 14 years but who passed the hearing screen at their other measurement occasion (i.e., at age 9 or 14 years) were included in this study.

Materials, Procedures, and Test Characteristics

Data collection on twins between target ages 9 and 14 years, in person with a trained examiner, took place between July 2008 and May 2017. Throughout data collection, deidentified data were shipped in batches from Australia to Kansas between one and three times per month. Upon arrival, each assessment was checked off for each participant in a central database. Assessments were then scored, checked for scoring accuracy, entered in the software system (IBM SPSS Statistics, Versions 17–25), and checked for entry accuracy. Data were entered within 1–3 weeks of their arrival to avoid falling behind on data processing. The data processing cycle for ages 9 and 14 years began in August of 2008 and ended in August of 2017.

Standardized, age-normed tests of receptive vocabulary, omnibus language, and nonverbal IQ were collected. These included the PPVT-III (Dunn & Dunn, 1997) as an assessment of receptive vocabulary and the CELF-4–Australian Standardized Edition (Semel et al., 2006) Core Language score to assess omnibus language. The CELF-4 Australian was developed as part of the "CELF-4 Australian Standardization Project," conducted by the School of Psychology at the University of Western Sydney on behalf of Harcourt Assessment (see Chapter 2 of the CELF-4 Australian Examiner's Manual; Semel et al., 2006). The CELF-4 Australian was adapted from the US edition of the CELF-4 in order to replace language and stimulus items that were culturally inappropriate in Australia. Norms for the CELF-4 Australian are based on a representative sample of 825 children, adolescents, and young adults that were stratified based on age, gender, parental education, geographic region, and type of schooling according to the 2001 Australian census data. The CELF-4 Core Language score is composed of four clinically sensitive subtests that differ by age group. For ages 9–12 years, the Core Language score includes Concepts and Following Directions, Recalling Sentences, Formulated Sentences, and Word Classes–Total. For ages 13–21 years, the subtests include Recalling Sentences, Formulated Sentences, Word Classes–Total, and Word Definitions. Recall that some children were tested at 8 years of age in this study; these children also received the CELF-4 Core Language score subtests for 9- to 12-year-olds. The age-normed scores from the PPVT-III and CELF-4 are standard scores, with an expected M = 100, SD = 15. We used the basic version of the Raven's Progressive Matrices test, known as the Standard Progressive Matrices (Raven et al., 2000) as a measure of nonverbal IQ, hereafter referred to as “Raven Nonverbal IQ.” Raven Nonverbal IQ scores are provided as percentiles based on norms for children and adolescents in Australia, in which a percentile of 50 corresponds to a standard score of 100.

The CELF-4 Core Language and the Raven Nonverbal IQ were new to the protocol at ages 9 and 14 years. Previously, at ages 4 and 6 years (Rice et al., 2018, 2020), the Columbia Mental Maturity Scale (Burgemeister et al., 1972) was used to measure twins' nonverbal IQ, and the TOLD-P:3 (Newcomer & Hammill, 1997) was used to measure omnibus language. Figure 3 shows the distribution of CELF-4 Core Language score by age; the average standard scores were 97.34 (SD = 13.77) at age 9 years and 94.55 (SD = 14.94) at age 14 years. The decrease in standard score across age was significant using a paired-samples t test corrected for twin dependency, t(1088) = −10.48, p < .0001, d = 0.64. Figure 4 shows the distribution of Raven Nonverbal IQ scores at ages 9 and 14 years. Relative to most other standardized measures, the Raven has a limited score range, such that there are seven possible outcome scores, ranging from the 5th to the 95th percentile, shown in the seven histogram bars in Figure 4. At age 9 years, the average Raven Nonverbal IQ percentile score was in the expected range of 50 (M = 47.42, SD = 28.69), and at age 14 years, it was lower than 50 (M = 36.63, SD = 27.86). The decrease in Raven Nonverbal IQ percentile score from ages 9 to 14 years was also statistically significant: paired-samples t test corrected for twin dependency, t(1118) = −13.32, p < .0001, d = −0.80.

Figure 3.

Distribution of standard scores on Clinical Evaluation of Language Fundamentals–Fourth Edition (CELF-4) Core Language at 9 and 14 years of age.

Figure 4.

Distribution of percentile scores on Raven Nonverbal IQ at 9 and 14 years of age. Raven Nonverbal IQ = Raven Progressive Matrices (Standard Progressive), with Australian norms.

Analytic Strategy

We followed the analytic method used on this twin sample at 4 and 6 years of age, as described in Rice et al. (2018, p. 83), in which general linear mixed-effects models were estimated using maximum likelihood. The significance of fixed effects was evaluated by their Wald test p values, and the significance of random-effects variances and covariances was evaluated via likelihood ratio tests between nested models (i.e., comparing −2 times the differences in model log likelihood to a χ² distribution with degrees of freedom equal to the difference in the number of parameters). Each phenotype outcome was predicted separately in two-level models: Individual twins (Level 1) were nested within twin pairs (Level 2), and their responses at 9 and 14 years of age were predicted simultaneously as correlated multivariate outcomes. Nested model comparisons indicated significant improvements in model fit for most outcomes when allowing heterogeneous variances and covariances for age and zygosity; thus, we used this model consistently across all outcomes to allow for any heterogeneity in variance. This flexible modeling method allows for individual twins in the analysis given that dependency within pairs is accounted for by random-effects variances and covariances (Guo & Wang, 2002); it also does not restrict the sample to only those with outcomes at both ages. Inspection of the Level 1 residuals indicated plausible normality across phenotypes, indicating no need for transformations (Hoffman, 2015; Snijders & Bosker, 2012).

Empty means (i.e., intercept only) models were first estimated to examine the effects of twinning (RQ1) and zygosity (RQ2), followed by conditional models to examine the pair-level and twin-level predictor effects (RQ3). Total R² was calculated for each model as the squared correlation between the predicted fixed effects and obtained outcome values (Hoffman, 2015). These models were all estimated within SAS PROC MIXED software (Version 9.4; SAS Institute). We calculated effect sizes for significant predictor effects using the formulas below (Darlington & Hayes, 2016), in which DF stands for denominator degrees of freedom as estimated using the Satterthwaite method. Cohen's d was used for two-group variables, such as zygosity and child sex, and partial eta (η) was used for continuous variables, such as maternal age. Cohen's d is interpreted as a standardized mean difference metric for grouping variables, whereas partial eta is interpreted as a correlation metric for continuous variables. “Partial” refers to the unique contribution of each predictor after controlling for its overlap with other predictors in the model (which is also the case for Cohen's d).

$$d=\frac{2\times t\text{value}}{\sqrt{\mathrm{DF}}}$$ (1)

$$\mathrm{\eta }=\frac{t\text{value}}{\sqrt{t{\text{value}}^2+\mathrm{DF}}}$$ (2)

In order to obtain intraclass correlations (ICCs) and heritability estimates (RQ4), the same conditional models were then reestimated using structural equation modeling software Mplus Version 8.11 (Muthén & Muthén, 1998–2017), which provided estimates, standard errors, and 95% confidence intervals for these additional effects as combinations of the model-estimated variance components.

Results

RQ1 and RQ2: Unconditional Models, Twinning, and Zygosity Effects at Ages 9 and 14 Years

Table 1 shows the estimated means, standard errors, and 95% confidence intervals for three phenotypes (vocabulary, omnibus language, and nonverbal IQ) at ages 9 and 14 years, separately for MZ and DZ twins, for a total of 12 sample means. To identify twinning effects (RQ1), we must first establish what the expected score would be in the general population. For the PPVT-III Vocabulary and CELF-4 Core Language phenotypes, the expected standard score is 100 (equivalent to the 50th percentile). The Raven is normed in a percentile format in which the expected score in the general population is 50. In the third to last column in Table 1, positive and negative symbols indicate whether the confidence interval range is above (+) or below (−) the expected population mean, in which the latter indicates a twinning effect for that phenotype. For CELF-4 Core Language and Raven Nonverbal IQ, twinning effects were present for seven of eight sample means. For PPVT-III Vocabulary, one sample mean was lower than expected (MZ twins at age 9 years), one was higher than expected (DZ twins at age 14 years), and the remaining two were as expected. The last two columns in Table 1 show two types of significance at the α = .01 level: age differences within zygosity and zygosity differences within age. Regarding differences across age, scores significantly changed across age for all three phenotypes—but not in the same direction. Twins' scores increased with age on PPVT-III Vocabulary but decreased with age on CELF-4 Core Language and Raven Nonverbal IQ. Note that, since standard scores are adjusted for age, a reduction in the mean standard score across age indicates less growth than the average level relative to same-age peers. Regarding zygosity (RQ2), there were no significant differences between MZ and DZ twins across the three phenotypes.

Table 1.

Unconditional models: standard score means, standard errors, and 95% confidence intervals (CIs) by phenotype, age, and zygosity.

Phenotype	Age (years)	Zygosity	n	M	SE	95% CI		CI excludes population mean	Age difference	Zygosity difference
						LL	UL
PPVT-III Vocabulary	9	DZ	807	99.90	0.52	98.87	100.93
		MZ	370	97.64	0.84	95.99	99.28	−
	14	DZ	755	103.95	0.48	103.02	104.89	+	*
		MZ	330	102.57	0.86	100.87	104.27		*
CELF-4 Core Language	9	DZ	805	98.08	0.59	96.91	99.25	−
		MZ	370	95.67	0.97	93.76	97.58	−
	14	DZ	755	95.15	0.64	93.90	96.41	−	*
		MZ	330	92.56	1.11	90.37	94.74	−	*
Raven Nonverbal IQ	9	DZ	806	48.71	1.19	46.37	51.04
		MZ	370	44.47	1.83	40.85	48.08	−
	14	DZ	754	36.84	1.14	34.60	39.07	−	*
		MZ	330	36.04	1.99	32.12	39.96	−	*

Note. PPVT-III Vocabulary and CELF-4 Core Language are standard scores with an expected population mean of 100. Raven Nonverbal IQ score is a percentile score with an expected population mean of 50. LL = lower limit; UL = upper limit; PPVT-III = Peabody Picture Vocabulary Test–Third Edition; DZ = dizygotic; MZ = monozygotic; CELF-4 = Clinical Evaluation of Language Fundamentals–Fourth Edition–Australian Standardized Edition; Raven Nonverbal IQ = Raven Progressive Matrices (Standard Progressive), with Australian norms.

^* p < .01.

RQ 3: Conditional Models

To control for the imbalance in the timing of the measurement occasions, we included linear and quadratic effects of the offset of the actual assessment age from the target age of 9 or 14 years for each twin (offset M = −2.07 months, SD = 10.21, range: −13.15 to 41.87). The ICCs for the age offset variables were ~1.00, indicating that most cotwins were tested within a few days of each other. At age 9 years, 97.5% of twins were tested within 7 days of each other, and at age 14 years, 92.8% of twins were tested within 7 days of each other. A conservative alpha level of .01 was used to determine significance for all fixed effects, as described below.

Conditional models were then specified for each phenotype to examine the effects of predictors at each age. The predictors included zygosity (DZ coded as 0 vs. MZ coded as 1), maternal education (centered at a value of 8 for an education level of trade or apprenticeship), maternal age (centered at age 30 years), child sex (boys coded as 0 vs. girls coded as 1), and linear and quadratic effects of age offset at ages 9 and 14 years. Preliminary analyses examined effects of sex composition of the twin pairs and individual sex, as well as moderation of each effect by zygosity. None of these effects were significant, and so all predictor effects were constrained to be equal across MZ and DZ twins. The intercept was allowed to differ by zygosity to test for differences between MZ and DZ twins; this effect was not significant in any of the conditional models. Results for the predictor effects for each of the three conditional models are shown in Table 2. Significant differences in the unstandardized effects across age are shown in the last column, labeled “AD.” Total R² was similar across phenotypes; from lowest to highest, these were Raven Nonverbal IQ = .088, CELF-4 Core Language = .121, and PPVT-III Vocabulary = .124.

Table 2.

Conditional models: pair-level and twin-level predictor effects by phenotype and age.

Predictors by phenotype	Age 9 years					Age 14 years					AD
	Est	SE	95% CI		p <	Est	SE	95% CI		p <
			LL	UL				LL	UL
PPVT-III Vocabulary (R2 = .124)
DZ (0) vs. MZ (1)	−1.52	0.95	−3.38	0.35	.110	−0.87	0.94	−2.73	0.98	.356
Maternal education	1.40	0.19	1.03	1.78	.001	1.26	0.18	0.90	1.61	.001
Maternal age	0.38	0.09	0.21	0.55	.001	0.28	0.08	0.12	0.44	.001
Boy (0) vs. girl (1)	−1.79	0.65	−3.07	−0.50	.007	−1.90	0.65	−3.17	−0.63	.003
Linear age offset	−0.04	0.08	−0.19	0.12	.632	−0.09	0.04	−0.17	−0.01	.034
Quadratic age offset	0.00	0.00	0.00	0.01	.309	0.01	0.00	0.00	0.01	.028
CELF-4 Core Language (R2 = .121)
DZ (0) vs. MZ (1)	−2.06	1.08	−4.18	0.07	.058	−1.86	1.21	−4.24	0.51	.123
Maternal education	1.63	0.21	1.21	2.05	.001	2.02	0.24	1.56	2.48	.001	*
Maternal age	0.30	0.10	0.11	0.49	.002	0.35	0.11	0.14	0.56	.001
Boy (0) vs. girl (1)	1.23	0.72	−0.18	2.64	.087	1.48	0.80	−0.10	3.05	.067
Linear age offset	−0.39	0.07	−0.53	−0.24	.001	−0.08	0.05	−0.17	0.02	.111	*
Quadratic age offset	0.01	0.00	0.01	0.02	.001	0.00	0.00	−0.01	0.01	.952	*
Raven Nonverbal IQ (R2 = .088)
DZ (0) vs. MZ (1)	−3.91	2.16	−8.16	0.35	.072	0.21	2.23	−4.18	4.59	.927
Maternal education	1.87	0.45	0.99	2.74	.001	2.55	0.43	1.71	3.40	.001
Maternal age	0.49	0.20	0.10	0.88	.013	0.67	0.19	0.29	1.05	.001
Boy (0) vs. girl (1)	1.29	1.64	−1.93	4.51	.433	−0.02	1.66	−3.29	3.24	.989
Linear age offset	−0.25	0.21	−0.66	0.16	.237	−0.24	0.10	−0.45	−0.04	.021
Quadratic age offset	0.00	0.01	−0.02	0.02	.849	0.01	0.01	0.00	0.03	.111

Note. Predictor reference groups and centering points: maternal education (0 = 8; trade/apprenticeship) and maternal age (0 = 30 years). Bold values indicate p < .01. CI = confidence interval; Est = estimate; LL = lower limit; UL = upper limit; AD = age difference; PPVT-III = Peabody Picture Vocabulary Test–Third Edition; DZ = dizygotic; MZ = monozygotic; CELF-4 = Clinical Evaluation of Language Fundamentals–Fourth Edition–Australian Standardized Edition; Raven Nonverbal IQ = Raven Progressive Matrices (Standard Progressive), with Australian norms.

Maternal Education and Maternal Age

Both maternal variables were included as between-pairs (Level 2) predictors, as twins share the same mother. Maternal education was a significant positive predictor of higher scores on all phenotype outcomes at both 9 and 14 years of age, and its effect differed significantly across age for the CELF-4 Core Language outcome. Maternal age was also a significant positive predictor of phenotype scores across outcomes at both ages, except for Raven Nonverbal IQ at age 9 years. For the significant effects of maternal education, effect sizes ranged from η = .17 to η = .34; for maternal age, they ranged from η = .13 to η = .18.

Sex and Age Offset

Child sex was a significant predictor of PPVT-III Vocabulary, in which girls scored significantly lower than boys at ages 9 and 14 years. As shown in Figure 5, girls were predicted to have a vocabulary standard score ~2 points less than boys at both ages; Cohen's d effect size at age 9 years was −0.19, and at age 14 years, it was −0.20. The linear and quadratic effects of age offset were significant only for the CELF-4 Core Language phenotype at age 9 years. In this model, the linear age offset effect was negative (effect size, η = −.21), and the quadratic age offset effect was positive (effect size, η = .14). This resulted in a decelerating negative effect of being older than expected at age 9 years, which was no longer significant at age 14 years. Accordingly, the comparisons across twin age for the linear and quadratic age offset effects were significant, with the age offset effects at age 9 years larger in magnitude than age 14 years.

Figure 5.

Conditional model for Peabody Picture Vocabulary Test–Third Edition (PPVT-III) Vocabulary: effect of boys versus girls at 9 and 14 years. *p < .01.

RQ4: ICCs and Heritability Estimates at Ages 9 and 14 Years

Conditional model ICCs reflecting the total variance attributable to pair-level random intercept variation (i.e., mean differences between twin pairs) in each phenotype by zygosity and age are shown in Table 3. Overall, ICCs differed significantly by zygosity but not by age. ICCs ranged from .25 to .46 for DZ twins and from .59 to .82 to MZ twins; for both twin types, the lower value was for Raven Nonverbal IQ, and the higher value was for CELF-4 Core Language. Consistent with past twin studies, MZ ICCs were significantly higher than DZ ICCs for all comparisons except for PPVT-III Vocabulary at age 9 years. Note that larger differences between MZ and DZ ICCs should translate to higher estimates of heritability (as discussed next).

Table 3.

Conditional models: intraclass correlations by phenotype, age, and zygosity.

Phenotype	Age	Zygosity	Est	SE	95% CI		ZD	AD
					LL	UL
PPVT-III Vocabulary	9	DZ	.44	.05	.35	.53
		MZ	.60	.08	.44	.77
	14	DZ	.38	.05	.28	.48	*
		MZ	.62	.07	.49	.76
CELF-4 Core Language	9	DZ	.46	.04	.39	.54	*
		MZ	.82	.03	.77	.88
	14	DZ	.44	.04	.36	.52	*
		MZ	.80	.03	.74	.86
Raven Nonverbal IQ	9	DZ	.33	.05	.24	.42	*
		MZ	.59	.05	.49	.70
	14	DZ	.25	.05	.15	.35	*
		MZ	.68	.05	.59	.77

Note. Est = estimate; CI = confidence interval; LL = lower limit; UL = upper limit; ZD = zygosity difference; AD = age difference; PPVT-III = Peabody Picture Vocabulary Test–Third Edition; DZ = dizygotic; MZ = monozygotic; CELF-4 = Clinical Evaluation of Language Fundamentals–Fourth Edition–Australian Standardized Edition; Raven Nonverbal IQ = Raven Progressive Matrices (Standard Progressive), with Australian norms.

^* p < .01.

Table 4 displays the results for heritability by age for each phenotype (RQ4) at 9 and 14 years of age. In order to interpret these results within the full developmental context of this longitudinal study, we also provide previously published heritability outcomes at ages 4 and 6 years as a reference (Rice et al., 2018). The variance components from the conditional models were used to calculate the proportion of variance due to heritability (h²), common environment (c²), and unexplained variance (e²) using the formulas provided below.

Table 4.

Conditional models: heritability per phenotype by age.

Phenotype		Age	Heritability (h2)				AD	Common environment (c2)				Unique environment (e2)
			Est	SE	95% CI			Est	SE	95% CI		Est	SE	95% CI
					LL	UL				LL	UL			LL	UL
PPVT-III Vocabulary		4	.20	.13	.00	.45		.38	.09	.20	.56	.42	.05	.32	.51
6	.54	.12	.32	.77		.13	.09	.00	.31	.33	.04	.25	.41
9	.33	.19	.00	.71		.27	.12	.03	.51	.40	.08	.23	.56
14	.49	.17	.15	.83		.13	.12	.00	.37	.38	.07	.24	.52
Omnibus Language	TOLD-P:3	4	.40	.10	.21	.59	*	.34	.08	.19	.50	.26	.03	.19	.32
		6	.70	.09	.52	.88		.11	.08	.00	.27	.19	.03	.14	.24
	CELF-4	9	.72	.10	.53	.91		.10	.08	.06	.26	.18	.03	.12	.23
		14	.71	.10	.50	.91		.09	.09	.09	.26	.20	.03	.14	.27
Nonverbal IQ	CMMS	4	.21	.15	.00	.50		.27	.11	.06	.48	.52	.06	.40	.63
		6	.59	.13	.33	.85		.00	.10	.00	.20	.41	.05	.32	.50
	Raven	9	.53	.14	.25	.80		.07	.11	.00	.27	.41	.05	.30	.51
		14	.85	.14	.58	1.00		.00	.11	.00	.05	.32	.05	.23	.41

Note. All values for ages 4 and 6 years are reprinted from the previously published paper by Rice et al. (2018; see Table 5). Negative c² estimates were replaced with 0. CI = confidence interval; Est = estimate; LL = lower limit (truncated at 0); UL = upper limit (truncated at 1); AD = age difference; PPVT-III = Peabody Picture Vocabulary Test–Third Edition; TOLD-P:3 = Test of Oral Language Development–Primary: Third Edition; CELF-4 = Clinical Evaluation of Language Fundamentals–Fourth Edition–Australian Standardized Edition; Raven Nonverbal IQ = Raven Progressive Matrices (Standard Progressive), with Australian norms; CMMS = Columbia Mental Maturity Scale.

^* p < .01.

$$h^2=2\times \left({\mathrm{ICC}}_{\mathrm{MZ}}–{\mathrm{ICC}}_{\mathrm{DZ}}\right)$$ (3)

$$c^2={\mathrm{ICC}}_{\mathrm{MZ}}–h^2$$ (4)

$$e^2=1–\left(h^2+c^2\right)$$ (5)

Heritability estimates were highest for Raven Nonverbal IQ at age 14 years (h² = .85) and CELF-4 Core Language at ages 9 and 14 years (h²s = .71–.72) and lowest for PPVT-III Vocabulary at age 9 years (h² = .33). Heritability did not significantly differ between age 9 and 14 years for any phenotypes. The highest estimate for common/shared environment was on PPVT-III Vocabulary at age 9 years (c² = .27), and the lowest was for Raven Nonverbal IQ at age 14 years (c² = 0). The c² value of zero is due to large differences in ICCs between MZ and DZ twin pairs, as shown previously in Table 3.

Discussion

This study examined the performance of 9- and 14-year-old twins on age-appropriate phenotypes of receptive vocabulary, omnibus language, and nonverbal IQ. We explored the following: effects of twinning (RQ1) and zygosity (RQ2), effects of predictors on overall levels of performance (RQ3), and heritability of phenotype outcomes (RQ4). Unconditional models for twinning (RQ1) and zygosity (RQ2) effects revealed that twin scores differed across ages 9 and 14 years for all three phenotypes, although the patterns of change were different across the three outcome variables (i.e., PPVT-III scores increased with age, CELF-4 Core Language Score decreased with age, and Raven Nonverbal IQ also decreased with age). These test scores are adjusted for age, suggesting that the twins' PPVT scores gained in average levels relative to their age peers, whereas some of the twins' CELF-4 and Raven scores dropped relative to age peers. These imbalances were controlled in the conditional models, as were zygosity effects. Subsequent analyses of predictors of overall levels of performance (RQ3) were controlled in subsequent heritability estimates of phenotype outcomes (R4).

An interesting and perhaps unexpected outcome is the persistent advantage for boys in vocabulary development at both 9 and 14 years of age. This contrasts with twins' performance at 4 and 6 years of age, in which girls outperformed boys on omnibus language and nonverbal IQ (Rice et al., 2018). However, these findings are consistent with longitudinal reports of vocabulary across age. For example, in a longitudinal paper from our lab examining PPVT scores in children with and without SLI, from ages 2;6 to 21 years (Rice & Hoffman, 2015; see p. 355), gender differences favored girls at ages 3 and 4 years; however, by age 10 years, girls scored significantly lower than boys and also had significantly less positive linear rate of growth.

With regard to RQ4 (heritability), as shown in the bottom two rows for each phenotype in Table 4, the heritability estimates at ages 9 and 14 years were .33 and 49 (receptive vocabulary), .72 and .71 (omnibus language), and .53 and. 85 (nonverbal IQ), respectively. These heritability values replicate other reports in the literature (Harlaar et al., 2016; Tosto et al., 2017) and are relatively high, although lowest for vocabulary, and are consistent with an expectation of higher levels of heritability over childhood. Heritability comparisons across ages 9 and 14 years were not statistically different for any of the phenotype outcomes, suggesting that the contribution of heritability to language- and nonverbal-related constructs levels off at around 9 years of age.

The heritability estimate of .85 for nonverbal IQ at age 14 years is consistent with the well-known “Wilson effect,” based on the research of Ronald Wilson, first reported in 1977 (Wilson, 1977), subsequently in 1978 (Wilson, 1978), and again in 1983 (Wilson, 1983), and as also discussed by Bouchard (2013). Various methods were used in the twin studies, which largely focused on adults, with the conclusion that “ … the heritability of IQ reaches an asymptote at about 0.80 at 18–20 years and continuing at that level well into adulthood.” Furthermore, “ … shared environmental influence decreases across age, approximating about 0.10 at 18–20 years of age and continuing at that level into adulthood” (Bouchard, 2013). These generalizations also hold for the study reported here, as shown in Table 4. Note that in the studies Bouchard cited, most focused on IQ as the phenotype; language was intermingled with other cognitive tasks. As is clear in Table 4, overall language scores yield notably larger heritability estimates than vocabulary scores, which aligns with other replicated findings of vocabulary scores associated highly with mother's education (which is often interpreted as environmental effects).

The study reported here considered the full range of outcomes on the phenotypes of interest, that is, vocabulary, language, and nonverbal IQ, for all children in the sample. Another generalization of interest is the possibility of higher indices of heritability for children at low levels of performance relative to their same-age peers, as reported for this sample of children who were studied when they were younger, that is, 4 and 6 years of age (Rice et al., 2018). A group of particular interest is children with SLI, whose low language levels persist with age but are not likely to be identified for speech pathology services (Rice et al., 2023, 2025). Our earlier published report on this sample of twins when they were 4 and 6 years of age concluded: “ … nonverbal IQ is not on the same causal pathway as language impairments; twinning effects on language acquisition affect classification of 4- and 6-year-old children as SLI and NLI, and heritability is most consistent in the SLI group” (Rice et al., 2020, p. 793). The next steps in this program of research will be to investigate if these patterns persist or change as the children move into adolescence.

Although our analyses on the full sample at ages 9 and 14 years were extensive, they did not illuminate heritability of language and nonverbal IQ for children at relatively low levels of performance. Furthermore, we did not evaluate possible effects generated by children with low nonverbal IQ. In our previous studies of typically developing children (Rice & Hoffman, 2015) and twin children (Rice et al., 2014, 2018, 2020), it was informative to estimate the extent to which overall patterns were influenced by levels of language or nonverbal IQ. We have also conducted studies of children who met the criteria for SLI (Rice et al., 2025). To better understand the outcomes of this study, investigations of potential effects of low language, with or without low nonverbal IQ, are needed.

Data Availability Statement

The data sets analyzed as part of this study are not publicly available due to pending arrangements for deposit in a public repository.

Funding Statement

This study was funded by National Institutes of Health (NIH) Grant R01DC005226, awarded to Mabel L. Rice, Stephen R. Zubrick, and Catherine L. Taylor as principal investigators (PIs). Preparation of this article was supported by NIH Grant R01DC001803, awarded to Mabel L. Rice (PI); NIH Grant P30DC005803, awarded to Mabel L. Rice (PI); NIH Grant R42DC013749, awarded to Mabel L. Rice as co-PI; and NIH Grant P30HD002528, awarded to Mabel L. Rice as affiliated researcher. Stephen R. Zubrick and Catherine L. Taylor were supported by the Australian Research Council Centre of Excellence for Children and Families Over the Life Course (Grant CE140100027). The authors would like to thank the children and families who participated in the study and the staff at the Western Australian Data Linkage Branch and the Maternal and Child Health Unit, as well as members of the Australia research team: Sarah Beveridge-Pearce, Bradley Calamel, Tanya Dickson, Julie Fedele, Lucy Giggs, Antonietta Grant, Jennifer Hafekost, Erika Hagemann, Jessica Hall, Anna Hunt, Alicia Lant, Stephanie McBeath, Megan McClurg, Alani Morgan, Virginia Muniandy, Elke Scheepers, Leanne Scott, Michaela Stone, and Kerry Van de Pol.