Times are displayed in (UTC-05:00) Central Time (US & Canada) Change
About this paper symposium
| Panel information |
|---|
| Panel 11. Language, Communication |
| Paper #1 | |
|---|---|
| Disentangling the Relation between Conversational Turns and Children’s Talkativeness | |
| Author information | Role |
| Alexus Ramirez, University of Maryland, College Park, United States | Presenting author |
| S. Alexa McDorman, University of Maryland, College Park, United States | Non-presenting author |
| Gavkhar Abdurokhmonova, University of Maryland, College Park, United States | Non-presenting author |
| Rachel R. Romeo, University of Maryland, College Park, United States | Non-presenting author |
| Abstract | |
| Back-and-forth verbal interactions, known as conversational turns (CTs), have garnered substantial attention, as more active turn-taking with others has been shown to predict children’s language development over and above the number of words children hear. Many researchers have turned to Language ENvironment Analysis (LENA), a small recording device that captures CTs among other metrics like adult word count and child vocalizations (CVs). Although LENA provides daily averages of these metrics, there is an increasing number of researchers that concentrate on specific times of the day (e.g., dinner time). The current study investigates how various methods of sampling LENA data change the relationship between CTs and CVs. Data were collected as part of a larger study that examined the development of executive functioning. The current dataset comprises only children between the ages of 4- to 7- years (M= 5.69, SD = .70, 45 males) who completed at least one day-long recording (n = 68, Mhours = 30.1, SD = 4.2, Range = 12-20). Most children were typically developing (n = 56, 82.4%) and monolingual (n = 44, 65.7%). Parents identified children’s race as non-White Hispanic (33.8%), White (33.8%), Black (27.9%), and Asian (4.4%). Lastly, a composite variable of SES was created and z-scored using the mother’s education, the father’s education, and the income-to-needs ratio (Range = -1.51–1.95). First, averaged across entire LENA recordings, a linear regression was conducted with the number of average CVs per minute as the predictor and the number of average CTs per minute as the outcome variable. Children’s age and SES were inserted as control variables because they were correlated with the outcome. This model accounted for a significant 68.7% of the variance in children’s exposure to CTs per minute, F(3, 64) = 46.77, p < .001. Specifically, CVs per minute was the most robust factor (β = .86, SE = .02). Next, we considered how the relation between CTs and CVs changes when using a different mode of sampling within the LENA recordings. Thus, we selected separate 15-minute samples with the highest number of CTs (i.e., CT max) and CVs (i.e., CV max) across both days of a child’s recordings. A linear regression was conducted with CV max as the predictor and the CT max as the outcome variable. Children’s developmental status was inserted as a control variable because it was correlated with the outcome. This model accounted for a significant 15.7% of the variance in CT max, F(2, 65) = 7.23, p = .001. Similarly, CV max was the most robust factor (β = .36, SE = .03), but with far less contribution than in the first model. These findings indicate that the relation between the LENA variable may change according to the sampling method of the data. Ultimately, there is no single “correct” sampling method; however, researchers must carefully consider their selection when addressing their research questions. Future work will examine whether these relationships change when randomly sampling, another common way to sample LENA data. | |
| Paper #2 | |
|---|---|
| Complexity in Bilingual Infant-Directed Speech: Establishing Reliable Random Sampling and Segmentation Strategies for Daylong Recordings | |
| Author information | Role |
| Nicola Phillips, McGill University, Canada | Presenting author |
| Jillian Rodin, McGill University, Canada | Non-presenting author |
| Marcos E. Domínguez-Arriola, McGill University, Canada | Non-presenting author |
| Linda Polka, McGill University, Canada | Non-presenting author |
| Abstract | |
|
Introduction: The number of different words (lexical diversity or LD) in infant-directed speech (IDS) has been linked with infant vocabulary growth [e.g., Pan et al., 2005]. Adults vary the LD of IDS depending on children’s ages, with lower LD in the first year of life, giving way to higher LD in the second (Newman et al., 2016; Huttenlocher et al., 2010). Analyzing LD in speech to infants acquiring two languages requires nuanced consideration. For example, in bilingual homes, different adult speakers provide different quantities of input in each language (Orena et al., 2020). This matters because maternal and paternal IDS show LD differences (Pancsofar & Vernon-Feagans 2006; Tamis-LeMonda et al., 2012). Daylong Audio Recordings [DARs] provide large, ecologically valid samples of adult speech in bilingual homes but are unwieldy, triggering difficult decisions about how to reliably sample from them. It is common to extract language samples from portions of DARs with high densities of adult words, but this can introduce systematic biases. Although random sampling can reduce such biases, many LD measures vary depending on sample length [Zenker & Kyle, 2021]. So, in this pilot project we asked, what is the lowest DAR sampling rate for reliably measuring LD and are there LD measures robust to sampling rate? Relatedly, before LD can be measured, adult speech must be segmented into utterances. Prior publications have defined utterance offsets in IDS based on intonational shifts (e.g., Rowe, 2008); such definitions are ambiguous, so we also asked whether fixed durations between utterance offsets/onsets offer higher inter-rater reliability. Methods: We chose a subset of 4 families from a longitudinal DAR corpus of French-English learning infants in Montreal, Canada. Each DAR was split into 30-second chunks containing speech; we randomly sampled these chunks at four different rates (5%, 10%, 15%, 20%), and segmented and transcribed adult speech. Files were randomly assigned to one of three segmentation conditions: intonation-based offset, and 500 ms or 1000 ms offset/onset intervals. At each sampling rate we computed Guiraud’s index, type-token-ratio (TTR), and moving-average TTR (MATTR, across 10-, 50-, and 100-word windows). Hypotheses: (1) LD measures will differ across sampling rates, interacting with language (English and French), speaker type (Mom, Dad, Other Adults), and infant age (10 and 18 months). (2) Guiraud’s Index and MATTR will be more stable across sampling rates than TTR. (3) Using a fixed duration offset/onset definition to segment adult utterances will be more reliable than intonation shift. Results: (1) LD measures differed across sampling rates, with rates <15% showing greater variability than >15%. LD measures were most consistent across infant age. French showed more variability than English and speaker type was most profoundly impacted by sampling rate. (2) TTR decreased systematically with sampling rate, while Guiraud’s Index increased. MATTR10 and MATTR50 were consistent across sampling rates >10%, while MATTR100 showed greater variability across all sampling rates, age, language, and speaker type. (3) Segmenting adult utterances using fixed duration offset/onset intervals led to higher inter-rater reliability compared to variable, intonation-based segmentation (p < . 01). |
|
| Paper #3 | |
|---|---|
| Finding Signal in Noise: Human vs. Machine-learning Approaches to Cleaning Daylong Recordings of Bilingual Preschoolers | |
| Author information | Role |
| Dr. Sarah Surrain, Ph.D., The Children’s Learning Institute at the University of Texas Health Science Center at Houston, United States | Presenting author |
| Kelly Vaughn, The Children’s Learning Institute at the University of Texas Health Science Center at Houston, United States | Non-presenting author |
| Yarián García Alayón, The Children’s Learning Institute at the University of Texas Health Science Center at Houston, United States | Non-presenting author |
| Kimberli Izaguirre, The Children’s Learning Institute at the University of Texas Health Science Center at Houston, United States | Non-presenting author |
| Kevin Rosales, The Children’s Learning Institute at the University of Texas Health Science Center at Houston, United States | Non-presenting author |
| Abstract | |
| Children’s language development is shaped by daily experiences with language(s) in the home. Child-centered daylong recordings can provide more ecologically valid measures of language exposure and use across contexts and caregivers. This is particularly important for bilingual environments, as language practices may be more variable across speakers and activities than in monolingual settings. A challenge in analyzing daylong recordings is whether and how to exclude periods of sleep. Removing these periods should theoretically increase the signal-to-noise ratio, reduce measurement error, and result in stronger correlations with child language skills. Because human coding of full recordings is rarely feasible and parent reports can be incomplete, the current study compares three efficient approaches: 1) including the full recording, 2) relying on parent report alone, and 3) using an automatic sleep classifier tool (Bang et al., 2023). We predicted that estimates of adult word counts (AWC), child vocalizations (CV), and conversational turns (CT) would significantly differ by approach and that the classifier approach would result in stronger correlations with children’s expressive vocabulary skills. We expected that CV would be particularly sensitive to differences in approach, given that children who talk more would be expected to have larger vocabularies in their home language. We draw on a new corpus of daylong recordings (M=8.82 hours, SD=2.22) with 89 Spanish-speaking preschoolers (Mage= 4.80 years, SD=.44) in the greater Houston area. The children were recruited from public bilingual preschool classrooms as part of a larger intervention study. Parents consented to have their child wear a Language ENvironment Analysis (LENA) recorder on one non-school day. Children were assessed in Spanish and English on two different days at their preschool. EOWPVT scores were selected from a larger battery to index expressive skills in each language. LENA-generated measures of AWC, CV, and CT were calculated separately using each approach and converted to hourly averages to account for differences in recording length. Estimates of AWC, CV, and CT were significantly different, yet highly correlated, across the three approaches (r’s 0.926–0.996). LENA measures were significantly, positively correlated with child expressive vocabulary skills in Spanish but not English, as expected. Surprisingly, correlations were stronger for the LENA measures derived from full recordings compared to the other two approaches (see Table 1). To explore this further, 12 children with parent-reported sleep, permission to listen, and EOWPVT scores were selected for additional human coding. Human coders identified boundaries of sleep periods. A hybrid approach was used to exclude segments based on these annotations. In this subset, CV estimates using the hybrid and classifier approaches correlated more strongly with EOWPVT Spanish scores than the other two approaches, as expected (Fig. 1). Our findings suggest that for preschool-aged children, who nap less and may have more self-directed quiet time, using an automated classifier trained on younger children does not improve overall estimates. However, for the small subset that do report sleep, the classifier can help determine which segments to exclude, reducing the burden on parents and human coders. | |
| Paper #4 | |
|---|---|
| Approaches to Predicting Expressive Language Growth in Infants at High and Low Likelihood for Autism | |
| Author information | Role |
| Jennifer Markfeld, Vanderbilt University, Vanderbilt Frist Center for Autism and Innovation, United States | Presenting author |
| Jacob I. Feldman, Vanderbilt Frist Center for Autism and Innovation, Vanderbilt University Medical Center, United States | Non-presenting author |
| Grace Pulliam, Vanderbilt University, Vanderbilt Brain Institute, United States | Non-presenting author |
| Caroline Braun, Boston College, United States | Non-presenting author |
| S. Madison Clark, Vanderbilt University Medical Center, United States | Non-presenting author |
| Ruoxi Guo, Boston College, United States | Non-presenting author |
| Insung Kim, Vanderbilt University Medical Center, United States | Non-presenting author |
| Suzanny Dias Kuhlmann, Franklin and Marshall College, United States | Non-presenting author |
| Kelsea McClurkin, Vanderbilt University, United States | Non-presenting author |
| Catherine Bush, Vanderbilt University Medical Center, United States | Non-presenting author |
| Kristen Bottema-Beutel, Boston College, United States | Non-presenting author |
| Bahar Keçeli-Kaysılı, Vanderbilt University Medical Center, United States | Non-presenting author |
| Tiffany G. Woynaroski, Vanderbilt Frist Center for Autism and Innovation, Vanderbilt University Medical Center, Vanderbilt Kennedy Center, UH Manoa, United States | Non-presenting author |
| Abstract | |
|
Identifying predictors of early language development is a top priority in autism research, as language abilities by age 5 are known to predict long-term outcomes for children on the autism spectrum. Studying early language development is difficult due to the instability of autism diagnoses in infancy and toddlerhood. Researchers now often study infants who have an older diagnosed sibling (i.e., Sibs-autism), as these infants are more likely to have autism and developmental language disorder relative to infants with only non-autistic older siblings (Sibs-NA). This study aims to characterize the early expressive language development of Sibs-autism and Sibs-NA and to assess automated and conventional indices of caregiver talk as putative predictors of expressive language growth across these groups. We examined whether human-coded indices of caregiver talk are superior predictors of children’s language growth as compared to automated indices of caregiver talk from daylong recordings. We recruited 97 Sibs-autism and Sibs-NA, matched on chronological age and sex assigned at birth, for a longitudinal study (see Table). Sibs-autism had at least one older sibling with autism, and Sibs-NA had only non-autistic older siblings and no first-degree relatives with an autism diagnosis. All infants lived in primarily English-speaking households. Infants were seen at age 12-18 months (i.e., Time 1), nine months later, and again around their third birthday. Across timepoints, expressive language was assessed using the Mullen Scales of Early Learning (MSEL). At Time 1, caregiver talk was measured and operationalized in two different ways. Specifically, we collected two daylong audio recordings via Language ENvironment Analysis (LENA) recorders for each participant and two, 15-minute free play sessions for each caregiver–child dyad using a standardized toy set. Adult word count (AWC), averaged across the two days of LENA recordings, was derived via LENA software as an automated index of caregiver talk. A previously developed coding system is currently being utilized to derive an index of caregiver follow-in talk to child bids as averaged across free play sessions; coding will be completed by the time of the conference. Latent growth curves were fit to model growth in expressive language over time across Sibs-autism and Sibs-NA. In preliminary analyses considering only the currently available automated index of caregiver talk, AWC and sibling group were subsequently entered as putative predictors of language growth. Expressive language increased over time for MSEL scores across groups. AWC was a significant predictor of the growth of MSEL scores (p = .008), whereas sibling group was not (p = .140; see Figure). In the model, AWC positively predicted language growth; whereas the nonsignificant estimate for sibling group predicting language growth was such that slopes in Sibs-autism tended to be reduced relative to Sibs-NA. Preliminary results demonstrate that automated indices of caregiver talk predict growth in expressive language over the first 3 years of life in Sibs-autism and Sibs-NA. Additional analyses will be run in advance of the conference to investigate the predictive validity of conventionally coded indices of caregiver talk. Implications for research, theory, measurement, and clinical practice will be discussed. |
|
⇦ Back to session
Methodological Insights for Analyzing Children’s Diverse Language Environments and Development with Daylong Audio Recordings
Submission Type
Paper Symposium
Description
| Session Title | Methodological Insights for Analyzing Children’s Diverse Language Environments and Development with Daylong Audio Recordings |