How early maternal language input varies by race and education and predicts later child language

Studies of language development in young children have consistently shown that mothers with higher incomes, higher education levels, and more prestigious careers (markers of higher socioeconomic status, or SES) appear to use a greater amount of language, more diverse language, and more complex language with their young children than mothers from lower-SES backgrounds (Hart & Risley, 1995; Hoff, 2003; Rowe, 2017). Characteristics of early maternal language input used by higher-SES mothers have been positively associated with children’s concurrent and later language development, especially child vocabulary early and later in early childhood (Hart & Risley, 1995; Hoff, 2003; Rowe, 2017). These studies generally included small samples of mothers and children and were often not representative of the diverse populations of mothers and children in the United States. In addition, many of these maternal language input studies used samples where there was a confound between race and SES, with almost no studies examining maternal language input in diverse rural families where poverty rates are higher and where college graduation rates are lower. Rural children comprise 25% of the school age population in the United States, and they often live in ethnically diverse communities that are further from language-rich resources like libraries and schools (O’Hare, 2009; Vernon-Feagans, Burchinal, & Mokrova, 2015). On the other hand, rural families have more connections to extended family, teachers, and religious organizations that may buffer their children from poorer outcomes, including poorer school achievement (Vernon-Feagans et al., 2015; Vernon-Feagans, Gallagher, & Kainz, 2010). Thus, this group is particularly important to examine.

The present study contributes to the literature on the relations among maternal education, maternal language input, and child language by presenting results from a large representative sample of African American and non-African American children and their mothers living in low-wealth rural areas in the United States. The study examines maternal language input, using a variety of maternal language input variables and measures at four time points across early childhood. This study represents a particularly significant contribution by examining race differences within maternal education levels on important maternal language input variables and child language outcomes, as no other study has been able to do. In addition, this study explores how a variety of maternal language input variables in early childhood might mediate the relationship between maternal education and later child language at school age as well as whether the mediation models are similar for both African American and non-African American mothers and children.

SES, Race, and Maternal Language Input

The landmark Hart and Risley (1995) study was likely the most influential study of child language development related to SES. Although this study used a small sample of 42 children and their mothers, it was very focused on the language interactions between the mothers and their children over early childhood. The authors recorded monthly hour-long sessions in the home where the maternal language input data were collected over 21/2 years, starting when the children were about 9 months of age. The detailed language transcripts obtained from these language interactions made the Hart and Risley study one of the most intensive studies of early language input in a diverse sample of young children. Hart and Risley used a measure of occupational prestige as their measure of SES, with the finding that mothers who had high occupational prestige provided more language input, and especially more words, to their young children than mothers who had low occupational prestige. Extrapolating their results over 21/2 years to the sheer amount of maternal words used over the entire early childhood period suggested that children of professional families heard 30 million more words than children whose families were on welfare. This SES difference in children’s language experience in early childhood has now been called the “word gap.” Further, the SES differences in maternal language input in the Hart and Risley study translated into better child vocabulary and literacy skills years later for children of professional families as compared to children whose families were on welfare or had lower-prestige professions (Walker, Greenwood, Hart, & Carta, 1994). Yet, all six families in the study who were on welfare, which was the lowest occupational prestige category, were African American. In contrast, only one of the 13 families in the highest occupational prestige category was African American. These sample differences highlight the confound between poverty and race in the United States, where many more African American families live in poverty compared to other families.

Although the Hart and Risley (1995) study was not interested in race differences in maternal language input or child language, the unexplored confound between race and SES in the study, as well as in other studies (Dollaghan, Campbell, Paradise, Feldman, Janosky, Pitcairn & Kurs-Lasky,1999; Huttenlocher, Waterfall, Vasilyeva, Vevea, & Hedges, 2010), may have led to the implication that African American mothers talk less with their young children (Brooks-Gunn & Markman, 2005; Tamis-LeMonda, Braumwell, Cristofaro, 2012). For instance, Brooks-Gunn and Markman (2005) stated in their review of ethnicity and parenting: “Most striking are differences in language use. Black and Hispanic mothers talk less with their young children than do white mothers” (p. 139). However, this conclusion was based largely on reviewing studies that confounded race and SES (Hart & Risley, 1995; Huttenlocher et al., 2010). Although these previous studies did not mention race differences in maternal language input, the confound between race and SES in their samples resulted in most of the mothers in the low education/low SES groups being African American. Although these studies did find that lower SES mothers spoke less and less complexly to their children, it was not investigated as to whether that difference was due to race or SES.

More recent studies have tried to disentangle race and SES, but have not always had samples that actually collected maternal language input. Hindman et al. (2014) examined coded exchanges between mothers and preschool-aged children, but not actual language transcripts, in the large ECLS-B sample. This study found only small associations between maternal communicative exchanges, home language, and ethnicity. Instead, they uncovered a strong relation between maternal communicative exchanges and maternal education. Tamis-LeMonda et al. (2012), in more complicated findings, reported there were no race differences in the total amount of language mothers used with their 14-month and then 2-year-old children. Yet, African American mothers used comparatively less regulatory language (utterances that directed infants’ attention or corrected infants’ actions) and more referential language (utterances that provided or elicited information, such as the use of questions) compared to Mexican and Dominican mothers. Further, referential language, but not regulatory language, was related to children’s better language skills at 2 years of age. In an older but important study (Anderson-Yockel & Haynes, 1994), ten working-class African American mothers and ten working class White mothers were matched on occupation, education, and income. Maternal language to their 18 to 30 month old child was obtained during a book reading task, including measures of labeling, giving feedback, providing attentional cues, asking questions, and using directives. There were no significant differences between the African American and White mothers on any measures except that White mothers asked more questions than African American mothers. Yet, this finding has not been supported in more recent studies of African American mothers, where African American low-income mothers asked more wh questions than White and Latino mothers when their children were 36 months of age (Cristofaro & Tamis-LeMonda, 2011). Thus, overall, there are still questions about whether there are race differences in maternal language input, or whether the reported differences may be driven by SES-race confounds. Larger more representative samples like the one used in the current study may be able to help uncover if race and/or maternal education differences exist in maternal language input as well as whether maternal language input is related to children’s later language at school age.

Maternal Education and Maternal Language Input

Particularly important in many of these studies of maternal language input has been the use of maternal education as the most important SES marker variable related to mothers’ use of language with their children (Hoff, 2003, 2013). For instance, Hindman et al. (2014) suggested that maternal “education might be the key driver of how SES relates to book-related talk” (p. 306). Other studies have found that more-educated mothers have a larger vocabulary, better knowledge of child development, and more engaging parenting styles when their children were toddlers (Bornstein, Hayes, & Painter, 1998; Hoff, 2003; Rowe, 2008; Rowe, Denmark, Harden, & Stapleton, 2016; Vernon-Feagans, Cox, & Family Life Project Key Investigators, 2013). In turn, maternal language of mothers with more education has often been positively related to their children’s language skills in early childhood (e.g., Vernon-Feagans et al., 2015).

The exact reason why maternal education has been related to maternal language and children’s early language and later achievement has been hypothesized to be related to a number of factors. Maternal education is likely a social address variable (Bronfenbrenner & Evans, 2000) that stands for important parenting behaviors that maximize children’s early language and achievement. Similarly, because maternal education is related to income, it likely bestows resources in the home and elsewhere that can foster children’s optimal development. In addition, higher levels of education have been related to maternal language increases in the lexical diversity and grammatical complexity of their speech. Hoff-Ginsberg has argued that maternal education affects maternal language to their children because it likely impacts mothers’ verbal style that results in more language and more complex language to their children (Hoff-Ginsberg, 1991; Hoff, Burridge, Ribot & Giguere, 2017). Accordingly, Hoff-Ginsberg (1991) found SES differences in mothers’ talk with adults that paralleled the differences in their talk with their children. A somewhat different argument has been elaborated in more recent work by Rowe (2008) that suggested that mothers’ knowledge of child development fully mediated the relationship between maternal education and toddler child language, suggesting that mothers’ knowledge of the importance of talk with children may lead to better language interactions with her own children. Thus, maternal education may be a more pervasive marker of parenting that is related to the way mothers verbally interact with their young children. Maternal education was used in the current study as the measure of SES, with controls for other SES-related variables such as income.

SES, Race, and Child Language

Even though there are still questions about the interplay of maternal language input, race, and SES, considerable evidence has demonstrated that children from low-SES backgrounds score more poorly on standardized language measures and tests as early as three years of age (Golinkoff, Hoff, Rowe, Tamis-LeMonda, & Hirsch Pasek, 2018). In addition, no matter what the SES is of the child, there appear to be child race differences on standardized language measures, with African American children’s scores lower than other children (Magnuson & Duncan, 2006). These differences between African American low-income children and other children have been found to emerge as early as three years of age, and have even been found even when children are matched on SES (Burchinal et al., 2011; Pungello, Iruka, Dotterer, Mills-Koonce, & Reznick, 2009). In their review and analysis of the Black-White achievement gap on standardized tests, Magnuson and Duncan (2006) found this racial gap persistent and difficult to explain based on SES and other environmental factors.

Importantly, in more naturalistic and ethnographic language studies of low-income African American children in early childhood and during the transition to school (Gardner-Neblett & Iruka, 2015; Gardner-Neblett, Pungello, & Iruka, 2012; Heath, 1983; Sperry, Sperry, & Miller, 2018a, 2018b; Vernon-Feagans, 1996), African American children have been found to be exposed to rich and complex narrative language by both adults and children in their neighborhoods and homes. Some of this language has been called “bystander talk” because it was not directly addressed to the child but was part of the overall verbal environment around the child that likely impacted the child’s language. In fact, African American children have been found to be extremely proficient in complex narrative and storytelling language by school age. In fact, Vernon-Feagans (1996) found that African American boys from poverty backgrounds in kindergarten were far superior in the amount and complexity of their talk used in their neighborhood after school compared to their mostly non-African American higher-SES peers. It was also clear from other ethnographic studies (Heath, 1983; Sperry et al., 2018a) that African American children from North Carolina and Alabama received a greater amount and complexity of language by caregivers (e.g., father, mother, or grandmother) and other adults and children in the naturalistic environment compared to other children (Sperry et al., 2018b). Gardner-Neblett and colleagues, in a series of papers and reviews (Gardner-Neblett & Iruka, 2015; Gardner-Neblett et al., 2012), have found that African American children show strong early narrative skills that are predictive of later reading in school.

Unfortunately, there has not been good evidence whether the complex and diverse language exposure in more naturalistic environments experienced by many African American children, as well as these children’s highly developed early narrative skills, have translated into better later language scores and/or success in school compared to their peers. In the Vernon-Feagans’ study (1996), African American children’s experiences with rich language environments and stronger storytelling skills did not appear to translate into better achievement in school. Thus, there is still a need to clarify the issue of race and SES with respect to maternal language input and child language in more ecologically valid settings as is attempted in this study. Critically important is to better understand whether maternal language input by African American mothers is different from maternal language input from mothers of the same SES who are not African American, and whether their language variability might be implicated in their children’s performance on language measures.

Which Maternal Language Input Measures Are Related to SES and Child Language?

Although frequently not addressing SES-race confounds in maternal language input, a variety of studies have now replicated the important Hart and Risley (1995) SES differences on maternal language input. Further, many studies have expanded the scope of maternal language input constructs to include other types of maternal language measures that appear to be more related to child language than maternal total words, as was used by Hart and Risley. Many studies have shown that the number of different words (a measure of vocabulary diversity) and mean length of utterance (a measure of grammatical complexity) as well as related complexity measures differentiated mothers of different SES backgrounds. These measures were better predictors of children’s later language than maternal total words (Hoff, 2003; Huttenlocher et al., 2010; Rowe, 2012; Rowe, Raudenbush, & Goldin-Meadow, 2012; Vernon-Feagans et al., 2013; Vernon-Feagans et al., 2008).

More recently, a number of additional maternal language measures have been found to be important constructs in relation to SES and to later child language performance. One such variable has been wh-questions, which likely elicit a greater amount of language and more complex language from the child. In a study of low-income mothers, maternal wh-questions with their children at 36 months was related to the children’s prekindergarten (pre-k) performance in language (Cristofaro & Tamis-LeMonda, 2011). In a study of both mothers and fathers, mother and father wh-questions at and before 36 months of age were related to school readiness in kindergarten (Reynolds, Vernon-Feagans, Bratsch-Hines, & Baker, 2018). In another recent study, fathers’ use of wh-questions with their 24 month child was related to child language one year later (Rowe, Leech, & Cabrera, 2017). Another construct of recent interest is the number of conversational turns, which measures the number of back and forth language exchanges between the mother and child (Romeo et al., 2018). This language construct in toddlerhood and later early childhood has been argued to be an important indicator of the quality of mother-child language interactions (Hirsh-Pasek et al., 2015, Romeo et al,., 2018). Measures of conversational turns have now been linked to later child preschool language and academic achievement, but these initial studies were not able to measure the construct with actual language transcripts but rather through coding of mother-child interactions (Hirsh-Pasek et al., 2015) or through a specialized listening device called LENA left in the home for long periods of time (Gilkerson, Richards, Warren, Oller, Russo & Vohr, 2018; Romeo et al,,2018). Thus, the current study added these two additional measures of maternal language (wh-questions and conversational turns) as well as complex conjunctions to examine their predictive power in addition to more traditional measures of maternal language input, like vocabulary diversity and grammatical complexity measures.

Maternal Language Input as a Possible Mediator between SES and Child Language

A growing literature has examined whether maternal language input might be an important mediator of the relationship between maternal education/SES and child language skills (Hoff, 2003, 2013; Rowe et al., 2012). Studies have argued that maternal education is related to better child vocabulary and language through the proximal variable of how mothers talk to their children. Mothers are generally considered the most important adult who talks with and shapes children’s language early in life; as such, mothers have been the focus of most of the language input research on children’s early language development (Snow, Burns, & Griffin, 1998; Vernon-Feagans et al., 2013). Hoff and Naigles (2002) presented a social-pragmatic perspective on the role of maternal language input that argues for the importance of recurrent social interactions between mother and child, especially in regular and routine interactions. Bronfenbrenner and Evans (2000) argue for what they call ‘proximal processes’ that are the major drivers of development. Proximal processes are the one-on-one relationships that children have with adults in their microsystems, such as the home. It has been argued that it is the continuous and sustained proximal processes over time that help scaffold optimal developmental processes in childhood. Thus, rich and varied maternal language interactions with young children over time appear to be an important proximal process that promote children’s language development.

A number of studies have examined whether maternal language input fully or partially mediated the relationship between maternal education/SES and preschool child language. Hoff (2003) in a sample of 63 mothers and their toddlers examined the mediating role of maternal language input in the relation between maternal education and child vocabulary growth over a 10-week period. Although Hoff found maternal education differences on almost all the maternal word diversity and language complexity input variables, only maternal mean length of utterance fully mediated the relation between maternal education and child vocabulary. Huttenlocher et al. (2010) examined a combination of maternal education and income as the SES variable to examine child language growth from 14 to 46 months in a sample of 47 mother-child dyads. The study found that SES was a strong predictor of maternal language input and child language, whereas maternal language diversity and complexity were only partial mediators of the relationship between SES and child language growth. Rowe et al. (2012) partially replicated the Huttenlocher study. Her study demonstrated that maternal word diversity was a marginal mediator of the relationship between SES and child vocabulary on a standardized vocabulary test in a sample of 62 children at 18, 30, and 42 months followed for one year. Thus, maternal SES was still a strong predictor of children’s language in both the Huttenlocher et al. and Rowe et al. studies, even in the presence of maternal language input. Since there is still not consensus about which maternal language input variables are the most important mediators of the relationship between maternal SES and later child language, Rowe and colleagues called for further investigation of these relationships in a larger sample. There is also not conclusive evidence that early maternal language remains an important mediator of children’s language when children transition into elementary school.

The Current Study

The Family Life Project (FLP) has published a series of articles on maternal language input in a representative sample of 1,292 children followed from birth who were living in rural communities in the US. FLP oversampled for families in poverty and African American families. Previous FLP papers have focused on the early childhood period up to three years of age. For instance, maternal language composites were partial mediators of the relation between a cumulative risk measure and child language at 3 years of age and maternal education was a strong predictor of maternal language input in infancy (Vernon-Feagans et al., 2008, 2013, 2015). Overall, these previous studies have found measures of SES and risk measures predicted early maternal language input and/or child language.

The current study contributes to the understanding of race and education differences in maternal and child language that have not been reported in any other previous FLP papers. Importantly, the large FLP sample afforded us the opportunity to descriptively disentangle race and SES on our variables of interest. We first examined race differences (African American, non-African American) within maternal education levels (high school degree or less vs. more than high school degree) on both maternal language and school age child language variables. We also were able to examine if there were overall maternal education differences in maternal language input and child language. Our larger goal was to examine whether maternal language early in a child’s life, from 6–36 months, mediated the relationship between maternal education and later child language, during pre-k and kindergarten, which has not been examined in previous studies as well as whether there were different mediation paths by race. Thus, this study focused on the following research questions, controlling for family income and child IQ as well as a host of demographic variables.

• (1)In a representative sample of families living in high-poverty rural areas in the US, do two measures of maternal language quantity from 6–36 months (number of different words and number of conversational turns), three measures of maternal language complexity from 6–36 months (mean length of utterance, wh-questions and complex conjunctions) and three measures of child language outcomes at school age (teacher ratings and two standardized child word knowledge measures) differ by race within maternal education levels? If not, does maternal education differentiate these maternal language input and child language measures?
• 2)Do these five measures of maternal language quantity and maternal language complexity from 6–36 months mediate the relationship between maternal education and child language at pre-k and kindergarten, and do these mediation models differ by race?

Given our overall method and previous studies, we did not believe we would find race differences within education levels, but we did believe that mothers with more education would use better maternal language quantity and complexity than mothers with less education. It was also predicted that maternal education would be a powerful driver of both early maternal language and child language at school age and that early maternal language input would be at least a partial mediator of the relationship between maternal education and child later language at school age. We did not have any evidence to hypothesize that mediation models would differ by race.

Method

Participants

Data for the current study were drawn from the Family Life Project (FLP), a representative sample of every baby born to a mother who lived in one of six high-poverty rural counties (three in North Carolina and three in Pennsylvania) over a one-year period. FLP used a developmental epidemiological design which included daily hospital visits and searches of birth records over a one-year period to identify all births in the target counties. To ensure adequate representation of African American and low-income families in these geographic areas, FLP used a complex sampling design that oversampled for African American families in North Carolina and low-income families in both states (for a detailed description of the sampling design and recruitment strategy, see Vernon-Feagans et al., 2013). The current study used the full sample of FLP children (N = 1,292).

Procedures

FLP data were collected through home, child care, pre-k, and school visits. Trained research assistants, who lived in the same communities as the children and were of the same race as the families, conducted home visits when target children were approximately 2, 6, 15, 24, 36, and 60 months of age. At each home visit, the research assistants interviewed family members, conducted parent and child assessments, administered questionnaires, and video recorded mother-child shared picturebook interactions. At the pre-k visit, the research assistants conducted child assessments in the child’s pre-K setting in the spring before kindergarten. Once children transitioned to school, research assistants conducted child assessments in their schools in the spring of the kindergarten year.

Measures

Maternal language.

To capture maternal language to their child, FLP used a wordless picturebook task during the 6-, 15-, 24-, and 36-month home visits. The mother was asked to sit in a comfortable chair or couch with her child and was given a different wordless picturebook at each age level that was appropriate for that age child. The wordless picturebooks were developed from standard picturebooks, with any and all words removed from the text of the book. These books were piloted for acceptability and interest by both African American and non-African American parents from different economic levels. Pictures in the books were altered to make sure the characters in the books could represent children from a number of different ethnic backgrounds. Books included Baby Faces (Miller, 1998); No, David! (Shannon, 1998); and the Frog and Boy book series (Mayer, 1967; 1969; 1971; 1973). Before the task began, the mother and the child were outfitted with high-quality wireless microphones and the session was video recorded. The mother was told that we were interested in how children and mothers share a book together as they normally would. The mother was first given the picturebook so she could look through it and ask the research assistant any questions she had about the story or the task prior to beginning. The mother was then asked to go through the book with her child and to alert the research assistant when she was finished. The home visitors were told to end the session after about ten minutes if the mother had not signaled she was finished by that point. Thus, the time of the picturebook session varied somewhat but was usually not more than 10 minutes.

The picturebook sessions between the mother and the child were transcribed using the software Systematic Analysis of Language Transcripts (SALT; Miller & Chapman, 1985). A graduate research student and a research associate spent one year learning SALT conventions and developing a training manual in consultation with SALT developers and based on the official SALT training manual (Miller, Long, McKinley, Thormann, Jones, & Nockerts, 2005). Transcribers learned the specific conventions of SALT using this training manual. Training lasted at least three months as transcribers learned the conventions and definitions of codes during transcription. Transcribers completed 20 sessions, which were reviewed by the senior transcriber to ensure accuracy. As an ongoing check, transcripts were regularly reviewed, and any issues were discussed and resolved at weekly research group meetings.

Maternal language quantity.

The first domain of maternal language input was maternal language quantity, which was measured by two variables: number of conversational turns and number of different words. The number of conversational turns was the number of verbal exchanges between the mother and child during the picturebook session. For all descriptive and inferential analyses of number of conversational turns, we averaged values from the 15-, 24-, and 36-month time points because the variable did not adequately represent the concept of conversational turns at the 6-month time point, as infants were not active conversational partners at that time. Number of different words was based on the number of free morphemes (word roots) used across the entire picturebook session in accordance with the definitions provided by the SALT manual guide (Miller et al., 2005). Affixes added to the beginning or end of words were not counted as separate root words. For instance, “talk” and “talked” would be considered the same root word. Omitted and unintelligible words were not included. For all descriptive and inferential analyses of number of different words, we averaged values from the 6-, 15-, 24-, and 36-month time points.

Maternal language complexity.

The second domain of maternal language input was maternal language complexity, which was measured by three maternal language variables: mean length of utterance, number of wh-questions, and complex conjunctions. Mean length of utterance in morphemes was calculated by dividing the total number of morphemes by the total number of utterances. For all descriptive and inferential analyses of mean length of utterance, we averaged values from the 6-, 15-, 24-, and 36-month time points. Number of wh-questions included utterances by the mother that began with “how,” “what,” “whatcha,” “when,” “where,” “which,” “who,” “whose,” or “why” and ended with a question mark. For all descriptive and inferential analyses of number of wh-questions, we averaged values from the 6-, 15-, 24-, and 36-month time points. Complex conjunctions was the number of complex words that signaled some form of embedding of clauses. Words included such conjunctions as “before,” “after,” “while,” “since,” “although,” and because,” but did not include simple conjunctions that signaled compound sentences such as “and” and “but.” These conjunctions together created a marker variable for the number of complex sentences in the picturebook sessions since SALT did not have an option for calculating the number of complex sentences. For all descriptive and inferential analyses of complex conjunctions, we averaged values from the 6-, 15-, 24-, and 36-month time points.

Child language outcomes.

Children’s vocabulary and language skills were measured just prior to school entry and at the end of kindergarten through one teacher-reported measure of child language skills and two standardized measures of child vocabulary and word knowledge.

Teacher-rated child language.

In the spring of pre-k, child language skills were assessed using the Adaptive Language Inventory Scale (ALI; Feagans & Farran, 1979). The ALI measured children’s use of language related to narrative and discourse skills that are needed for learning in the classroom. Pre-k teachers responded to 18 items (e.g., child is easily understood when talking to peers, child talks spontaneously and easily to adults, and child responds to questions asked of him/her in a thoughtful logical way) on a scale from 1 to 5, with 1 = well below average and 5 = well above average. The internal consistency estimate for the ALI was α = 0.98. We used the mean score in analyses.

Child receptive vocabulary.

In the spring of pre-k, child receptive vocabulary was assessed using the Peabody Picture Vocabulary Test, 4^th Edition (PPVT; Dunn & Dunn, 2007). PPVT provides a norm-referenced index of children’s single-word receptive vocabulary knowledge. During the task, children were asked to select the picture considered to best illustrate the meaning of a stimulus word presented orally by the examiner out of four possible answers. Test-retest reliability for the PPVT was .93 and internal consistency estimates averaged between .92 and .97 (Dunn & Dunn, 2007). We used the standard score in analyses.

Child word knowledge.

In the spring of kindergarten, child word knowledge was assessed using the Woodcock Johnson Picture Vocabulary (WJ PV; Woodcock, McGrew, & Mather, 2001). WJ PV provides a norm-referenced index of children’s lexical (word) knowledge. During this task, children were asked to identify pictured objects. Although a few receptive items were offered at the beginning of the test, this subtest was primarily an expressive language task at the single-word level. The items became increasingly difficult as the selected pictures appeared less and less frequently in the environment. WJ PV had a median reliability of .98 for children ages 5–7 (McGrew, Schrank, & Woodcock, 2007). We used the standard score in analyses.

Maternal education.

Maternal education was collected from the 2-, 6-, 15-, 24-, and 36- month home visits. We created two education dummy variables: high school degree or less vs. more than a high school degree. The dummy variables represented the highest level of reported education level between 2 and 36 months. For descriptive analyses, we used both dummy variables. For inferential analyses, we used the dummy variable of more than high school degree, with high school degree or less as the omitted reference group.

Control variables.

A number of child- and family-level control variables were included to account for potential selection bias and other confounding factors. Child-level control variables were reported by the child’s mother upon entry into the study and included the following: gender, race, mental development index, and weekly hours in child care. Gender was coded as 0 = girls, 1 = boys and race as 0 = non-African American [predominantly European American], 1 = African American. The mental development index (MDI) was measured using the Bayley Scales of Infant Development (Bayley, 1993) at the 15-month home visit. The measure included a series of developmental tasks (e.g., retrieves toy, scribbles spontaneously), which were scored by a trained research assistant who gave the child “credit” or “no credit” based on their ability to complete of the task. The internal consistency estimate for the MDI was α = 0.85. We used the standard score in analyses. Weekly hours in child care were collected from the 6-, 15-, 24-, and 36-month home visits. Mothers reported whether their child was in non-parental child care, and for how many hours each week. If a child was not participating in non-parental child care at one or more time points, we assigned them zero hours in child care. We used the average of weekly hours in child care from 6–36 months in analyses. Family-level control variables included state, income-to-needs ratio (INR), and father presence in the home. State (0 = PA; 1 = NC) was included to account for site differences among FLP families. Family INR was measured at the 6-, 15-, 24-, and 36-month home visits, and was reported as the total annual household income for the family divided by the federal poverty threshold for a family of that size and composition. We used the average INR from 6–36 months in analyses. Father presence in the home (0 = present; 1 = not present) was measured at the 6-, 15-, 24-, and 36-month home visits, and was reported by the mothers. Fathers were coded as present in the home if they resided in the home at each time point. These men could be biological fathers who were married to the mother or who resided with the mother.

Analytic Strategy

The first research question asked the following: Do measures of maternal language quantity and maternal language complexity from 6–36 months and measures of child language outcomes at school age differ by race within maternal education levels? If not, does maternal education differentiate these maternal language input and child language outcomes? To examine these differences, we ran two analysis of covariance (ANCOVA) models, which included all control variables described above; number of conversational turns; number of different words, mean length of utterance, wh-questions, and complex conjunctions from 6–36 months; and child language outcomes at pre-k and kindergarten. In the first model, the focal variable was race within maternal education (African American, non-African American; high school degree or less, more than high school degree). In the second model, the focal variable was maternal education (high school degree or less vs. more than high school degree).

The second research question asked the following: Do measures of maternal language quantity and complexity from 6–36 months mediate the relationship between maternal education and child language at pre-k and kindergarten, and do these mediation models differ by race? To examine mediated relations, we created two structural equation models in which the latent child language outcome was regressed separately on the mediating variables comprising maternal language quantity or maternal language complexity. Tests of mediation were conducted by examining indirect effect estimates (MacKinnon, 2008). Child and family covariates were included at each path of the analyses in both models. To examine whether there were race differences in the mediation models, we conducted multiple group analysis in which the structural equation models were estimated for African American and non-African American families.

Prior to mediation analyses, we examined the measurement model for the latent child language construct. This model was just identified, which meant that no degrees of freedom were left to assess its fit to the data. However, all factor loadings and intercepts were statistically significant (p < .001), as were the latent residual variances, implying meaningful variation on the latent construct. Next, we examined the unadjusted association of maternal education with the latent child language variable, which had a value of r = 0.28, p < .001.

All descriptive analyses were run using SAS 9.3, and all inferential analyses were run using Mplus 7. Inferential models used weighting and stratification variables to account for the complex sampling design (oversampling of children from low-income and, in North Carolina, African American households). To account for this sampling design, the MLR estimator was used, which calculated maximum likelihood parameter estimates and standard errors that were applicable to non-independent observations. Model comparisons were computed using a scaled chi-square difference test (Satorra & Bentler, 2010) because of the use of the MLR estimator. To handle missing data (0–25% of outcome variables and 0–15% of predictor variables), full information maximum likelihood estimation (FIML) was employed to reduce potential bias that could result from using listwise deletion (Acock, 2012). Standardized coefficients were reported and model fit was assessed from the chi-square test, root mean square error of approximation (RMSEA), comparative fit index (CFI), and standardized root mean square residual (SRMR).

Results

Demographic Information

Demographic information for the sample is presented in Table 1. Half of the children were boys (51%), and 60% of families lived in North Carolina. Slightly less than half of children were African American (43%), with the remaining children primarily European American. Children averaged 21 hours of nonparental child care per week from 6–36 months. Over half of the mothers had more than a high school degree (59%), and 41% of mothers had a high school degree or less. The income-to-needs ratio from 6–36 months was 1.78, indicating that, on average, families had incomes above the federal poverty threshold of 1.00, but still low enough to qualify for many public assistance programs. Fathers were present in these families 42% of the time from 6–36 months.

Table 1.

Descriptive results (n = 1292)

Variable	M	SD
Control variables
Gender (male)	0.51	0.50
Race (African American)	0.43	0.49
Mental development index (MDI), 15 mos	96.26	10.69
Weekly hours in child care, 6–36 mos	20.83	16.53
State (North Carolina)	0.60	0.49
Income-to-needs ratio (INR), 6–36 mos	1.78	1.48
Father in household (presence), 6–36 mos	0.42	0.49
Maternal education
More than high school degree	0.59	0.49
Maternal language
Maternal language quantity, 6–36 mos
Number of conversational turns	61.29	35.22
Number of different words	89.91	31.03
Maternal language complexity, 6–36 mos
Mean length of utterance	3.39	0.65
Wh-questions	12.04	8.08
Complex conjunctions	7.56	7.80
Child language outcomes
Teacher-rated child language (ALI), pre-k	3.41	0.87
Receptive vocabulary (PPVT), pre-k	93.96	15.80
Word knowledge (WJ PV), kindergarten	99.07	9.97

Note: Number of conversational turns was measured from 15–36 months. ALI = Adaptive Language Inventory. PPVT = Peabody Picture Vocabulary Test. WJ PV = Woodcock Johnson Picture Vocabulary. MDI = mental development index. INR = income-to-needs ratio.

Differences by Race within Maternal Education Levels

Table 2 presents mean differences by race within the two levels of maternal education. There were no significant differences on any five of the maternal language variables for African American versus non-African American mothers within either of the two education levels. However, one significant mean difference by race emerged in both maternal education groups for children’s receptive vocabulary at pre-k, as measured by PPVT. For the high school degree or less group, African American children scored significantly lower than non-African American children on PPVT, t(9) = −3.56, p = .002. For the more than high school degree group, African American children again scored significantly lower than non-African American children on PPVT, t(9) = −5.84, p < .001. There were no significant mean differences on teacher-reported language as measured by ALI or on word knowledge as measured by WJ PV.

Table 2.

Comparisons of maternal language (maternal language quantity and maternal language complexity) and child language outcomes by race within maternal education groups

	High school degree or less				More than high school degree
	African American (n = 243)		Non-African American (n = 292)		African American (n = 283)		Non-African American (n = 474)
	M	SE	M	SE	M	SE	M	SE
Maternal language
Maternal language quantity, 6–36 mos
Number of conversational turns	62.46	2.72	55.44	2.33	66.70	2.39	60.01	1.94
Number of different words	84.60	2.25	82.86	1.92	95.40	1.98	96.76	1.60
Maternal language complexity, 6–36 mos
Mean length of utterance	3.36	0.05	3.31	0.04	3.45	0.04	3.48	0.04
Wh-questions	11.51	0.59	10.52	0.51	12.51	0.52	13.69	0.42
Complex conjunctions	6.47	0.57	6.77	0.49	8.31	0.50	8.68	0.40
Child language outcomes
Teacher-rated child language (ALI), pre-k	3.26	0.08	3.36	0.07	3.45	0.06	3.57	0.06
Receptive vocabulary (PPVT), pre-k	88.61	1.15	94.34**	1.00	91.11	0.99	99.51***	0.83
Word knowledge (WJ PV), kindergarten	97.20	0.74	99.01	0.64	99.02	0.64 1	100.94	0.54

Note:

^*p < .05.

^**p < .01.

^***p < .001.

Number of conversational turns was measured from 15 to 36 months. ALI = Adaptive Language Inventory. PPVT = Peabody Picture Vocabulary Test. WJ PV = Woodcock Johnson Picture Vocabulary.

Differences by Maternal Education Levels

Given our finding that there were no significant maternal language differences by race within the two education levels, Table 3 presents mean difference comparisons by maternal education only. Mothers with more than a high school degree were significantly higher on all maternal language input variables, including number of conversational turns, t(8) = 1.99, p =.05; number of different words, t(8) = 6.79, p < .001; mean length of utterance, t(8) = 3.17, p =.002; wh-questions, t(8) = 4.55, p < .001; and complex conjunctions, t(8) = 4.01, p < .001. Additionally, mothers with more than a high school degree had children who were higher on all child language measures, including teacher-reported language as measured by ALI, t(8) = 3.10, p = 0.002; as well as children’s receptive vocabulary as measured by the PPVT, t(8) = 4.11, p < .001; and word knowledge as measured by WJ PV, t(8) = 3.05, p = .002.

Table 3.

Comparisons of maternal language quantity, maternal language complexity, and child language outcomes across maternal education groups (high school degree or less, more than high school degree)

	High school degree or less (n = 535)		More than high school degree (n = 757)
	M	SE	M	SE
Maternal language
Maternal language quantity, 6–36 mos
Number of conversational turns	58.33	1.68	62.77*	1.33
Number of different words	83.70	1.39	96.21***	1.10
Maternal language complexity, 6–36 mos
Mean length of utterance	3.34	0.03	3.47**	0.02
Wh-questions	10.99	0.37	13.22***	0.29
	6.66	0.35	8.52***	0.28
Child language outcomes
Teacher-rated child language (ALI), pre-k	3.32	0.05	3.52**	0.04
Receptive vocabulary (PPVT), pre-k	92.06	0.71	95.92***	0.69
Word knowledge (WJ PV), kindergarten	98.26	0.46	100.12**	0.36

Note:

^*p < .05.

^**p < .01.

^***p < .001.

Number of conversational turns was measured from 15 to 36 months. ALI = Adaptive Language Inventory. PPVT = Peabody Picture Vocabulary Test. WJ PV = Woodcock Johnson Picture Vocabulary.

Maternal Language Quantity as Mediators

Bivariate associations for maternal language quantity.

The unadjusted bivariate associations of maternal language quantity were estimated in relation to the latent child language outcome. Number of conversational turns was not significantly correlated with child language, r = 0.03, p = .43, whereas number of different words was significantly correlated with child language, r = 0.38, p < .001.

Maternal language quantity as a mediator.

Fit indices for the model with maternal language quantity were good, Χ² (20, N = 1292) = 96.45, p < .001; RMSEA = 0.05 (C.I. = 0.04 to 0.07), CFI = 0.94, SRMR = 0.04. R² values and standardized path coefficients are provided in Figure 1. Maternal education and demographic controls explained 7% and 18% of the observed variation in number of conversational turns and number of different words, respectively. Moreover, maternal education, demographic controls, and maternal language quantity explained 42% of the observed variation in the latent construct of child language. Number of conversational turns did not mediate the relation between maternal education and child language while Number of different words did partially mediate the relation between maternal education and child language (β_{education → number of different words → child language} = 0.038, p < .001).

Figure 1.

Structural equation model (SEM) predicting child language in prekindergarten and kindergarten from direct and indirect effects of maternal language output and maternal education. Model fit statistics: χ²(20, N= 1292) = 96.45, p < .001; RMSEA = 0.05 (C.I. = 0.04 to 0.07), CFI = 0.94, SRMR = 0.04. Standardized path coefficients (interpretable as effect sizes for given relations) are provided on the single-headed arrow. Boldface type and solid lines indicate coefficients significant at the .05 level. Dashed lines indicate non-significant coefficients.

Note. ALI = Adaptive Language Inventory. PPVT = Peabody Picture Vocabulary Test. WJ = Woodcock Johnson. Covariates included gender (male), race (African American), mental development index, site (North Carolina), income-to-needs ratio, hours in child care, and father presence in the home. The significant indirect effect was maternal education → number of different words → child language, β= 0.038 (0.01), p< .001.

Mediation differences by race for maternal language quantity.

The above findings were tested to determine if the maternal language quantity model would be equivalent for the two race groups (African American, non-African American). The unconstrained model in which the mediation paths were freely estimated across groups had an adequate fit, Χ² (42, N = 1292) = 146.80, p < .001; RMSEA = 0.06 (C.I. = 0.05 to 0.07), CFI = 0.92, SRMR = 0.06. The models which constrained the mediation paths involving number of conversational turns or number of different words to be equal across groups were not significantly different from the unconstrained model, ΔΧ² (2) = 0.24, p = .89 and ΔΧ² (2) = 0.26, p = .88, respectively. These findings signified that the mediated relations for the maternal language quantity model were equivalent for African American and non-African American families.

Maternal Language Complexity as Mediators

Bivariate associations for maternal language complexity.

The unadjusted bivariate associations of maternal language complexity were estimated in relation to the latent child language outcome. Mean length of utterance, r = 0.26, p < .001, and wh-questions, r = 0.33, p < .001, were significantly correlated with child language, whereas complex conjunctions were not significantly correlated with child language, r = −0.04, p = .42.

Mediation effects for maternal language complexity.

Fit indices for the model with maternal language complexity were good, Χ² (22, N = 1292) = 102.23, p < .001; RMSEA = 0.05 (C.I. = 0.04 to 0.06), CFI = 0.95, SRMR = 0.03. R² values and standardized path coefficients are provided in Figure 2. Maternal education and demographic controls explained 11%, 12%, and 11% of the observed variation in mean length of utterance, wh-questions, and complex conjunctions, respectively. Moreover, maternal education, demographic controls, and maternal language complexity explained 44% of the observed variation in the latent construct of child language. Maternal mean length of utterance partially mediated the relation between maternal education and child language (β_{education → mean length of utterance → child language} = 0.019, p = .006) and maternal wh-questions also partially mediated the relation between maternal education and child language, (β_{education → wh-questions → child language} = 0.029, p < .001). Complex conjunctions did not mediate maternal education in predicting to child language.

Figure 2.

Structural equation model (SEM) predicting child language in prekindergarten and kindergarten from direct and indirect effects of maternal language complexity and maternal education. Model fit statistics: χ²(22, N= 1292) = 102.32, p < .001; RMSEA = 0.05 (C.I. = 0.04 to 0.06), CFI = 0.95, SRMR = 0.03. Standardized path coefficients (interpretable as effect sizes for given relations) are provided on the single-headed arrow. Boldface type and solid lines indicate coefficients significant at the .05 level. Dashed lines indicate non-significant coefficients.

Note. ALI = Adaptive Language Inventory. PPVT = Peabody Picture Vocabulary Test. WJ = Woodcock Johnson. Covariates included gender (male), race (African American), mental development index, site (North Carolina), income-to-needs ratio, hours in child care, and father presence in the home. Significant indirect effects were maternal education → mean length of utterance → child language, β= 0.019 (0.01), p= .006 and maternal education → wh-question words → child language, β= 0.029 (0.01), p< .001.

Mediation differences by race for maternal language complexity.

The above findings were tested to determine if the maternal language complexity model would hold equivalent for the two race groups (African American, non-African American). The unconstrained model in which the mediation paths were freely estimated across groups had an adequate fit, Χ² (46, N = 1292) = 157.06, p < .001; RMSEA = 0.06 (C.I. = 0.05 to 0.07), CFI = 0.93, SRMR = 0.06. The models which constrained the mediation paths involving mean length of utterance, wh-questions, or complex conjunctions to be equal across groups were not significantly different from the unconstrained model, ΔΧ² (2) = 0.59, p = .74; ΔΧ² (2) = 3.28, p = .19; and ΔΧ² (2) = 2.30, p = .32, respectively. These findings signified that the mediated relations for the maternal language complexity model were equivalent for African American and non-African American families.

Discussion

The goal of this study was twofold. First, this study added to the existing literature on mothers’ language to their young children by trying to disentangle the confound between race and SES, which has largely not been able to be analyzed in previous smaller studies of maternal language input and child language. This disentanglement also helps to understand the real scope of the “word gap” research (Hart & Risley, 1995; Hindman, Wasik, & Snell, 2016; Pungello, Iruka, Dotterer, Mills-Koonce & Reznick, 2009; Rowe, 2017; Tamis-LeMonda et al., 2012). We examined these maternal language input in a much larger and diverse sample of mothers and children than in previous language studies and examined five measures comprising maternal language quantity and maternal language complexity. We also obtained language samples at more time points than previous large scale studies, with maternal language variables from 6–36 months and measuring child language just before and at the end of kindergarten in over 1200 children. Second, we also wanted to replicate and extend other smaller studies with a larger variety of maternal language input variables to examine whether early maternal language input from 6–36 months might mediate the relation between maternal education and later child language during the transition to school. Finally, we examined whether the mediation models were similar or different for African American versus non-African American families.

Importantly in addressing our first question, this study appears to be the first study to help disentangle maternal race and education with respect to maternal language input. This study found there were no race differences on the five maternal language input variables within maternal education levels, helping to dispel previous reviews that had claimed African American mothers speak less to their children (Brooks-Gunn & Markman, 2005). We also found no support for African American mothers asking fewer questions as some other smaller studies have found (Anderson & Yockel, 1994) but also no support that African American mothers of lower SES ask more wh-questions than other low SES mothers (Cristofaro & Tamis-Lemonda, 2011). This might be the case for a variety of reasons including the large and representative nature of our study. This finding of no differences between African American and non-African American maternal language within the same education levels is all the more surprising and important because there were large income differences between African American and non-African American mothers in this sample even within the same education level. The income-to-needs ratios were about half as great for the African American families compared to non-African American families. The high school or less group had an income-to-needs ratio of .82 for African American families and 1.40 for non-African American families. The more than high school group had an income-to-needs ratio of 1.40 for the African American families and 2.75 for the non-African American families. Thus, despite much lower incomes, African American mothers were just as verbal with their children within the same education levels. This again suggests the critical importance of maternal education in understanding maternal language input. Our findings support the claim by Hindman et al. (2014) that maternal education might be the driver of early maternal language input and supports the “word gap” literature about the importance of maternal education as a differentiator of maternal language input (Hindman et al., 2016; Rowe, 2017).

Although there may be numerous reasons for the similarity of maternal language input by race within maternal education levels in this study, part of the reason may have been because the shared books in the current study were extensively piloted to make sure the books were equally relevant and acceptable to both African American and non-African American mothers. It also appears that when ethnographic studies have been conducted that examine language in more ecologically valid contexts (Heath, 1983; Sperry et al., 2018; Vernon-Feagans, 1996), few race differences have been found like the current study. There is also the possibility that the rural nature of this study may have contributed to the similarity of African American and non-African American maternal language within education levels (Vernon-Feagans et al., 2015; Vernon-Feagans & Swingler, in press).

We did find race differences within maternal education levels on one child standardized measure of vocabulary at pre-k, with African American children performing more poorly than non-African American children on the PPVT. Yet, we did not find race differences within maternal education levels on teacher-rated child language (ALI) or a standardized measure of word knowledge (WJ PV). Given no differences in maternal language input by race within education, our findings argue against previous claims that maternal language input by African American mothers may be a possible cause of their children’s lower vocabulary scores (Pungello et al., 2009). Instead, a myriad other factors might account for the lower vocabulary scores by African American children on the PPVT. It has been argued that vocabulary tests like the PPVT use vocabulary items that are culturally biased against African Americans (Hammer, Farkas, & Maczuga, 2010; Scheffner-Hammer, Pennock-Roman, Rzasa, & Tomblin, 2002). Another possible explanation for the lower scores on the PPVT may be that the PPVT is tapping into what has been called “academic language.” The term “academic language,” also called the “language of schooling,” refers to the language typically used in academic texts, scientific communication, and school learning that is generally decontextualized and in narrative form (Snow & Uccelli, 2009). Academic language may be area of language input and child language use that may be important to study in future maternal language and child language studies related to SES.

In answering our second question, even after controlling for a host of control variables (such as race, gender, child IQ, and family income), maternal language quantity (number of different words) and complexity (mean length of utterance and number of wh-questions) were significant partial mediators of the relationship between maternal education and later child language during the transition to school as measured by standardized language measures and teacher ratings. Using moderation analysis, we did not find mediation model differences between our African American and non-African American sample. Thus, these overall findings replicate previous studies of studies with younger samples of children that suggest that both maternal mean length of utterance and number of different words are important mediators of child language even into elementary school (Hoff, 2003; Huttenlocher et al., 2010). We also found that maternal education was still an important predictor of later child language, even in the presence of the mediating maternal language variables. Because our study controlled for a host of demographic controls that other studies have not been able to include (Hirsh-Pasek et al., 2015; Hoff, 2003; Huttenlocher et al., 2010; Tamis-LeMonda et al., 2012), these findings make a stronger case for the specific influence of both maternal education and maternal language input as predictive of later child language.

Although some of the maternal language measures used in this study have been used in previous studies, somewhat unique to this study was the inclusion of maternal wh-questions as one of the language complexity input measures. Wh-questions have recently been highlighted as important for better early child language (Cristofaro & Tamis-LeMonda, 2011; Rowe, Coker, & Pan, 2004; Rowe et al., 2016). For instance, Cristofaro and Tamis-LeMonda (2011) found that mothers’ wh-questions at 36 months of age predicted children’s receptive vocabulary scores at 36 months, which in turn predicted children’s later school readiness. Rowe et al. (2016) found that wh-questions (but not other types of questions) during father-child play in toddlerhood in a low-income small sample of African Americans was related to both child vocabulary and reasoning outcomes a year later. These authors suggested that posing wh-questions to toddlers may provide children with the opportunity to talk more and help build child vocabulary. Question asking may allow for more back and forth conversation between mother and child that further helps children elaborate on their language skills. Our study supports and expands these findings on wh-questions by demonstrating that early maternal wh-questions can predict child language not just concurrently but two to three years later during children’s transition to school. The other more recently-researched variable of number of conversational turns was not a significant mediator in our models (Romeo et al., 2018; Gilkerson et al., 2018). Thus, although this variable differentiated maternal language input across education levels, at least in this study it was not an important mediator of the relationship between maternal education and child language.

Although maternal language partially mediated the relationship between maternal education and child language, maternal education still accounted for significant variance in child language during the transition to school. Even in the presence of these maternal language mediators, there was a direct relationship between maternal education and child language outcomes. Partial mediation may only have been found because there are other factors that may be mediators of the relationship between maternal education and child language, like access to better resources such as books and other experiences children have with other adults (Pungello et al., 2009). For instance, in previous studies of fathers (Baker, Vernon-Feagans, & Family Life Project Key Investigators, 2015; Pancsofar & Vernon-Feagans, 2006, 2010), it has been found that father language input may be more important than maternal language input in predicting child language and achievement in early childhood and in kindergarten. Thus, fathers may be an important additional source of language input in future studies (Cabrera & Tamis-LeMonda, 2013). Rowe et al. (2016) found that maternal knowledge of infant development was an important mediator of the relationship between maternal education and child language. Thus, other possible mediators need to be explored in understanding the relationship between maternal education and child language.

Finally, there are limitations to this study. We gathered language samples during a picturebook task in the home with the mother. Thus, we did not sample the other important ecological settings in which the child was exposed to rich and varied talk by both other children and adults (Sperry et al., 2018a). Our effect sizes were also quite small for our mediation models, although given the large number of controls, the effect sizes were likely still meaningful. Even though this was a representative sample of families in rural low-wealth communities and identification of these families and their acceptance into the study was not different by race (see Vernon-Feagans et al., 2013, for more detail), there were very few African American mothers with a college education living in these counties. Thus, more studies need to be conducted that include larger numbers of African Americans and non-African Americans across many education levels. In addition, this study was conducted in the rural United States, with no non-rural comparison group. Thus, it is not clear whether these results would be replicated in a more urban sample.

In summary, this study contributes to our knowledge of maternal language input and child language related to race and SES, in what has been called the “word gap” research. Our findings support previous work that maternal education differentiates maternal language input, with more highly educated mothers providing more language output and language complexity to their young children, which in turn was related to children’s better language skills during the transition to school. We found no race differences within maternal education levels on five measures of maternal language input, even though our African American mothers within the same education levels were twice as poor as the non-African American mothers. We only found one child race difference within maternal education on receptive vocabulary. There was support for both maternal language quantity and maternal language complexity as partial mediators of the relationship between maternal education and child language but even with these mediators in our models, maternal education was still a powerful predictor of child language.