Anda di halaman 1dari 22

Developmental Psychology Copyright 2007 by the American Psychological Association

2007, Vol. 43, No. 5, 1062–1083 0012-1649/07/$12.00 DOI: 10.1037/0012-1649.43.5.1062

The Varieties of Speech to Young Children

Janellen Huttenlocher Marina Vasilyeva


University of Chicago Boston College

Heidi R. Waterfall Jack L. Vevea


University of Chicago and Cornell University University of California, Santa Cruz

Larry V. Hedges
University of Chicago

This article examines caregiver speech to young children. The authors obtained several measures of the
speech used to children during early language development (14 –30 months). For all measures, they found
substantial variation across individuals and subgroups. Speech patterns vary with caregiver education,
and the differences are maintained over time. While there are distinct levels of complexity for different
caregivers, there is a common pattern of increase across age within the range that characterizes each
educational group. Thus, caregiver speech exhibits both long-standing patterns of linguistic behavior and
adjustment for the interlocutor. This information about the variability of speech by individual caregivers
provides a framework for systematic study of the role of input in language acquisition.

Keywords: caregiver speech, caregiver education, speech to children.

Supplemental materials: http://dx.doi.org/10.1037/0012-1649.43.5.1062.supp

This article presents a longitudinal study of caregiver speech to available mechanisms (Baker, 2001; Chomsky, 1986; Lidz &
young children from 14 to 30 months of age. During this period, Gleitman, 2004). As Chomsky (1986) argued, “It is plausible to
children progress from single-word utterances to utterances in suppose that apart from pathology . . . such variation as there may
which words are combined to form sentences. Clearly, to acquire be is marginal and can be safely ignored across a broad range of
a language, children must be exposed to that language, both the linguistic investigation” (p. 18). Further, caregiver speech has been
words and the ways they can be combined. To determine what role said to involve errors, false starts, and so forth and may not be
input plays in acquisition, it is necessary to empirically examine sufficient for inducing grammatical rules (e.g., Lidz, Gleitman, &
caregiver speech and its relation to the development of language in Gleitman, 2003; Lidz, Waxman, & Freedman, 2003; Newmeyer,
the child. Yet claims about the role of language input have been 2003).
based more on theoretical assumptions than on systematic obser- Some recent investigators have adopted a different theoretical
vation. It has been assumed that the role of input is a limited perspective, one in which input is seen as the source of child
one—that there is little variation in the speech of different care- language (e.g., Goldberg, 1995; Jackendoff, 2002; Tomasello,
givers or in the syntactic development of different children. Ac- 2003). For example, Tomasello (2003) has argued that language
quisition has been seen as emerging from innate and universally input, when coupled with domain-general learning skills, is suffi-
cient to derive a full adult grammar. There is empirical work
indicating that there are substantial variations in language input
and that these are related to differences in children’s language
Janellen Huttenlocher, Department of Psychology, University of Chi-
cago; Marina Vasilyeva, Lynch School of Education, Boston College; development. While such findings are correlational, some of the
Heidi R. Waterfall, Department of Psychology, University of Chicago, and studies strongly suggest that, at least in part, observed relations
Department of Psychology, Cornell University; Jack L. Vevea, Department reflect a causal role of the input in acquisition (e.g., Hoff-
of Psychology, University of California, Santa Cruz; Larry V. Hedges, Ginsberg, 1998; Huttenlocher, Vasilyeva, Cymerman, & Levine,
Department of Sociology, University of Chicago. 2002). Also, recent findings have revealed powerful learning
Larry V. Hedges is now at the Department of Statistics and the Depart- mechanisms that allow infants to extract statistical regularities
ment of Education & Social Policy, Northwestern University. from language input (e.g., Saffran, 2001). Together these findings
This research presented was supported by National Institutes of Health
suggest that caregiver speech may be a driving force in syntactic
Grant PO1 HD40605. We thank Susanne Gahl, Susan Levine, Stella
development (e.g., Andersen, 1973; Bybee, 1998; Elman, 1993).
Lourenco, Nora Newcombe, and Mary C. Potter for their helpful comments
on the manuscript. The long-term goal of our program of research is to establish the
Correspondence concerning this article should be addressed to Janellen role of input in children’s language development. However, the
Huttenlocher, Department of Psychology, University of Chicago, 5848 investigation of caregiver speech itself involves a major research
South University Avenue, Chicago, IL 60637. E-mail: hutt@uchicago.edu effort, and it is the focus of the present article. We propose to
1062
SPEECH TO YOUNG CHILDREN 1063

determine whether caregiver speech changes systematically as the of mother speech (number of clauses per utterances and mean
child develops and if there are substantial variations in speech length of utterance) increased slightly and repetitions decreased,
among caregivers that persist over time. Such findings would the authors concluded that “the mothers’ usage does not change
indicate that it is important to examine the relation of input to dramatically during the child’s learning period from one to three
language development. Before presenting our study, we briefly years” (p. 65). Rowland, Pine, Lieven, and Theakston (2003)
review findings from prior research on differences in caregiver found no change in maternal use of wh questions over a 1-year
speech over time and individual differences across speakers. period from 2 to 3 years of age, although Theakston, Lieven, Pine,
and Rowland (2005) found fewer questions as children approached
Does Caregiver Speech Change Over Time? 3 years than at 2 years.
In summary, existing longitudinal studies of caregiver speech do
It has been claimed that caregiver speech becomes more com- not provide a clear picture of change over time. The findings of
plex as children develop (e.g., Snow, 1972). If it does, that would different studies do not always agree, as might be expected given
suggest that parents are adjusting their speech to changing char- the small samples used. The largest samples included 12 partici-
acteristics of the child. General questions concerning the adjust- pants at each age (i.e., Gleitman et al., 1984; Snow, 1972). Furrow
ment of caregiver speech to children’s language can be examined et al. (1979) examined only seven families, and Kavanaugh and
by investigating just the speech of caregivers, either in cross- Jirkovsky (1982) included only four. Only the earliest stages of
sectional or in longitudinal studies. This type of investigation can syntactic development have been examined, and, further, studies
provide information as to what aspects of caregiver speech do or have used different measures. Clearly, existing studies do not
do not change systematically with child age. However, questions permit assessment of the relation of caregiver language to child
as to whether caregivers fine tune their speech to the child’s age. In the present study, we systematically examine caregiver
language level will require examination of parent– child interaction speech using a broad range of measures over an extended period in
(e.g., Berko-Gleason, 1977; Snow, Perlmann, & Nathan, 1987). a diverse and sizeable group of families.
Findings from cross-sectional studies suggest that caregiver
speech varies as a function of child age. For example, Snow (1972)
compared speech to 2-year-olds and 10-year-olds. She found that Are There Individual Differences in Caregiver Speech?
parental speech to the younger group involved more redundancy as
well as fewer clauses per utterance. Other studies have examined There is suggestive evidence that there are individual differ-
caregiver speech over a narrower range of child age than that ences in the speech of caregivers related to demographic factors
studied by Snow (1972). Phillips (1973) found differences in (e.g., socioeconomic status [SES]) and that these are related to the
maternal mean length of utterance in speech addressed to 18- speech of their children (e.g., Elardo, Bradley, & Caldwell, 1977;
month-olds versus 28-month-olds. Rondal (1980) examined input Hart & Risley, 1992; Price & Hatano, 1991). Lower SES mothers
to children at 18 months and 36 months of age and found differ- talk less and spend less time in mutual activities with their children
ences in “lexical diversity” (as measured by the type/token ratio), than do middle-SES mothers, and their speech is less contingent on
syntactic complexity, as well as utterance length. Further, de- the child’s speech (e.g., Bee, Van Egeren, Streissguth, Nyman, &
creases were found in imperatives, repetitions, and expansions. Leckie, 1969; Farian & Haskins, 1980; Heath, 1982; Hess &
However, it should be noted that the samples in these studies were Shipman, 1965). Further, middle-SES mothers include more
small, which is problematic for cross-sectional comparisons of language-teaching speech during play with children than do lower
speech to children of different ages since child age may be con- SES mothers (Hammer & Weiss, 1999). Hoff (2003a, 2003b)
founded with other characteristics. found that several measures of mother speech (utterance length,
Although it has been claimed that speech to very young children number of word types, and number of word tokens) were corre-
involves simplified syntax (e.g., Snow, 1972), the evidence from lated with SES and also were predictive of child vocabulary.
longitudinal studies is mixed as to whether complexity of caregiv- Similarly, Pan, Rowe, Spier, and Tamis-Lemonda (2004) found
ers’ speech actually increases as children become older. Kaye that maternal educational level was associated with children’s
(1980) found that speech to young infants (less than 26 weeks) vocabularies (as measured by the Peabody Picture Vocabulary
included shorter utterances and was more repetitive than speech to Test, 3rd ed.; Dunn & Dunn, 1997).
2-year-olds. However, Snow (1977) found no difference in utter- Caregiver speech is not generally examined longitudinally in
ance length for mothers when speaking to 3-month-olds versus these studies. Hence, direct evidence is lacking with respect to
18-month-olds. Also, Kavanaugh and Jirkovsky (1982) found no whether observed individual differences are long standing. How-
differences in utterance length in parents’ speech when children ever, there is suggestive evidence that SES differences may indeed
were 9 months, 12 months, and 15 months, although exact repe- be relatively permanent since speech to other adults, like that to
titions decreased. young children, varies with SES. Two studies have found differ-
Furrow, Nelson, and Benedict (1979) examined the speech of ences in adult-to-adult speech in different SES groups. Hoff
mothers over a slightly older age range—at 18 months and 27 (2003b) found differences among mothers from different SES
months old—and found no significant differences in syntactic groups when talking both to adults and to children. Van den
characteristics of speech such as number of clauses per utterance Broeck (1977) found that syntactic complexity was related to
or use of different types of questions, auxiliaries, and so forth. educational level for specific contexts in adult-to-adult discourse.
However, exact repetitions decreased with age. Gleitman, New- Although these studies are suggestive, systematic research on the
port, and Gleitman (1984) examined speech to children from 18 to nature of the syntax of caregiver speech in different SES groups
21 months and again from 24 to 27 months. While the complexity remains to be carried out.
1064 HUTTENLOCHER ET AL.

Studies that have examined particular aspects of syntax show points, every 4 months during this period. The set of measures we
substantial individual differences in caregiver speech. These stud- developed capture a variety of important characteristics of speech
ies have shown an association with corresponding aspects of to young children. Three kinds of measures were used: indicators
children’s language. For example the proportion of auxiliary- of the composition of speech (measures of clausal and constituent-
fronted questions varies across parents and is associated with more level complexity), indicators of the diversity of speech (number of
rapid growth of auxiliaries in children (e.g., Furrow et al., 1979; different words and the number of different kinds of sentences),
Newport, Gleitman, & Gleitman, 1977). Naigles and Hoff- and indicators of quantity (numbers of words, utterances, and
Ginsberg (1998) found that the relative verb frequency and the sentences). Using these measures of quantity, diversity, and com-
diversity of syntactic environments in child-directed speech were position, we examined the nature of individual differences in
strong indicators of early verb use. Hoff-Ginsberg (1986) found caregiver speech and the pattern of change over time.
that the average number of noun phrases per utterance in mothers’
speech was a positive predictor of the same measure in children’s Participants
speech. Lastly, the proportion of multiclause sentences by care-
givers is related to children’s comprehension and production of The present article includes data on 50 families from the greater
multiclause sentences (Huttenlocher et al., 2002). Again, the data Chicago area who are a subset of 64 who are participating in a
from caregivers have not been longitudinal so that direct evidence larger longitudinal study. Recruitment for the larger study was
is lacking as to whether observed differences are long lasting. based on direct mailing to roughly 5,000 families living in targeted
There is evidence of situationally based variations in caregiver zip codes and an advertisement in a free, monthly parent magazine.
speech. For example, mothers vary speech to a target child de- Parents who responded to the mailing or advertisement were asked
pending on who else is present. Snow (1982) found that toddlers to participate in a screening questionnaire over the phone. Infor-
received fewer than half as many utterances from caregivers when mation gathered included child gender; parents’ income, educa-
an older sibling was present. Jones and Adamson (1987) also tion, and occupation; as well as race and ethnic identification. To
found that quantity of speech (number of utterances) was affected ensure a diverse sample, we chose families to match as closely as
by the presence of an older sibling. Only measures of quantity possible the 2000 census data on family income and ethnicity for
were affected (number of tokens and number of utterances); mea- the greater Chicago area.
sures such as mean length of utterance were not (Oshima-Takane The criteria for drawing the sample used in the present study
& Robbins, 2003). Further, there is evidence that mothers speak were the following. First, the study was limited to families in
differently when addressing their firstborn versus their later-born which the primary language was English. Second, the sample was
children. They use longer utterances and address more metalin- limited to families where one parent was the primary caregiver
guistic utterances to their children than do mothers of later borns over the entire period being studied. Third, we included only
(Hoff-Ginsberg, 1998; Jones & Adamson, 1987). families that participated in at least four out of the five observation
In summary, fundamental questions about whether there are sessions. The resulting sample consisted of 48 mothers plus 2
long-lasting individual differences in the speech of different care- fathers (the children in these families included 26 boys and 24
givers have not been answered by research on caregiver speech to girls). The families were subdivided into four educational levels
date. Given the sparsity of longitudinal studies of caregiver speech, and six income levels. The numbers of families in different edu-
it is not yet clear to what extent there are substantial and long-term cational, income, and racial groups are shown in Table 1. The
differences among different caregivers. Further, existing studies numbers of families in which the target child was the firstborn,
have not examined a wide range of characteristics of caregiver second born, or had more than one older sibling also is shown in
speech. Table 1.

The Present Study Method


The review of existing literature above indicates a need for Families were visited once every 4 months at home. Included
further longitudinal data on caregiver speech. Information on long- here are data from caregivers during five visits at child age of 14,
standing characteristics of language to children is critical to deter- 18, 22, 26, and 30 months. During each visit, the caregiver and
mining how input may be related to development. The present child were videotaped for a 90-min period during which they
study explores the nature of caregiver speech during the period of engaged in their ordinary daily activities. After the session was
early syntactic growth. We intend to examine the nature of indi- completed, the tapes were transcribed at our lab. Transcription
vidual differences among caregivers in child-directed speech, de- involved breaking the stream of speech into distinct utterances.
termine whether these are stable over time, and whether they are These utterances were then analyzed grammatically according to
systematically related to other variables such as child gender and the coding system presented below to characterize the composition
family income. We have developed a broad set of measures to of the utterances. We did not use the formats for the CHILDES or
examine the characteristics of caregiver speech over time and to SALT databases because they were not well-suited to answering
explore variability across caregivers. We examine the factors that our research questions.
may be associated with characteristics of their speech, including The analysis of caregiver language is based on speech to target
both relatively stable family characteristics such as income and children. One reason for restricting our study to child-directed
education and contextual factors such as presence of older siblings. speech is that this speech sample can be reliably determined.
The families in our study vary widely in socioeconomic back- Although, in principle, children can learn from other-directed
ground (education and income). Families were visited at five time speech, it is often unclear whether children are attending to speech
SPEECH TO YOUNG CHILDREN 1065

Table 1 Quantity Measures


Distribution of Participants Across Social Factors
We used three measures to assess the quantity of caregiver
Social factor Frequency speech. The first measure was the total number of words (i.e.,
tokens) produced by the caregiver. The remaining two were the
Educational level
High school only 6 total number of utterances and the total number of sentences.
Some college 9 Tokens. To calculate the number of tokens, we counted the
Bachelor’s degree 18 number of instances of words in the entire 90-min transcript. For
Advanced degree 17 example, if a given mother said the word shoe 50 times in a
Income level
⬍$15,000 4 transcript but said nothing else, she would have 50 tokens; like-
$15,000–$34,999 10 wise, a mother who said 50 different words also would have 50
$35,000–$49,999 6 tokens. We excluded specific classes of words from our token
$50,000–$74,999 8 analysis, and by extension, the same classes were excluded from
$75,000–$99,999 11
all other measures as well. These classes were animal noises (e.g.,
ⱖ$100,000 11
Race/ethnicity baaa, bow wow, etc.), letters of the alphabet (except a and i), as
African American 10 well as interjections such as ooooh, ouch, and uh-oh. Lastly,
Asian 2 parental imitations of infant babbling were also exempted (e.g.,
Hispanic 4 ka ka ka ⫽ zero tokens). Utterances that contained only one of
White 34
Birth order of target child the above forms were deleted before analysis began on the
Firstborn 31 transcript.
Second born 11 Utterances. For each participant, we calculated the total num-
Third or later born 8 ber of utterances in the entire transcript. To arrive at this measure,
we divided the flow of speech into utterances based on intonation
and pauses, as well as conversational turn taking. An utterance
that is not directed to them. Further, children’s attention to other- consisted of a single intonational contour within a conversational
directed speech may vary with their language levels. Indeed, very turn. Intonational contour frequently includes falling or rising pitch
young children seem to require special prosodic features (mother- (as in declaratives and questions, respectively), and often there is
ese) to attend to speech. Finally, findings showing substantial a pause preceding and following it. An utterance may include a
relations to growth of syntactic skills have been based on speech single word (e.g., Stop!), an isolated phrase (e.g., big boy), or a
directed to target children (e.g., Barnes, Gutfreund, Satterly, & single or multiclause sentence. Two independent clauses not con-
Wells, 1983). nected either by intonation or lexical items (e.g., and, because)
Reliability of transcription was established by having a second were considered two utterances, even if they occurred within the
individual independently transcribe 20% of the videotapes. The same conversational turn. A sentence that contained short word-
reliability criterion was set at 95%: That is, the two transcribers searching pauses was considered a single utterance (e.g., Bring me
had to be in agreement on 95% of the utterances. In rare cases that [pause] shoe).
where the reliability coder disagreed with the transcriber for more Sentences. To calculate the number of sentences, we first
than 5% of utterances, disagreements were resolved with a third characterized utterances as to whether they contained zero, one, or
judge. If the reliability coder and the transcriber agreed on at least more than one clause. Zero-clause utterances were those that did
95% of utterances, the original transcriber’s data were used. Once not contain a verb; these utterances were not counted as sentences.
reliability was established, each transcript was subjected to lin- Typically zero-clause utterances contained just a noun (bear), a
guistic analysis, as described below. Reliability was also investi- noun phrase or proper noun (your bear, Jenny), a prepositional
gated on 50% of the syntax-coded transcripts, and reliability be- phrase (in your room), a preposition (up), or an interjection (yeah,
tween the syntax reliability coder and the primary syntax coder no, alright, thank you, etc.). An utterance was coded as having one
was again set at 95%. The same procedure was used as above. clause if it contained a single verb phrase. In cases where the
copula be was omitted, the utterance was also coded as having one
clause (e.g., You tired? You big boy now!). We refer to one-clause
Treatment of Caregiver Speech utterances as simple sentences.
Our measures were designed to capture major characteristics of
caregiver speech, spanning diversity as well as quantity of speech.
The measures address both the lexicon and syntax. Diversity Diversity Measures
measures capture the different kinds of words and the different
kinds of syntactic structures caregivers produced. At least at the We used two different measures of linguistic diversity: word
extremes, diversity depends on quantity; the number of different types, which captures the number of different lexical items used by
words (types) cannot exceed the total number of words (tokens), a caregiver, and sentence types, which indicates the number of
and the number of different kinds of sentences cannot exceed the different types of complex sentences that a given caregiver used.
total number of sentences. The compositional measures capture the The number of word and sentence types is neither a straight
syntactic complexity of caregiver speech (e.g., the proportion of compositional nor a straight quantity measure. At least for the
complex sentences a caregiver produces). Composition is distinct extremes of frequency, the number of different types will depend
from quantity of speech. on quantity of speech.
1066 HUTTENLOCHER ET AL.

Word types. Word types were calculated by counting the num- complement serving as object of the main clause and an object-
ber of unique words in the transcript. A single type includes all of relative clause in the subordinate clause.
the inflectional variations of a given word (e.g., jump, jumps,
jumping, jumped ⫽ one type). Words with irregular inflectional
morphology were considered to constitute one type (e.g., goose/ Compositional Measures
geese, run/ran). Derivationally related words, however, were
The measures of syntactic composition are of central impor-
treated as distinct words (slowly, slow ⫽ two types). A proper
tance. They include the proportion of multiclause sentences and
name and its nickname were also treated as one type (Jenny,
two measures that are distinct from clausal structure: the average
Jennifer ⫽ one type). Likewise, proper names, song titles, and
book titles that contained more than one word were also treated as number of noun phrases per sentence and the average number of
a single type (e.g., Sponge Bob Square Pants ⫽ one type, The words per sentence. These two measures capture constituent-level
Pokey Little Puppy ⫽ one type). Commonly occurring “baby” or complexity.
motherese words were standardized and treated as examples of the Multiclause sentences. To calculate the proportion of multi-
same type. For example, nummy, yummy, and yummers were all clause sentences, we divided the total number of multiclause
treated as examples of yummy. sentences by the total number of sentences. We described how
Sentence types. Structural coding of complex sentences we determined whether a sentence contained multiple clauses
yielded seven types of two-clause sentences, as indicated below. above.
We determined, for a particular caregiver, whether all or just a Noun phrases per sentence. We coded the number of noun
subset of the types were used. It also was possible to characterize phrases (noun, pronoun or proper noun, with optional adjectives
caregiver speech further in terms of the number and kinds of and articles) that either filled an argument position in the syntax
three-, four- and five-clause sentences that they used. Our diversity (e.g., I want the ball) or were used in adjuncts (e.g., Mommy needs
of syntax measure involved the number of distinct kinds of com- to lay down on the couch). Tags, however, were excluded (e.g.,
plex sentences that a given caregiver used in a particular session. aren’t you?). Use of the number of noun phrases addresses a level
For all multiclause sentences, we categorized each sentence of syntactic complexity that is not captured by the number of
based on the structural–syntactic relations between the clauses. We clauses. This measure of noun phrases per sentence provides an
did not consider serial verb constructions (e.g., go get it), modals index of constituent-level complexity; it includes prepositional
(e.g., going to do it), or tags (e.g., isn’t it?) as sufficient to phrases, locative expressions, and so forth. The more constituents
constitute a multiclause utterance. Because subordinate clauses can that are present in a sentence, the more complex the overall
vary morphologically (e.g., bearing infinitive or gerundive mark- structure is. Noun phrases include those that are obligatory in the
ing), we categorized two-clause sentences according to their struc- syntax (e.g., subjects and direct objects) as well as those that are
tural relations. These included the following: coordination, adjunc- optional (e.g., the “by-phrase” in the passive and the prepositional
tion (preceding the main clause), adjunction (following the main phrases). A sentence that includes optional noun phrases can be
clause), subordinate clauses with object as complement, sentences considered more complex than one that does not. Note also that
with subordinate clauses as subjects, object-relative clauses, and imperatives (e.g., Come here!) would be considered less complex
subject-relative clauses. according to this metric. However, since we were measuring the
The first relation is coordination, where the two clauses are number of noun phrases, this seems correct.
conjoined by and or or (e.g., Jimmy went to the store and bought It should be noted that locative expressions like here and there
milk). We also coded for two types of adjunct clauses: one where were coded as noun phrases only when they served as objects of
the adjunct precedes the main clause (e.g., Before you go outside, prepositional phrases (e.g., in here). Likewise, this, that, and what
put on your coat) and one where the adjunct clause follows the were counted as noun phrases only when they replaced the subject,
main clause (e.g., Put on your coat before you go outside). Next, the object, or the object of a preposition. Who was counted as a
we coded for subordinate clauses fulfilling the role of object for noun phrase, except when it served as a complementizer in a
the main clause (i.e., the main verb subcategorizes for a clausal relative clause (e.g., the girl who lives next door). Lastly, posses-
complement; e.g., I thought that you were tired) and for subordi- sive noun phrases (e.g., Mommy’s shoes, the little girl’s toys) were
nate clauses fulfilling the role of subject of the main clause (e.g., coded as two separate noun phrases because the leftmost noun
What you need is a nap). Next, we divided the relative clauses into phrase (Mommy, the little girl) has the internal structure of an
those that modify the main clause subject (e.g., The doll that independent noun phrase (cf. Anderson, 1992).
Grandma gave you is all dirty; The boy that likes ice cream is Words per sentence. We also calculated the average number of
here) and those that modify the main clause object (e.g., Hand me words per sentence, including both complex and simple sentences.
the piece that goes over here; I know the one you want). In order By calculating number of words rather than morphemes, we found
to keep this measure parallel to the other types of subordinate that our results would be more comparable cross-linguistically for
clauses discussed above and to avoid proliferation of subordinate languages in which many words in a sentence are inflected (cf.
clause types, we did not further subdivide relative clauses based on Nelli, 1998; Slobin & Bever, 1982). Like number of noun phrases,
the role of the head noun phrase within the relative clause. Sen- number of words addresses a level of syntactic complexity that is
tences that contained more than two clauses were then coded for not captured by the number of clauses. Because tags were excluded
each relation holding among the clauses. In other words, each of from this measure, the words that are counted are structurally part
the seven basic types could be combined with any other, thus of a sentence, including modifications of nouns and verbs through
forming a new type of complex sentence. For example, I know that the use of adjectives, adverbs, and so forth. Thus, this measure,
you want the one that Grandma gave you contains both a clausal like number of noun phrases, provides an index of constituent-
SPEECH TO YOUNG CHILDREN 1067

level complexity, indicating the number of lexical items in the ences in growth parameters. A fourth question is whether the
syntactic tree. differences in caregivers remain stable, that is, whether individuals
tend to retain the same rank ordering relative to one another over
Group-Level Predictors of Variation in Caregiver Speech time. The stability of rank ordering of caregivers’ speech is as-
sessed using Kendall’s coefficient of concordance.
We examined four group-level variables associated with char-
acteristics of parent speech. Three of these are constant over time. Change Over Time: Overall Trends
Two are measures of SES, namely family income and education
(of the primary caregiver). The family income measure divided Consider first how caregivers as a group change over time with
participants into six groups; the frequencies are listed in Table 1. respect to each of the measures of caregiver speech. Table 2
We preserved the level of detail reported in the table despite some presents tests of the form of the overall change trajectory for each
sparse frequencies to capture the pronounced differences between of the eight measures. For the three measures that characterize
income groups. The education measure divided participants into quantitative aspects of speech—number of word tokens, number of
four groups, as reported in Table 1. Again, we preserved the level utterances, and number of sentences—there is no significant over-
of reported education despite sparse frequencies to capture differ- all change across time. In contrast, all five complexity measures
ences between individuals with a high school education and those show change over time. Two of these complexity measures char-
with some college. The third variable is child gender. It is known acterize the diversity of speech—number of word types and num-
that girls tend to have higher language levels in the early stages of ber of sentence types—and the remaining three complexity mea-
development; hence, it is of interest to determine whether there are sures are compositional measures—number of complex sentences
associated properties of caregivers’ speech. The fourth variable, relative to the total number of sentences, number of noun phrases
the presence of older siblings during a visit, is not constant over per sentence, and number of words per sentence. As shown in
time. As we have noted, it has been found that quantity of care- Table 2, there is significant linear change for all five measures
giver speech to a target child varies with older sibling presence, across all five time points. Words per sentence is the only measure
whereas the composition of their speech does not. We examined where the increase has a significant quadratic component: The rate
whether this pattern was found in our data, and, if so, determined of increase grows over time. However, this quadratic component is
whether the presence of older siblings affected diversity measures, small, having an almost negligible impact on the trajectory of
which are sensitive to both quantity and variety of speech. change. In general, then, the complexity of caregiver speech in-
creases linearly over the entire age range studied.

Results
Change Over Time: Individual Differences
The treatment of results is concerned with four fundamental
Let us now consider whether there are systematic individual
questions. To answer these questions, we consider patterns of
differences in the speech of different caregivers. We address this
speech for each caregiver across time in the form of trajectories of
question by examining variation across individuals in intercept and
change (which can include no change as a special case). These
slope. For each measure, Table 3 presents the standard deviations
trajectories or change curves can be characterized by an intercept
(square roots of the estimated variance components) together with
that describes parent speech at the initial observation point and one
the chi-square test statistics for each measure. There is substantial
or more other parameters that describe the nature of change over
variability across individuals for both intercept and slope for all
time. We use two additional growth parameters: one describes the
measures except one, sentence types, where there is no significant
rate of linear growth over time and the other describes quadratic
individual variation in slope.
change over time, that is, acceleration (or deceleration) of growth
While there is no overall change across time for the three
over time. We employ hierarchical linear modeling (HLM) proce-
quantity measures, there are nevertheless large variations in linear
dures (Raudenbush, Bryk, Cheong, & Congdon, 2000) to carry out
the statistical analyses. We present summary information about the
HLM analyses here; for more complete information about coding,
model selection, and parameter estimates, see the Appendix. Table 2
One of our fundamental questions is whether caregiver speech Tests for Shapes of Change Trajectories
changes over time. To address this question, we used HLM to
t(49) statistic for t(49) statistic for
examine whether the mean trajectory of caregiver speech across all
Measure linear change quadratic change
caregivers on each measure changes over time. The second ques-
tion is whether there are individual differences among caregivers. Quantitative
To address this question, we used HLM to examine whether there Word tokens 1.84 1.02
is variation in individual caregivers’ growth parameters. Thus, in Utterances ⫺1.06 0.24
Sentences ⫺0.27 0.65
addition to tracking the overall characteristics of caregivers over Complexity
time, the model allows us to examine the nature of individual Word types 4.90*** 1.43
differences in the pattern of language use. If there are substantial Sentence types 7.97*** 1.46
individual differences, a third question arises concerning the char- Complex sentences 10.22*** 1.47
Noun phrases per sentence 8.41*** 0.86
acteristics of caregivers that may account for those differences Words per sentence 7.48*** 2.41*
(e.g., level of education). Here, too, we used HLM to model the
association between explanatory variables and individual differ- *
p ⬍ .050. ***
p ⬍ .001.
1068 HUTTENLOCHER ET AL.

Table 3
Root Variance Components (VC) for Growth Components

Intercept Linear slope

Measure 公VC ␹ (49)


2 公VC ␹2(49)

Quantitative
Word tokensa 1,638.43 965.34*** 224.19 75.44**
Utterancesa 355.70 767.88*** 49.36 72.08*
Sentencesa 248.22 788.86*** 39.88 74.70***
Complexity
Word types 106.52 430.58*** 15.75 97.54***
Sentence typesa 2.29 161.32*** 0.48 58.53
Complex sentencesb 0.31 1,147.17*** 0.10 276.72***
Noun phrases per sentence 0.18 265.42*** 0.04 98.17***
Words per sentence 0.52 290.84*** 0.24 82.61**
a b
Inference reported in square-root metric. Inference and estimate reported in logit metric.
*
p ⬍ .050. ** p ⬍ .010. *** p ⬍ .001.

slopes, as shown in Table 3. That is, although the overall (average) analyses not reported in detail here show that family income is not
slopes are negligible, the standard deviations reflecting individual a significant predictor of any characteristic of growth for any of the
differences are highly significant; the average slope is a mixture of measures when caregiver education is controlled. It should be
individual caregiver slopes that are positive, flat, and negative. For noted that educational level does not account for all of the indi-
all complexity measures, there are large variations among individ- vidual differences in intercept; that is, a significant variance com-
uals for the intercepts as shown in Table 3. For all complexity ponent remains even after education is accounted for.
measures except sentence types, there are also large variations For a subset of measures, we initially found that child gender
among individuals for the linear slopes. was a significant predictor of the intercept. However, gender
effects vanished when presence of older siblings at the particular
Change Over Time: Subgroup Differences data collection session was controlled. Table 5 shows that older
siblings were more frequently present for girls than for boys in our
Now let us consider whether some of the individual differences visits to children. Table 6 presents results showing the effect of
that characterize change over time are associated with caregiver older siblings being present. The presence of older siblings is
education, family income, gender of child, or the presence of older strongly predictive of differences in the three measures of speech
siblings at a session. Given the large individual variation, it is quantity (tokens, utterances, and sentences), as well as the two
reasonable to ask whether this variation is associated with explan- measures of diversity. However, for compositional measures, no
atory variables. such association with presence of older siblings was found.
In general, caregiver education is associated with individual Both education and presence of siblings significantly predict
differences in intercepts but not slopes for all but one of our characteristics of change for several measures. The figures show
measures (see Table 4). The only exception to this pattern is the modeled growth curves based on the statistically significant
proportion of complex sentences, where education predicts the growth parameters for each measure separately. Figures 1, 2, and
slope rather than the intercept, ␹2(3) ⫽ 9.97, p ⬍ .05. In contrast, 3 show curves for each level of education and sibling condition on
the quantity measures. Tables 7, 8, and 9 show means, standard
deviations, and sample sizes for these quantity measures, by time
Table 4 and education level. The modeled curves depicted in the figures
Tests for Differences in Intercept by Educational Level indicate the structure of change in the means over time. For
␹2(3) statistic for example, the mean number of tokens at every age is higher for
Measure intercept differences parents with graduate degrees than for any other group; this is
reflected in the fact that the model intercept is highest for that
Quantitative group and consequently the growth curve for that group (see
Word tokensa 16.94*
Utterancesa 19.47***
Sentencesa 15.45*
Complexity Table 5
Word types 12.50** Total Number of Older Siblings Present at Each Visit
Sentence typesa 9.47*
Complex sentencesb 4.22
Visit number
Noun phrases per sentence 13.11**
Words per sentence 9.51*
Gender 1 2 3 4 5
a b
Inference reported in square-root metric. Inference and estimate re- Girls (n ⫽ 23) 7 4 10 14 10
ported in logit metric.
Boys (n ⫽ 27) 2 4 5 2 5
*
p ⬍ .050. ** p ⬍ .010. *** p ⬍ .001.
SPEECH TO YOUNG CHILDREN 1069

Table 6 For diversity (word types and sentence types), Figures 4 and 5
Tests for Effect of Presence of Older Siblings show separate curves for each level of education and sibling
condition. Tables 10 and 11 show the corresponding means, stan-
t(49) statistic for dard deviations, and sample sizes, by time and education. Again
Measure effect of siblings
the differences in intercept for the different educational groups and
Quantitative sibling conditions are substantial, but there were no differences in
Word tokensa ⫺3.89*** slope. However, for both of these measures, there was substantial
Utterancesa ⫺4.67*** change over time; the trajectories of increase over time for the
Sentencesa ⫺4.19***
Complexity diversity measures were parallel for different levels of education
Word types ⫺3.91*** and sibling conditions. These phenomena may be observed in the
Sentence typesa ⫺2.44* tabled means. Note that the means of word types for parents with
Complex sentencesb ⫺1.33 graduate degrees are consistently higher than for other groups, just
Noun phrases per sentence ⫺1.26
Words per sentence 0.46
as the means for parents with a high school education are consis-
tently lower. The significant linear growth reflects the fact that
a b
Inference reported in square-root metric. Inference and estimate re- means tend to increase with time, which may be observed by
ported in logit metric. comparing means from left to right in any row of the tables. (For
*
p ⬍ .050. *** p ⬍ .001. diversity, the differences associated with educational level are less
consistent; this is reflected both in the means and in the modeled
curves, where those with some college appear quite similar to
Figure 1) is higher than for any other group. Differences in those with graduate degrees.) For compositional measures, noun
intercept are substantial, but there are no differences in slope and, phrases per sentence and words per sentence, again, educational
in addition, amount of speech does not change over time for any group predicts only the intercept. Figures 6 and 7 show the pattern
subgroup. This is consistent with the observation that changes in of parallel increasing trajectories for these measures, and Tables 12
the means across time for any education group appear to be and 13 list the means, standard deviations, and sample sizes. The
random fluctuations rather than systematic growth. increase over time was parallel for different educational groups.

Figure 1. Modeled growth in tokens for different educational levels with and without older siblings present.
1070 HUTTENLOCHER ET AL.

Figure 2. Modeled growth in utterances for different educational levels with and without older siblings present.

The proportion of complex sentences is the one measure where two-clause utterances, there is significant variation within the
caregiver education is associated with differences in slopes rather one-clause group and within the two-clause group depending on
than intercepts. The change curves for the four education levels educational level. Table 16 reports the same breakdown for num-
given in Figure 8 show that the proportion of complex sentences is ber of words in one- and two-clause sentences. Supplementary
similar for all levels of education at 14 months. However, the rate analyses show that caregiver education is still a significant predic-
of increase is greater for more educated caregivers. Table 14 lists tor of the intercept when number of words and number of noun
the relevant means, standard deviations, and sample sizes. Most phrases are analyzed separately for one- and two-clause sentences.
complex sentences were two-clause sentences with subordinate That is, caregivers who provide a child with sentences of differing
clauses that fulfill the role of object for the main clause. The main levels of complexity at one syntactic level also do so at another.
clause involves verbs of cognition or motivation (e.g., think, want). Another issue concerns the potential for confounding the raw
A possible reason why this measure shows a steeper slope for more
amount of speech and complexity of speech. Even though our
educated groups is that, while educated caregivers generally tend
composition measures are relative to total numbers of sentences,
to use more multiclause sentences, they do not do so with children
some of the forms are relatively rare. Therefore, the more speech
too young to interpret them.
is sampled, the greater the likelihood of finding such forms. Thus,
As the complexity of speech was greater for more educated
differences in the quantity of speech may explain some variations
groups both at clausal and constituent levels, the question arises as
to whether differences at a constituent level may be due entirely to associated with education. For that reason, we have conducted
differences at a clausal level. While measures of constituent struc- additional analyses with our composition measures. For composi-
ture surely increase with the number of clauses, they also can tion measures, if we control for the amount of raw speech by
increase when the number of clauses is held constant, for example, including number of utterances as a covariate, the results of tests
by adding prepositional phrases or adjectival modification. Tables for the effect of education do not change. Education remains a
15 and 16 show that values on constituent-level measures are not significant predictor of the slope for proportion of complex sen-
based solely on the number of clauses in caregiver speech. Table tences and of the intercept for noun phrases per sentence and
15 shows the number of noun phrases separately for one-clause words per sentence. The growth curves are qualitatively similar to
and two-clause sentences for the four educational groups. While the patterns observed when number of utterances is not controlled.
the average number of noun phrases is, of course, greater in Thus, it is not possible to argue that educational differences in the
SPEECH TO YOUNG CHILDREN 1071

Figure 3. Modeled growth in sentences for different educational levels with and without older siblings present.

composition measures are due to differences in raw amount of speech. When number of utterances is included as a time-varying
speech. covariate to control for variation in amount of speech, education
For the diversity measures (word types and sentence types), we and siblings are no longer significant predictors for the diversity
obtained a somewhat different result in relation to quantity of measures. For word types, the best fitting model is one in which

Table 7
Summary Statistics for Number of Word Tokens

Age in months

Education 14 18 22 26 30

Graduate degree
M 4,093 3,881 3,887 3,697 3,984
SD 1,804 2,083 1,617 1,868 1,784
n 17 17 15 16 15
Bachelor’s degree
M 2,803 2,126 2,892 3,566 3,778
SD 1,178 1,089 1,388 1,350 1,688
n 18 18 17 18 16
Some college
M 2,802 3,211 3,139 2,895 3,231
SD 1,783 2,043 2,199 1,973 1,937
n 9 9 9 9 9
High school
M 1,572 1,381 1,563 2,145 1,686
SD 1,484 1,365 1,034 1,620 953
n 6 6 6 5 6
1072 HUTTENLOCHER ET AL.

Table 8
Summary Statistics for Number of Utterances

Age in months

Education 14 18 22 26 30

Graduate degree
M 1,089.4 1,016.0 973.5 873.8 902.1
SD 447.1 483.5 417.3 438.3 413.5
n 17 17 15 16 15
Bachelor’s degree
M 804.9 889.0 788.8 890.1 878.2
SD 323.6 296.0 373.1 330.7 393.8
n 18 18 17 18 16
Some college
M 757.2 877.3 786.4 718.9 765.8
SD 372.3 443.4 497.7 429.4 398.6
n 9 9 9 9 9
High school
M 413.8 413.3 454.5 539.2 436.5
SD 333.2 367.3 291.3 389.7 218.2
n 6 6 6 5 6

the expected value is 153.67 at 14 months, and the measure predict differences in intercept.) Second, there are major differ-
increases by 18.34 each 4 months and by 0.20 for each utterance. ences between quantitative and complexity measures. Only com-
For sentence types, the expectation at 14 months is 2.37, and the plexity increases over time. The fact that complexity increases
measure increases by 1.10 for each additional 4 months and by over time but that there are long-standing differences across edu-
0.004 for each utterance. The interpretation for these diversity cational groups suggests that two factors may determine the com-
measures is ambiguous; the differences observed in the curves in plexity of caregiver speech. One involves a sensitivity on the part
Figures 4 and 5 may be genuine education and sibling effects or of the caregiver to the listener’s maturity, and the other involves
may reflect the association of raw amount of speech with educa- long-term differences that are related to educational level. Two
tion and presence of siblings. This ambiguity is not surprising additional tests are relevant to evaluating the differences between
since diversity is, in part, a quantity measure. complexity and quantity.
Two conclusions are suggested by the analyses thus far. First,
differences in educational level affect most aspects of caregiver Rank Order of Caregivers: Stability Over Time
speech in a similar way. That is, while the intercept varies, the
slope generally does not. (The exceptions to this pattern are pro- The subgroup differences we have found suggest that caregivers
portion of complex sentences, where slope varies with education, may maintain a similar rank on particular measures over time. That
and possibly diversity measures, where education may or may not is, a caregiver who is high on a particular variable at the first

Table 9
Summary Statistics for Number of Sentences

Age in months

Education 14 18 22 26 30

Graduate degree
M 755.1 699.2 664.1 615.4 625.5
SD 332.0 372.0 277.6 293.1 268.1
n 17 17 15 16 15
Bachelor’s degree
M 519.4 565.5 511.6 596.4 601.6
SD 201.6 188.9 231.6 225.2 276.1
n 18 18 17 18 16
Some college
M 530.7 604.8 583.0 525.8 577.1
SD 280.8 337.7 384.8 316.8 306.1
n 9 9 9 9 9
High school
M 295.8 286.3 314.2 395.8 309.7
SD 242.9 256.0 189.7 262.7 155.4
n 6 6 6 5 6
SPEECH TO YOUNG CHILDREN 1073

Figure 4. Modeled growth in word types for different educational levels with and without older siblings
present.

observation also may be high on that variable at the other time that concordance is as high within educational groups as overall.
points. To investigate the stability of individual characteristics of Thus, the tendency to maintain relative position across caregivers
caregivers over time, we computed Kendall’s coefficient of con- is due to factors in addition to educational level.
cordance (W) for each measure, as shown in Table 17. High
concordance indicates that individuals tend to maintain their rank Interrelations Among Measures
order relative to one another across time points.
There is a marked contrast in stability over time for different To further address the question of whether the pattern of inter-
measures. In particular, for measures of quantity (number of to- relations among the different measures of caregiver speech are
kens, utterances, and sentences), rank order was not maintained stable across time, we used a principal-components analysis, a
well across sessions (for all measures, p ⬎ .30), whereas for method in which information in the correlations among measures
measures of complexity (composition and diversity), there is a is used to create a new set of derived variables (factors), which
high degree of concordance (for all measures, p ⬍ .01). That is, for explain the variation in the original variables. When the bulk of
complexity, but not quantity, individual caregivers tend to main- variation among the measures can be explained using a small
tain their positions relative to others over time. Thus, while all number of factors, we interpret these factors as the underlying
caregivers adjust the complexity of their speech to the growth of features that the original measures have in common.
their child, they do so within a certain range that identifies them as Although the sample size is rather small, our analyses suggest
individuals. that two principal factors are sufficient to account for most of the
The question arises as to whether the concordance we found variation in the measures. The first factor represents quantity
might be due entirely to differences in education. To evaluate this measures, and the second factor represents compositional mea-
possibility, we examined concordance within education groups. If sures. Table 18 gives the rotated loadings representing the relation
the consistency of order for caregivers were due solely to educa- of each measure to the two factors. The loadings for number of
tion, then there would be no reason to expect concordance within tokens, utterances, and sentences on Factor 1 are all much higher
educational groups. Table 17 presents the indices of concordance than their corresponding loadings on Factor 2 at every time point.
both overall and broken down by level of education. The data show Similarly, the loadings for proportion of complex sentences, num-
1074 HUTTENLOCHER ET AL.

Figure 5. Modeled growth in sentence types for different educational levels with and without older siblings
present.

ber of noun phrases, and words per sentence on Factor 2 are always terns. Word types, while having larger loadings on Factor 1 at all
much larger than their corresponding loadings on Factor 1. time points, also have nonnegligible loadings on Factor 2. Sen-
The pattern of loadings for the diversity measures, word types tence types also have relatively large loadings on both factors.
and sentence types, are intermediate between the other two pat- Thus, these measures have some characteristics of the quantitative

Table 10
Summary Statistics for Word Types

Age in months

Education 14 18 22 26 30

Graduate degree
M 385.6 371.5 397.7 397.1 425.5
SD 97.5 99.9 99.0 120.3 102.9
n 17 17 15 16 15
Bachelor’s degree
M 321.3 346.2 340.2 393.2 430.3
SD 99.2 86.5 103.7 107.5 104.5
n 18 18 17 18 16
Some college
M 321.4 357.7 363.4 351.0 389.3
SD 132.5 134.4 148.7 133.7 118.9
n 9 9 9 9 9
High school
M 232.5 211.2 242.8 289.0 265.5
SD 136.3 102.3 89.1 131.4 86.9
n 6 6 6 5 6
SPEECH TO YOUNG CHILDREN 1075

Table 11
Summary Statistics for Number of Sentence Types

Age in months

Education 14 18 22 26 30

Graduate degree
M 7.12 8.41 9.33 10.13 12.07
SD 2.89 3.79 3.62 5.29 5.65
n 17 17 15 16 15
Bachelor’s degree
M 5.89 6.83 6.82 9.78 11.13
SD 2.08 2.77 2.83 3.34 3.36
n 18 18 17 18 16
Some college
M 7.22 8.11 9.44 8.22 10.78
SD 3.93 5.04 5.98 5.72 5.76
n 9 9 9 9 9
High school
M 5.50 4.17 4.17 6.40 7.5
SD 4.55 2.48 0.98 3.29 3.73
n 6 6 6 5 6

measures of language input as well as some of the characteristics Discussion


of complexity measures. Because approximately the same pattern
appeared in the analysis of principal components for all time In this study, we have obtained longitudinal data on the speech
points, the relational structure of these measures appears to be of caregivers from a wide range of socioeconomic groups at
stable over time, a conclusion that is consistent with our interpre- several time points during the early period of language develop-
tation of the HLM analyses. ment. The study has allowed us to answer certain basic questions

Figure 6. Modeled growth in noun phrases per sentence for different educational levels.
1076 HUTTENLOCHER ET AL.

Figure 7. Modeled growth in words per sentence for different educational levels.

that remained unanswered because of lack of systematic data on complexity or diversity as we have found. Our findings suggest a
caregiver speech. It has been commonly believed that differences different picture of the nature of these differences.
in caregiver speech are unsystematic and not important to acqui- We developed measures to capture major characteristics of
sition. Any speech differences that endure over time, it has been caregiver speech and used these measures to examine whether
claimed, are differences in idiolect, not differences in syntactic there are systematic changes in that speech as children become

Table 12
Summary Statistics for Number of Noun Phrases per Sentence

Age in months

Education 14 18 22 26 30

Graduate degree
M 1.46 1.52 1.58 1.65 1.72
SD 0.12 0.18 0.15 0.20 0.22
n 17 17 15 16 15
Bachelor’s degree
M 1.43 1.47 1.51 1.58 1.70
SD 0.13 0.18 0.19 0.20 0.19
n 18 18 17 18 16
Some college
M 1.34 1.35 1.47 1.50 1.48
SD 0.21 0.29 0.24 0.17 0.25
n 9 9 9 9 9
High school
M 1.22 1.17 1.34 1.47 1.52
SD 0.34 0.28 0.11 0.12 0.16
n 6 6 6 5 6
SPEECH TO YOUNG CHILDREN 1077

Table 13
Summary Statistics for Number of Words per Sentence

Age in months

Education 14 18 22 26 30

Graduate degree
M 4.68 4.75 5.08 5.22 5.55
SD 0.49 0.41 0.44 0.61 0.74
n 17 17 15 16 15
Bachelor’s degree
M 4.56 4.64 4.82 5.18 5.54
SD 0.43 0.54 0.56 0.52 0.51
n 18 18 17 18 16
Some college
M 4.37 4.40 4.66 4.74 4.92
SD 0.64 0.78 0.62 0.63 0.80
n 9 9 9 9 9
High school
M 4.20 3.95 4.29 4.69 4.71
SD 0.93 0.64 0.34 0.42 0.76
n 6 6 6 5 6

older and to examine whether there are substantial individual diversity of speech increased as children became older, whereas
differences in speech among caregivers across time. We used quantity of speech remained constant. There were substantial
compositional measures that capture many important syntactic individual differences for all measures at all time points, and
characteristics of speech, diversity measures that capture vari- educational level was related to caregiver speech for these
ety of speech, and simple quantity measures. Complexity and measures.

Figure 8. Modeled growth in multiclause sentences for different educational levels.


1078 HUTTENLOCHER ET AL.

Table 14
Summary Statistics for Proportion of Complex Sentences

Age in months

Education 14 18 22 26 30

Graduate degree
M .084 .087 .117 .133 .151
SD .031 .026 .035 .040 .045
n 17 17 15 16 15
Bachelor’s degree
M .077 .078 .100 .118 .148
SD .024 .031 .035 .035 .038
n 18 18 17 18 16
Some college
M .078 .096 .098 .097 .130
SD .035 .042 .035 .054 .049
n 9 9 9 9 9
High school
M .075 .075 .083 .102 .100
SD .042 .014 .028 .033 .056
n 6 6 6 5 6

While the syntax of a language provides tools that, in principle, young children have reported differences from usual adult speech
permit construction of sentences of arbitrary levels of complexity, patterns. Indeed, such child-directed speech is referred to as mother-
the actual sentences people produce are limited in complexity. The ese and is said to involve a special register in which speech is
limitations have been ascribed to limits in planning, memory, and simplified and intonation is exaggerated (e.g., Newport et al., 1977;
so forth. Our findings provide new and systematic information Snow, 1972). Such alterations in speech presumably decrease with
about caregiver speech indicating that there are clusters of indi- age, reflecting caregiver adjustment to increasing skills in the child.
viduals whose speech falls into different complexity groups. Dif- However, studies that have examined speech to young children lon-
ferent educational groups can be characterized by the range of gitudinally have failed to find convincing evidence of such changes
complexity over which their speech varies. Thus, when one de- with age. Some studies have failed to find differences in complexity
scribes the language input that children are exposed to, it seems
(e.g., Furrow et al., 1979; Newport et al., 1977), and others have failed
most accurate to posit different working levels of complexity in
to find differences in utterance length (e.g., Kavanaugh & Jirkovsky,
different caregivers—levels that characterize their habitual forms
1982; Newport et al., 1977). Note, though, that the lack of significant
of verbal communication with young children.
results in the earlier longitudinal studies might be due to the fact that
the age ranges studied were too narrow. Cross-sectional studies often
Change in Caregiver Speech Over Time
do show changes with age, but the sample sizes they have used have
With respect to change in caregiver speech over time, earlier been too small to evaluate whether there are systematic changes in
studies have obtained mixed results. Some studies of speech to very caregiver speech related to child age.

Table 15
Average Number of Noun Phrases in One- and Two-Clause Utterances

Session

Educational group 1 2 3 4 5 All sessions

One clause

Graduate degree 1.37 1.42 1.44 1.50 1.53 1.45


Bachelor’s degree 1.34 1.38 1.39 1.44 1.53 1.41
Some college 1.26 1.23 1.36 1.39 1.37 1.32
High school 1.14 1.09 1.25 1.37 1.43 1.26
All 1.31 1.33 1.38 1.44 1.49 1.39

Two clauses

Graduate degree 2.38 2.43 2.49 2.54 2.58 2.48


Bachelor’s degree 2.41 2.47 2.51 2.66 2.68 2.55
Some college 2.26 2.40 2.34 2.41 2.39 2.36
High school 2.24 2.12 2.23 2.35 2.24 2.24
All 2.35 2.40 2.43 2.54 2.53 2.45
SPEECH TO YOUNG CHILDREN 1079

Table 16
Average Number of Words in One- and Two-Clause Utterances

Session

Educational group 1 2 3 4 5 All sessions

One clause

Graduate degree 4.37 4.42 4.62 4.69 4.91 4.60


Bachelor’s degree 4.28 4.36 4.41 4.67 4.91 4.52
Some college 4.11 4.04 4.32 4.38 4.39 4.25
High school 3.91 3.73 4.02 4.34 4.39 4.08
All 4.23 4.25 4.41 4.59 4.74 4.44

Two clauses

Graduate degree 7.79 7.75 8.16 8.24 8.39 8.07


Bachelor’s degree 7.60 7.62 8.00 8.45 8.50 8.03
Some college 6.90 7.42 7.38 7.69 7.81 7.44
High school 7.53 6.49 6.94 7.61 7.21 7.15
All 7.45 7.32 7.62 8.00 7.98 7.67

The present longitudinal study, involving a large sample of tery with age. Regardless of the explanation for caregivers’ ad-
families, shows substantial change in caregiver speech over time. justments of their speech, our data indicate that they are substan-
Whereas most previous studies have investigated only a handful of tial, indicating that questions as to whether parents fine tune their
syntactic structures, the present study has explored a wide range of speech are important to explore. More exact information about the
speech characteristics. We have found significant changes in syn- sources of change in caregiver speech can be obtained from future
tactic complexity and diversity over time, but no change in quan- studies that examine both caregiver and child speech, making it
tity. This pattern suggests that changes in caregiver speech are not possible to determine the extent to which variation in caregiver
simply due to factors such as greater motivation or interest in speech can be attributed to children’s language level versus their
talking to older children. Rather, while the caregivers produce age.
roughly the same quantity of speech in a conversation, the com-
position of their speech changes. The increase in syntactic com- Individual Differences Among Caregivers
plexity over time indicates a sensitivity of caregivers to children’s
language levels and, at the same time, provides children with With respect to individual differences among caregivers, we
exposure to a wider range of the syntactic devices available in the have found substantial differences that are maintained over time.
language as children proceed in mastering the syntax of their Earlier studies did not generally examine caregivers longitudinally,
language. so that questions about longstanding differences could not be
While the observed changes in caregiver speech provide evi- directly addressed. We observed differences on all measures at
dence of child effects on how caregivers talk, the specific causes every time point. When we modeled individual differences among
are not yet clear. Caregivers may be reacting to the particular caregivers on our diverse set of measures, including composition
production or comprehension levels of a child at a given time, or of syntax and the lexicon, syntactic and lexical diversity, and
they might only be making rough adjustments to more general quantity of speech, we found differences as a function of group-
child characteristics in anticipation of increases in syntactic mas- level factors, in particular, with educational level. More educated

Table 17
Kendall’s Coefficient of Concordance for Each Measure, Overall and by Education Level

Educational level

Measure Overall 1 2 3 4

Word tokens .017 .088 .032 .126 .078


Utterances .024 .064 .072 .086 .254
Sentences .013 .010 .032 .091 .182
Word types* .094 .136 .121 .204 .075
Sentence types** .274 .247 .163 .502 .325
Complex sentences** .415 .204 .416 .599 .561
Noun phrases per sentence** .398 .384 .503 .383 .449
Words per sentence** .551 .352 .521 .697 .555
*
p ⬍ .01. **
p ⬍ .001.
1080 HUTTENLOCHER ET AL.

Table 18
Varimax-Rotated Principal Components at Each Time Point

Loadings on two components by age

14 months 18 months 22 months 26 months 30 months

Factor Factor Factor Factor Factor

Measure 1 2 1 2 1 2 1 2 1 2

Word tokens 0.95 0.28 0.95 0.29 0.96 0.26 0.94 0.34 0.95 0.28
Utterances 0.98 0.14 0.97 0.15 0.98 0.07 0.98 0.15 0.98 0.07
Sentences 0.97 0.18 0.97 0.17 0.99 0.11 0.97 0.21 0.99 0.11
Word types 0.83 0.47 0.80 0.50 0.87 0.40 0.83 0.49 0.89 0.35
Sentence types 0.46 0.66 0.70 0.56 0.64 0.60 0.48 0.75 0.58 0.67
Complex sentences 0.04 0.92 0.06 0.84 0.09 0.88 0.11 0.90 0.16 0.90
Noun phrases per sentence 0.32 0.84 0.37 0.83 0.20 0.87 0.38 0.81 0.09 0.88
Words per sentence 0.26 0.91 0.32 0.88 0.22 0.92 0.20 0.90 0.21 0.94

caregivers spoke more and with greater syntactic complexity and that arise in examining the issue. Clearly, there are alternative
diversity than did less educated caregivers. Earlier studies have possible explanations for the substantial and longstanding differ-
treated SES as a single variable including both income and edu- ences in complexity of speech we have observed across caregivers.
cation. In the present study, we examined income and education These differences could reflect biological variations in ability,
separately and found that, while the two factors are highly related, variation in the input different caregivers received as children, or
education was more closely associated with characteristics of other factors such as general styles of speech, beliefs in children’s
parent speech. ability to understand, and so forth.
Variation in complexity of caregiver speech for different edu- When one considers the respective roles of biology versus early
cational groups occurred at all levels of syntactic complexity input in determining caregiver speech, it should be noted that the
examined. That is, caregivers who expressed themselves using sources of differences among caregivers and children should be
higher proportions of multiclause sentences also used more noun parallel. That is, if caregivers differ because of a biological (ge-
phrases and words per clause, and they produced a greater diver- netic) relation to their parents, that difference should also be seen
sity of clausal-level structures. When the number of clauses was in their relation to their children. If they differ because of variation
controlled, there still were substantial individual differences in in the input they received as children, the speech they direct to
constituent-level complexity (numbers of noun phrases and their children should be related to their own prior input. In short,
words). As in earlier studies, we found that the presence of older intergenerational links that connect caregivers to the previous
siblings at a session does not affect the composition of speech even generation (their parents) also connect caregivers to the following
though it has a major effect on the quantity of speech. generation (their children).
Our results suggest that differences in the composition and There are data to indicate that intergenerational links are, at least
diversity of syntax reflect long-standing characteristics of caregiv- in part, input driven, ruling out extreme biological interpretations.
ers. Other findings in the literature support such a view. In partic- Hoff-Ginsberg (1998) found that birth order is related to the
ular, there are studies showing that educational level is related to language caregivers direct to their children. Since the biological
a person’s speech to other adults (e.g., Hoff, 2003b; Van den
relation of caregiver and child is the same across siblings, this
Broeck, 1977). Miller and Weinert (1998) and Ravid and Tolchin-
finding indicates that the role of input in parent– child interaction
sky (2002) have found that amount of education is specifically
is distinct from genetic factors. Further, Huttenlocher et al. (2002)
related to overall complexity of speech. In future studies, samples
found correlations of teacher speech with children’s syntactic
of caregiver speech should be obtained when they address other
growth over a school year. The speech of the teachers was uncor-
adults as well as children. The participants should be from a broad
related with children’s levels at the start of the school year but was
range of backgrounds, and samples should include contexts in-
correlated at the end of the year. These findings clearly reflect
volving a variety of topics and situations. Note, though, that the
effects of input on syntactic growth.
important point here is that children encounter very different
In considering the role of input, imagine two caregivers whose
linguistic environments and that some children may not receive
speech differs in complexity. One caregiver uses mainly one-
sufficient exposure to complex forms to become proficient with
their use. clause sentences that rarely specify location or time of an event.
The other caregiver uses mainly multiclause sentences that mark
the time and place of events, for example, “The lady got the book
The Sources of Individual Differences
that had a beautiful binding at Barnes & Noble last week.” Clearly
The findings of significant long-lasting differences in caregiver the child who receives input from the latter caregiver will have
speech reported here bring up questions concerning the possible more experience with complex speech and may find it easier to
sources of such differences. While our empirical data do not allow understand and produce such speech. If repeated exposure affects
us to identify such sources, let us consider general logical issues the probability of using particular grammatical forms, it could
SPEECH TO YOUNG CHILDREN 1081

explain the relation we have observed between input and acquisi- input and acquisition (pp. 199 –205). Cambridge, England: Cambridge
tion. To be fluent with complex syntax, an individual may require University Press.
frequent embellishment of simple sentences into sentences that Bybee, J. (1998). A functionalist approach to grammar and its evolution.
combine clauses, use prepositional and adverbial phrases, and so Evolution of Communication, 2, 249 –278.
forth. Chomsky, N. (1986). Knowledge of language: Its nature, origin, and use.
New York: Praeger.
While existing research shows the importance of exposure,
Dunn, L. M., & Dunn, L. M. (1997). Peabody Picture Vocabulary Test (3rd
biological factors also may be important. Differences in caregiver ed.). Circle Pines, MN: American Guidance Service.
speech may reflect biological differences in the ease with which Elardo, R., Bradley, R., & Caldwell, B. (1977). A longitudinal study of the
different people acquire language from input. Our data are consis- relation of infants’ home environments to language development at age
tent with this possibility. That is, the caregivers we observed in the three. Child Development, 4, 595– 603.
present longitudinal study differed at the start (intercept differ- Elman, J. L. (1993). Learning and development in neural networks: The
ences), and these differences remained parallel as children became importance of starting small. Cognition, 48, 71–99.
older (common slope). The observed differences in start levels and Farian, D. C., & Haskins, R. (1980). Reciprocal influence in the social
the parallel increases over time could reflect both the genetic interaction of mothers and three-year-old children from different socio-
differences among parents and the genetic similarities of parents economic backgrounds. Child Development, 51, 780 –791.
and children. Furrow, D., Nelson, K., & Benedict, H. (1979). Mothers’ speech to
children and syntactic development: Some simple relationships. Journal
A model that allows differences in input to affect acquisition can
of Child Language, 6, 423– 442.
also accommodate biological factors without major revision. For
Gleitman, L., Newport, E., & Gleitman, H. (1984). The current status of the
example, some individuals may benefit more from exposure than motherese hypothesis. Journal of Child Language, 11, 43–79.
others; for these individuals, exposure effects might be longer Goldberg, A. (1995). Constructions: A construction grammar approach to
lasting and contribute more to likelihood of using complex forms argument structure. Chicago: University of Chicago Press.
and so forth. In short, the mechanisms involved in acquisition may Hammer, C., & Weiss, A. (1999). Guiding language development: How
be similar across children but vary in the relative weights of African-American mothers and their infants structure play interactions.
biological versus input factors. In any case, our findings indicate Journal of Speech, Language, and Hearing Research, 42, 1219 –1233.
long-lasting individual differences in the complexity of caregiver Hart, B., & Risley, T. R. (1992). American parenting of language-learning
speech. Thus, the relation of individual differences in the speech of children: Persisting differences in family– child interactions observed in
caregivers to child acquisition should be examined systematically. natural home environments. Developmental Psychology, 28, 1096 –
1105.
Heath, S. B. (1982). What no bedtime story means: Narrative skills at home
Conclusions and school. Language in Society, 11, 49 –76.
Hess, R. D., & Shipman, V. C. (1965). Early experience and the social-
The present study provides important information on caregiver ization of cognitive modes in children. Child Development, 36, 859 –
speech to very young children. Our behavioral data have estab- 886.
lished two different sources of variability in speech to young Hoff, E. (2003a). Causes and consequences of SES-related differences in
children. First, caregivers modify their speech depending on the parent-to-child speech. In M. H. Bornstein & R. H. Bradley (Eds.),
characteristics of the children they are addressing. Second, there Socioeconomic status, parenting, and child development (pp. 147–160).
are substantial and long-lasting differences in the speech of dif- Mahwah, NJ: Erlbaum.
ferent caregivers. The long-lasting individual differences involve Hoff, E. (2003b). The specificity of environmental influence: Socioeco-
variations in syntactic complexity that are tied to the caregivers’ nomic status affects early vocabulary development via maternal speech.
educational level. While caregivers make adjustments in their Child Development, 74, 1368 –1378.
Hoff-Ginsberg, E. (1986). Function and structure in maternal speech: Their
speech depending on the child, they nevertheless retain their indi-
relation to the child’s development of syntax. Developmental Psychol-
vidual speech patterns over time. The findings of the present study
ogy, 22, 155–163.
provide a framework for systematic naturalistic study of Hoff-Ginsberg, E. (1998). The relation of birth order and socioeconomic
caregiver– child relations to further evaluate the role of input in status to children’s language experience and language development.
language acquisition. Applied Psycholinguistics, 19, 603– 629.
Huttenlocher, J., Vasilyeva, M., Cymerman, E., & Levine, S. (2002).
Language input and child syntax. Cognitive Psychology, 45, 337–374.
References
Jackendoff, R. (2002). Foundations of language: Brain, meaning, gram-
Andersen, H. (1973). Abductive and deductive change. Language, 49, mar, evolution. Oxford, England: Oxford University Press.
765–793. Jones, C., & Adamson, L. B. (1987). Language use in mother– child and
Anderson, S. R. (1992). A-morphous morphology. New York: Cambridge other child–sibling interactions. Child Development, 58, 356 –366.
University Press. Kavanaugh, R., & Jirkovsky, A. (1982). Parental speech to young children:
Baker, M. (2001). The atoms of language. New York: Basic Books. A longitudinal analysis. Merrill-Palmer Quarterly, 28, 297–311.
Barnes, S., Gutfreund, M., Satterly, D., & Wells, G. (1983). Characteristics Kaye, K. (1980). Why we don’t talk “baby talk” to babies. Journal of Child
of adult speech which predict children’s language development. Journal Language, 7, 498 –507.
of Child Language, 10, 65– 84. Lidz, J., Gleitman, H., & Gleitman, L. (2003). Understanding how input
Bee, H. L., Van Egeren, L. F., Streissguth, A. P., Nyman, B. A., & Leckie, matters: Verb learning and the footprint of universal grammar. Cogni-
M. A. (1969). Social class differences in maternal teaching strategies and tion, 87, 151–178.
speech patterns. Developmental Psychology, 1, 726 –734. Lidz, J., & Gleitman, L. (2004). Yes, we still need universal grammar:
Berko-Gleason, J. (1977). Talking to children: Some notes on feedback. In Reply. Cognition, 94, 85–93.
C. E. Snow & C. A. Ferguson (Eds.), Talking to children: Language Lidz, J., Waxman, S., & Freedman, J. (2003). What infants know about
1082 HUTTENLOCHER ET AL.

syntax but couldn’t have learned: Experimental evidence for syntactic Ravid, D., & Tolchinsky, L. (2002). Developing linguistic literacy: A
structure at 18 months. Cognition, 89, B65–B73. comprehensive model. Journal of Child Language, 29, 417– 447.
Miller, J., & Weinert, R. (1998). Spontaneous spoken language: Syntax Rondal, J. (1980). Fathers’ and mothers’ speech in early language devel-
and discourse. Oxford, England: Clarendon. opment. Journal of Child Language, 7, 353–369.
Naigles, L., & Hoff-Ginsberg, E. (1998). Why are some verbs learned Rowland, C., Pine, J., Lieven, E., & Theakston, A. (2003). Determinants of
before other verbs? Effects of input frequency and structure on chil- acquisition order in wh-questions: Re-evaluating the role of caregiver
dren’s early verb use. Journal of Child Language, 25, 95–120. speech. Journal of Child Language, 30, 609 – 635.
Nelli, C. (1998). L’acquisizione della morfologia libera italiana: Fasi di un Saffran, J. (2001). The use of predictive dependencies in language learning.
percorso evolutivo [The acquisition of free Italian morphology: Devel- Journal of Memory and Language, 44, 493–515.
opmental phases]. Studi di Grammatica Italiana, 17, 329 –362.
Slobin, D. I., & Bever, T. G. (1982). Children use canonical sentence
Newmeyer, F. (2003). Grammar is grammar and usage is usage. Language,
schemas: A crosslinguistic study of word order and inflections. Cogni-
79, 682–707.
tion, 12, 229 –265.
Newport, E. L., Gleitman, H., & Gleitman, L. (1977). Mother, I’d rather do
Snow, C. E. (1972). Mothers’ speech to children learning language. Child
it myself: Some effects and noneffects of maternal speech style. In C. E.
Snow & C. A. Ferguson (Eds.), Talking to children: Language input and Development, 43, 549 –565.
acquisition (pp. 109 –150). Cambridge, England: Cambridge University Snow, C. E. (1977). The development of conversation between mothers
Press. and babies. Journal of Child Language, 4, 1–22.
Oshima-Takane, Y., & Robbins, M. (2003). Linguistic environment of Snow, C. E. (1982). Are parents language teachers? In K. Borman (Ed.),
secondborn children. First Language, 23(67, Pt. 1), 21– 40. Social life of children in a changing society (pp. 81–95). Hillsdale, NJ:
Pan, B. A., Rowe, M., Spier, E., & Tamis-Lemonda, C. (2004). Measuring Erlbaum.
productive vocabulary of toddlers in low-income families: Concurrent Snow, C. E., Perlmann, R., & Nathan, D. (1987). Why routines are
and predictive. Journal of Child Language, 31, 587– 608. different: Toward a multiple-factors model of the relation between input
Phillips, J. R. (1973). Syntax and vocabulary of mothers’ speech to young and language acquisition. In K. Nelson & A. van Kleeck (Eds.), Chil-
children: Age and sex comparisons. Child Development, 44, 182–185. dren’s language (Vol. 6, pp. 65–97). Hillsdale, NJ: Erlbaum.
Price, G., & Hatano, G. (1991). Toward a taxonomy of the roles home Theakston, A. L., Lieven, E. V. M., Pine, J. M., & Rowland, C. F. (2005).
environments play in the formation of educationally significant individ- The acquisition of auxiliary syntax: Be and have. Cognitive Linguistics,
ual differences. In S. Silvern (Ed.), Advances in reading/language re- 16, 247–277.
search: Vol. 5. Literacy through family, community and school interac- Tomasello, M. (2003). Constructing a language: A usage-based theory of
tion (pp. 37– 62). Greenwich, CT: JAI Press. language acquisition. Cambridge, MA: Harvard University Press.
Raudenbush, S., Bryk, A., Cheong, Y. F., & Congdon, R. (2000). HLM 5: Van den Broeck, J. (1977). Class differences in syntactic complexity in the
Hierarchical linear and nonlinear modeling [Computer software and
Flemish town of Maaseik. Language in Society, 6, 149 –181.
manual]. Lincolnwood, IL: Scientific Software International.

Appendix

Details of the HLM Models

In the body of the article, we have focused on general results and cation, which permits each educational group to have a different
have relied on the figures to present detail about the forms of the mean growth trajectory. In this system, we estimate the intercept
change trajectories. Here, we present information about the parame- for one group (the group with the highest level of education) and
ters that describe the curves, as well as the variance components that express the intercepts of other groups as changes or differences
describe individual variation. Each HLM model estimates the change from that reference group. In addition, we consider models in
trajectory for some measure of caregiver speech using a regression which education predicts the slope (and, in the case of words per
equation that predicts the measure with an intercept and a slope that sentence, the quadratic slope). We investigate these explanatory
reflects change over time. We code time so that its value is zero at the models for any growth parameter that exhibits significant vari-
first observation, one at the second, and so on, ending at four for the ation across individual participants, regardless of whether the
fifth observation. In this system, the intercept is the expected value for average value of the parameter is significantly different from
the speech measure at the time of the first observation (14 months), zero. This allows us to investigate the possibility that different
and the slope represents the expected change between adjacent time education groups exhibit different growth rates as well as
points. For the one variable (words per sentence) that exhibited a different starting points and guards against the possibility that
slight curve in the change trajectory, a more complex model was used the flat growth trajectories for some outcomes result from a
in which a second slope reflected quadratic change over time. In situation in which differential growth for different groups av-
addition to the model for time, we consider one time-varying covari- erages to zero.
ate: the presence or absence of older siblings during the visit. The Although education explains some of the variations in the pa-
HLM analyses are roughly equivalent to estimating regression equa- rameters (in most cases the intercept, but for proportion of multi-
tions separately for each caregiver and summarizing the values of clause sentences, the slope), there are generally still unexplained
growth parameters across people. individual differences. The amount of the unexplained variation is
Next, we consider how those linear or quadratic models vary quantified by variance component estimates, which are the esti-
across individuals. We employ a dummy coding system for edu- mated variances of the individual differences (in intercept or slope)
SPEECH TO YOUNG CHILDREN 1083

Table A1
Growth Curve Parameters and Standard Errors for the Eight Language Measures

Word Number of Number of Word Sentence Complex Noun phrases Words per
Parameter tokens utterances sentences types types sentences per sentence sentence

Intercepts
For highest level of education 4,097.88 1,033.55 713.01 374.49 7.39 ⫺2.48 1.46 4.69
(advanced degree) (342.21) (79.36) (56.67) (23.56) (0.68) (0.05) (0.04) (0.13)
Change for bachelor’s degree ⫺607.06 ⫺101.43 ⫺101.19 ⫺24.67 ⫺1.34 ⫺0.01 ⫺0.04 ⫺0.13
(468.25) (109.67) (77.79) (32.08) (0.91) (0.03) (0.05) (0.17)
Change for some college ⫺870.38 ⫺214.15 ⫺118.97 ⫺40.23 0.09 ⫺0.06 ⫺0.15 ⫺0.35
(570.86) (133.13) (94.57) (38.90) (1.11) (0.04) (0.07) (0.20)
Change for high school only ⫺2,307.71 ⫺563.29 ⫺374.44 ⫺156.35 ⫺2.98 ⫺0.12 ⫺0.25 ⫺0.68
(660.67) (153.45) (109.37) (44.82) (1.29) (0.04) (0.08) (0.24)
Change if older siblings present ⫺1,009.50 ⫺274.40 ⫺190.44 ⫺58.36 ⫺1.69
(251.66) (56.91) (44.75) (14.92) (0.69)
Slopes
Linear change 15.91 1.03 0.19 0.06 0.08
(3.06) (0.14) (0.02) (0.01) (0.05)
Quadratic change 0.03
(0.01)
Amount of unexplained individual
differences
Square root of variance component for 1,348.91 306.88 221.20 93.21 2.12 0.36 0.16 0.50
intercept
Square root of variance component for 15.11 0.53 0.10 0.04 0.23
slope

Note. Standard errors are in parentheses below each estimate. For complex sentences, education increments are adjustments to the slope, not to the
intercept.

within educational groups. We report the square root of these reason this may arise is because the variable is a proportion
variance components, so that results are in the more interpretable (e.g., the proportion of complex sentences), in which case we
standard deviation metric. employ a logistic regression model rather than a conventional
Table A1 provides a summary of the results of our final anal- regression. Another reason transformations may be necessary is
yses. Each column of the table provides results for a different that the distribution of errors of prediction is not constant across
language measure. For each of these measures, the first four rows the range of model-predicted values. This was true for several
of the table present the estimates of the average intercept for the of our variables. Accordingly, we analyze the square root of
highest educational group and the differences in mean intercept number of tokens, number of utterances, number of sentences,
between that group and each of the other groups, along with the and diversity—number of word types and sentence types. In-
standard errors of these estimates (in parentheses). The fifth row of ferences for these variables are conducted in the square-root
the table presents the effects of the presence of older siblings. The metric. However, for the variables that were transformed by the
sixth and seventh rows of the table present the linear and quadratic square root, the parameters that we present are in the untrans-
slopes (if these are statistically significant) along with their stan- formed metric, as these untransformed models are more easily
dard errors (in parentheses). The last two rows of the table present interpreted, and the transformed and untransformed models
the variation of individual differences in intercepts or slopes (when result in very similar predictions.
they are statistically significant) within educational groups, ex- The process of fitting the models and exploring moderating
pressed as standard deviations (square roots of the variance com- variables involved trying a number of different possible models for
ponents). Note that the differences in intercept associated with each of our measures. In the interests of parsimony, we sought to
caregiver education are very large in comparison with the remain- reduce these to models in which only significant coefficients
ing individual differences. For the quantitative measures, the dif- remained. Table A1 presents the parameter estimates for each of
ference in intercept between the highest and lowest educational the measures that resulted from that process; these are the param-
groups is approximately five standard deviations (five times the eter values that were used to produce the figures depicting the
variance component for the intercept). For the compositional mea- best-fitting change trajectories.
sures, the difference in intercept between the highest and lowest
educational groups is approximately 10 standard deviations (10
times the variance component for the intercept). Received September 2, 2005
For some of the variables, transformations or other special Revision received August 18, 2006
handling can improve the validity of statistical inference. One Accepted August 29, 2006 䡲

Anda mungkin juga menyukai