dearing et al 2009

8/13/2019 Dearing Et Al 2009

1/21

Does Higher Quality Early Child Care Promote Low-Income Childrens

Math and Reading Achievement in Middle Childhood?

Eric Dearing

Boston College

Kathleen McCartney

Harvard Graduate School of Education

Beck A. TaylorSamford University

Higher quality child care during infancy and early childhood (654 months of age) was examined as a moder-ator of associations between family economic status and childrens ( N= 1,364) math and reading achievementin middle childhood (4.511 years of age). Low income was less strongly predictive of underachievement forchildren who had been in higher quality care than for those who had not. Consistent with a cognitive advan-tage hypothesis, higher quality care appeared to promote achievement indirectly via early school readiness

skills. Family characteristics associated with selection into child care also appeared to promote the achieve-ment of low-income children, but the moderating effect of higher quality care per se remained evident whencontrolling for selection using covariates and propensity scores.

By fifth grade, poor children are as much as 2 timesmore likely to lack proficiency in math and readingskills than children who are not poor (U.S. Depart-ment of Education, National Center for EducationStatistics, 2007). In turn, these developmental disad-vantages contribute to the societal costs of childhoodpoverty, estimated to be as high as $500 billion per

year, through processes such as reduced lifetimeearnings (Holzer, Schanzenbach, Duncan, & Lud-wig, 2007). Importantly, achievement in middlechildhood appears to be a particularly strong predic-tor of earnings in adulthood for children who havegrown up poor (Feinstein & Bynner, 2004). Yet,higher quality early child care may act as an inter-vention, helping promote the achievement of low-income children through middle childhood.

Higher Quality Early Child Care: A CognitiveAdvantage for Low-Income Children?

Following more general theory on the signifi-cance of early rearing environments, child-care

researchers have argued that nonmaternal careexperiences may stimulate and support healthydevelopment if care is of high quality, as evidencedby: (a) high levels of language stimulation, (b) accessto developmentally appropriate learning materials,(c) a positive emotional climate with sensitive andresponsive caregivers, and (d) opportunities for chil-

dren to explore their environments (e.g., McCartney,1984; National Institute of Child Health and HumanDevelopment [NICHD] Early Child Care ResearchNetwork, 2000). Compared with children whoattend relatively lower quality child care duringearly childhood, children who attend higher qualitychild care do, in fact, display higher average levelsof mathematics and reading achievement from earlychildhood through adolescence as well as higherearnings in adulthood (e.g., Barnett, 1995; Campbell,Ramey, Pungello, Sparling, & Miller-Johnson, 2002;Lamb, 1998; NICHD Early Child Care Research

Network & Duncan, 2003; Schweinhart et al., 2005).Beyond these average group differences, higherquality child care may be an effective means ofpromoting the achievement of children growing uppoor; higher quality child care may help compen-sate for the otherwise deprived developmentalcontexts in which many poor children live.

The authors wish to thank Ariel Kalil, Pamela Morris, Eliza-beth Votruba-Drzal, and two anonymous reviewers for theirinsightful comments and suggestions on previous versions ofthis manuscript.

Correspondence concerning this article should be addressed toEric Dearing, Counseling, Developmental, and Educational Psy-chology, Lynch School of Education, Boston College, 140 Com-monwealth Ave, Chestnut Hill, MA 02467. Electronic mail maybe sent to [email protected].

Child Development, September/October 2009, Volume 80, Number 5, Pages 13291349

2009, Copyright the Author(s)Journal Compilation2009, Society forResearch in ChildDevelopment, Inc.

All rights reserved. 0009-3920/2009/8005-0003

8/13/2019 Dearing Et Al 2009

2/21

Low family income also has consequences forchildrens achievement, in part because povertyplaces constraints on families investments in mate-rial resources (e.g., books) necessary for cognitiveand language development (Becker & Tomes, 1986;Dearing & Taylor, 2007; Votruba-Drzal, 2003;Yeung, Linver, & Brooks-Gunn, 2002). Moreover,economic pressures associated with poverty impairparent psychological well-being, thereby decreasingpositive parenting behaviors (e.g., stimulation, sup-port, and responsiveness) and increasing negativeparenting behaviors (e.g., harsh andor inconsistentresponses; Conger, Rueter, & Conger, 2000; Dear-ing, Taylor, & McCartney, 2004; Elder & Caspi,1988). Yet, both of these pathways may be inter-rupted, or at least mitigated, by higher qualitychild-care experiences.

Higher quality child care may offer direct bene-

fits to low-income children through material (e.g.,learning materials) and psychosocial (e.g., stimulat-ing and responsive caregivers) investments thatcompensate for limited investments provided inhome environments. Moreover, higher quality childcare may have indirect benefits for children, pro-viding parents informal and formal social supportsthat, in turn, are translated into more material andpsychosocial investments in children within theirhome environments (McCartney, Dearing, Taylor,& Bub, 2007). Regarding math and reading achieve-ment, improved access to learning materials andlearning stimulation may be particularly crucial forlow-income children because deprivation in thisarea is the primary mechanism by which lowincome leads to underachievement (e.g., Yeunget al., 2002). More specifically, by directly andindirectly improving their learning environments,higher quality child care may provide a cognitiveadvantage for low-income children, promotingtheir acquisition of early skills and knowledge thatare a prerequisite for later success in math andreading (Burchinal, Peisner-Feinberg, Bryant, &Clifford, 2000; McCartney et al., 2007; Reynolds,Ou, & Topitzes, 2004). As has been posited for early

education interventions, higher quality child caremay promote childrens learning of basic numericalknowledge and literacy skills and thereby set thestage for a positive cycle of performance throughthe school years (Reynolds et al., 2004, p. 1300).

Evidence From Randomized, Matched-Groups, andNonrandomized Designs

Model child-care programs (e.g., AbecedarianProject and HighScope Perry Preschool) and early

education intervention programs have proven effec-tive for promoting the achievement of low-incomechildren (Barnett, 1995), with evaluation evidencefrom studies using random assignment (Campbell,Pungello, Miller-Johnson, Burchinal, & Ramey,2001; Love et al., 2005; Schweinhart, Barnes, Weik-art, Barnett, & Epstein, 1993) and nonrandomizedmatched-groups designs (Reynolds, Temple, Rob-ertson, & Mann, 2002). These positive effects onlow-income childrens achievement persist throughadolescence and into adulthood, as evidenced byintelligence scores as well as educational attain-ment, employment, and earnings (e.g., Campbellet al., 2002; Schweinhart et al., 2005; Smokowski,Mann, Reynolds, & Fraser, 2005). In addition,experimental evaluations of poverty reductioninterventions that include child-care subsidies havedemonstrated benefits for poor childrens achieve-

ment through middle childhood and into adoles-cence (Huston et al., 2006).

It is not clear, however, whether model child-careprograms could be effectively scaled up while retain-ing the quality these programs evidenced under thedirection of researchers. By design, these model pro-grams also included theoretically informed curriculatargeting childrens cognitive growth. It is difficult,therefore, to generalize from these programs to childcare more commonly available to low-income fami-lies, even though some low-income families maymanage to access relatively good care (Phillips,Voran, Kisker, Howes, & Whitebrook, 1994). In addi-tion, early education and poverty reduction interven-tions have generally included comprehensive childand family services that extend beyond child care(e.g., home visits); disentangling the unique effects ofchild-care quality per se from the cumulative effectsof combined services is also difficult.

In nonrandomized studies of child-care effects,results have been mixed; some studies report thathigher quality care is most beneficial for childrenin low-income families (e.g., Currie, 2001; Desai,Chase-Lansdale, & Michael, 1989; Lamb, 1998), butother studies report no variation in the effectiveness

of care across economic strata (e.g., Burchinal et al.,2000). When samples have been limited primarily tolow-income families, children in higher quality careachieve better than those in lower quality care (e.g.,Burchinal et al., 2000), although the achievementbenefits may be greater for some poor children (e.g.,those who come from higher quality home environ-ments) than others (Votruba-Drzal, Coley, & Chase-Lansdale, 2004). In addition, higher quality childcare has demonstrated apparent protective effectsfor mathematics achievement in middle childhood

1330 Dearing, McCartney, and Taylor

8/13/2019 Dearing Et Al 2009

3/21

for children exposed to social and contextual risksoften correlated with low income such as low par-ent education (Peisner-Feinberg et al., 2001). Indeed,when researchers have created risk indexes usinglow income and correlated risk factors, the esti-mated benefits of higher quality child care for mathachievement appear greatest for children facing thegreatest risk (e.g., Burchinal, Roberts, Zeisel, Hen-non, & Hooper, 2006). This is an important findinggiven that poverty is generally accompanied by ahost of other risks for childrens developmentbeyond low family income.

It is also worth noting two methodologicalapproaches that researchers have taken to answertwo closely related questions in the child-care andpoverty literatures. The first approach has been todetermine whether the association between qualityof child care and childrens achievement varies as a

function of family income (e.g., NICHD Early ChildCare Research Network, 2000, 2002). The secondapproach has been to determine whether the associ-ation between family income and child achieve-ment varies as a function of whether children arein higher quality child care (e.g., McCartney et al.,2007) or, similarly, to compare the developmentaloutcomes of poor children as a function of whetherthey are in higher quality child care (e.g., Campbellet al., 2001). Although these two approaches arevery similar, one key distinction between them isthat the first is limited by design to children inchild care and the second is not (i.e., the achieve-ment of children in maternal care may also beconsidered).

In the NICHD Study of Early Child Care andYouth Development (SECCYD), for example, only794 of 1,364 children (i.e., 58.2%) were included inanalyses examining variations in the effects of qual-ity across the income distribution because manychildren were in maternal care at 24 and 36 monthsof age (NICHD Early Child Care Research Net-work, 2002). Of those 570 children excluded fromthe analyses, over 40% were living in low-incomefamilies (i.e., family income less than 2 times the

federal poverty threshold). In the remaining sub-sample of 794 children and families, there was noevidence that the effects of child-care quality variedas a function of family socioeconomics. Because thisresearch team was interested primarily in familyrisk moderators of the association between child-care quality and child outcomes, including childrenin maternal care would not have been appropriate.However, for researchers interested in developmen-tal contexts that best promote the development ofchildren in low-income families, external validity is

limited substantially if children in maternal care arenot considered.

Another value of including children in higherquality care, lower quality care, and children whoare not in child care is the ability to disentanglewhether higher quality child care promotes thedevelopment of children in low-income families,whether lower quality child care poses a dual riskfor children in low-income families (i.e., exacerbat-ing the risk of low income), or whether both are true.If children in maternal care are excluded, it is notclear which of these three possibilities offers the bestexplanation for achievement differences betweenlow-income children in higher versus lower qualitycare. In short, the effect of income on children inmaternal care provides a useful comparison for theeffects of income on children in child care.

Using this strategy with data from the NICHD

SECCYD, McCartney et al. (2007) report evidenceconsistent with the hypothesis that higher qualitychild care protects low-income children, with noevidence that lower quality child care posed a dualrisk. More specifically, at 36 months of age, low-income children in higher quality child care dis-played higher cognitive and language achievementthan those in maternal care; on the other hand,low-income children in lower quality child careperformed better or at levels indistinguishable fromthose in maternal care. Note, however, that theseresults were limited to achievement during earlychildhood. Knowing whether they extend into mid-dle childhood would add to considerations of thedevelopmental value of higher quality child carefor low-income children.

Importantly, however, all nonrandomized stud-ies of child care are potentially limited by selectionbias. In these studies, families of children in higherquality care have chosen this context. Without ran-dom assignment, therefore, there is the concern thathigher quality child-care effects may not be dueto higher quality child care per se but instead dueto unmeasured characteristics of children and fami-lies that are associated with the propensity to use

higher quality care. With this in mind, we useddata from a nonrandomized longitudinal study toexamine variations in the effects of family econom-ics on math and reading achievement as a functionof higher quality early child-care dosage, takingspecial care to control for potential selection effects.

The Present Study

In the present study, we examined whetherhigher quality (i.e., above average) child care

Child Care and Achievement 1331

8/13/2019 Dearing Et Al 2009

4/21

8/13/2019 Dearing Et Al 2009

5/21

associations between family economic status andmiddle-childhood achievement were evident when:(a) adjusting for potential selection bias usingpropensity scores and (b) considering low familyincome in combination with other correlated socio-contextual risk factors.

Method

The present study was based on secondary analysesof the NICHD SECCYD, a prospective longitudinalstudy that was originally designed to examine thedevelopmental implications of early child care. Thedata used in our investigation were from the first,second, and third phases of the NICHD SECCYD,which covered childrens lives from birth throughfifth grade, including data on: (a) child-care quality

during infancy and early childhood, (b) family eco-nomics across the first 10 years of childrens lives,(c) early school readiness skills, and (d) child mathand reading achievement during middle childhood.To address potential selection effects, we also useda large variety of child, parent, and family variablesthat were collected as part of the SECCYD.

Sample and Study Phases

Shortly after giving birth in 1991, 1,364 womenand their recently born children living in or near 10urban and suburban sites in the United States wererecruited to participate in the SECCYD using a con-ditional random sampling method (for samplingdetails, see NICHD Early Child Care Research Net-work & Duncan, 2003). The sample is not statisti-cally representative of any population defineda priori, and families were excluded from thestudy if children had a disability or mothers were< 18 years old, were not fluent in English, andorlived in a very dangerous neighborhood. Nonethe-less, the sample is economically and geographicallydiverse.

The first phase of the study included assess-

ments of children and their developmental contextsat four ages: 6, 15, 24, and 36 months. Telephoneinterviews were used to update the status of child-care arrangements at 3- and 6-month intervalsfollowing the 6-, 15-, and 24-month assessments.The second phase of the study included assess-ments of these same children at 54 months, kinder-garten, and first grade as well as telephoneinterviews when children were 42, 46, 50, 60, and66 months of age. The third phase included assess-ments at third and fifth grades.

Measures

Child, maternal, and family characteristics. Whenchildren were 1 month of age, their mothersreported on several demographic characteristics,including child gender, ethnicity, and birth order.At this time, mothers also reported on their ageand level of education. At 10 time points duringthe study, mothers also reported on their partnerstatus and household size; we formed summaryindicators from these two variables using the pro-portion of assessment points at which motherswere married or partnered and the mean house-hold size across the study.

Data on maternal attitudes, behaviors, intelli-gence, and personality as well as the quality ofthe home environment were also collected. At1 month, mothers completed the Beliefs About the

Consequences of Maternal Employment for Chil-dren scale (Greenberger, Goldberg, Crawford, &Granger, 1988) and the Parental Modernity Scale(Schaefer & Edgerton, 1985), providing measuresof their beliefs about the potential benefits andcosts of maternal employment and childrearing. At6 months, maternal separation anxiety was assessedusing the Maternal Separation Anxiety Scale (Hock,Gnezda, & MacBride, 1983) and mothers completedthree subscales from the NEO Personality Inventory(Costa & McCrae, 1985) to assess extraversion,agreeableness, and neuroticism.

At 6 months, observations were also completedof maternal sensitivity to child distress and nondis-tress (NICHD Early Child Care Research Network,2007) and the quality of the home environment(Home Observation Measure of the Environment[HOME]; Caldwell & Bradley, 1984). From this lat-ter measure, we used the total quality score thatincluded the sum of subscale scores in the areas of:parental responsiveness toward the child, parentalacceptance of the child, organization of the environ-ment, presence of learning materials, parentalinvolvement, and the variety of experiences pro-vided to the child. In addition, at 36 months, moth-

ers completed the Peabody Picture VocabularyTestRevised (Dunn & Dunn, 1981), an indicator ofverbal intelligence. Descriptive statistics are dis-played in Table 1 for all child, maternal, and familycharacteristics.

Income-to-needs. At nine time points (i.e., 6, 15,24, 36, and 54 months as well as at kindergarten,first grade, third grade, and fifth grade), familiesreported on their annual household income fromall sources. From these data, an income-to-needsratio was computed at each time point, defined as


8/13/2019 Dearing Et Al 2009

6/21

family income divided by the poverty threshold forthe appropriate family size, as established by theU.S. Census Bureau. Note that an income-to-needsratio of 1.0 denotes the poverty level, 2.0 is oftenconsidered the threshold for near poverty or lowincome, and approximately 3.0 denotes middle-class status. For our analyses, we used familiesaverage income-to-needs across the study. Byexamining income-to-needs as continuous ratherthan as categorical (e.g., poor vs. not poor), wewere able to define more precisely the levels offamily economic resources across which higherquality care might promote achievement.

Descriptive statistics for average income-to-needs

are displayed in Table 1. Note, however, thatincome-to-needs levels were centered on the sam-ple mean (i.e., 3.68 was subtracted from each valueso that the mean of the new centered income-to-needs variable was equal to 0) for analyses, becauseit was included in statistical interaction terms (for adiscussion of the advantages to centering predictorsin interactions, see Dearing & Hamilton, 2006).

Child care. For children in nonmaternal care for10 or more hr per week, child-care quality wasassessed during two half-day visits to childrens

primary child-care setting at 6, 15, 24, 36, and54 months using the Observational Record of theCaregiving Environment (ORCE), a live observa-tional instrument designed for the SECCYD(NICHD Early Child Care Research Network, 1996,2000). Study children were observed for a total offour 44-min observation cycles at each age. At 6, 15,and 24 months, sensitivity to childs nondistressexpressions, positive regard, stimulation of cogni-tive development, detachment, and flat affect wereassessed using qualitative ratings. At 36 months,two additional categories were added: fosteringexploration and intrusiveness. At 54 months, rat-ings were focused on sensitivity and responsivity,stimulation of cognitive development, intrusive-ness, and detachment. A composite variable ofchild-care quality was formed by summing thequalitative ratings across domains.

Internal consistencies for the composite qualityindicators were above .80 at each age. Means andstandard deviations for these composites are dis-played in Table 2. Also presented in Table 2 are thepercentages of children who were in these samearrangements 6 months later, as indicated in tele-phone interviews. Median time in child-care settingranged from 7 months (at 12 and 42 months) to14 months (at 30 and 60 months).

For the present study, a dummy variable repre-senting higher quality child care was created foreach time point at which child-care quality wasobserved. Higher quality care was defined such thatchildren at or above the median (at that time point)on the quality composite were coded as 1, whereaschildren below the median as well as children innonmaternal care for < 10 hr per week were codedas 0. The resulting five dummy variables were thensummed across the five assessments to create anindicator of the number of episodes that childrenwere in higher quality care, with scores rangingfrom 0 to 5. We created dummy variables for lowerquality care (i.e., below the median at each timepoint) in a similar fashion; number of episodes inlower quality care was then computed by summing

across these dummy variables. The mean numberof episodes in higher quality child care was 1.58(SD = 1.38) and the mean number of episodes inlower quality child care was 1.31 (SD= 1.35).

Note that using the median as a quality cut-pointresulted in groupings that were theoretically mean-ingful. On the 4-point ORCE qualitative ratings,scores of 3.0 or higher (and 2.0 or lower on reversescored items) indicated care that was developmen-tally stimulating and supportive. At each timepoint, children with overall quality scores at or

Table 1

Descriptive Statistics for Child Characteristics, Maternal Characteris-

tics, and Income-to-Needs

Variables M(SD)% Missing (%)

Child characteristics

Gender = boy 51.69% 0

African American 12.9% 0

European American 80.43% 0

Latino American 6.09% 0

Birth order 1.83 (0.95) 0

Maternal characteristics

Age 28.11 (5.63) 0

Years of education 14.23 (2.51) < 1.0

Average partner status 0.69 (0.36) < 1.0

Traditional childrearing values 60.34 (15.21) < 1.0

Sensitivity 9.21 (1.78) 6.8

Separation anxiety 66.42 (13.84) 6.4

Neuroticism 29.77 (7.16) 6.7

Agreeableness 46.28 (5.29) 6.7Extraversion 42.49 (5.83) 6.7

Employment attitudes 0.85 (7.11) < 1.0

Intelligence (PPVT) 99.01 (18.35) 14.4

Household characteristics

Average household size 4.22 (1.01) 3.8

6-month HOME 36.55 (4.65) 6.2

Average income-to-needs 3.68 (2.98) 4.0

Note. PPVT = Peabody Picture Vocabulary Test; HOME = HomeObservation Measure of the Environment.


8/13/2019 Dearing Et Al 2009

7/21

above the median had scores of 3.0 or higher on themajority, if not all, of the individual items.

Sociocontextual risk. We created an index of so-ciocontextual risk from income-to-needs and corre-lated maternal and family risk characteristics.Because our purpose in computing this risk indexwas to examine low family income in combinationwith correlated risk factors, we first examinedintercorrelations between the child, maternal, andfamily characteristics (listed in Table 1) and familyincome-to-needs. Six variables were strongly associ-ated with income-to-needs (i.e., r > .40): maternalage, education, verbal intelligence, sensitivity, andtraditional parenting values (negatively correlated)as well as the 6th-month HOME score.

We then dichotomized each of these six indica-tors and family income-to-needs, with values of 1on each variable indicating high risk. For family

income-to-needs, we classified children who werelow income or poorer (i.e., income-to-needs < 2.0;29.7% of sample) as high risk. For maternal age, weused < 20 as the high-risk cut-point (7.0% of samplewas younger than 20). For maternal education, weused less than high school as the cut-point (10.2%of sample). For verbal intelligence, sensitivity, andthe HOME, participants with scores in the lowestthird of the sample were classified as high risk. Fortraditional parenting values, participants withscores in the highest third of the sample were clas-sified as high risk.

Next, we summed the dichotomized variables sothat risk index scores ranged from values of 0 to 7,with 7 indicating the highest total number of risks.However, < 6% of the sample had six or seven risksand a cross-tabulation of the risk index by episodesin higher quality child care revealed many emptycells (e.g., no children with seven risks were inmore than three episodes of higher quality childcare and no children in five episodes of higherquality care had more than four risks). Thus, welimited the risk index to four levels with similar cellsizes: 0 risks (approximately 31% of the sample), 1risk (approximately 23% of the sample), 23 risks(approximately 26% of the sample), and 4 or morerisks (approximately 20% of the sample).

School readiness skills. Childrens cognitive knowl-edge and skills were assessed by trained researchassistants when children were 36 months of age using

the School Readiness composite from the BrackenBasic Concept Scale (Bracken, 1984). This 51-itemcomposite assesses childrens abilities in areas suchas letter identification, number and counting skills,comparisons, and color and shape recognition. Themeasure has demonstrated excellent validity viacorrelations with intelligence measures and academicperformance in kindergarten (Laughlin, 1995); it hasalso demonstrated good split-half and testretestreliability (Bracken, 1984; Bracken, Howell, Harrison,Stanford, & Zahn, 1991). In the SECCYD, the inter-nal consistency was excellent (a = .93).

Table 2

Descriptive Statistics for Child-Care Quality

Quality score

M(SD)

In same arrangement

6 months later,a %

Missing

(%)

Percentage of children in higher quality careb

Low income

(n = 405)

Middle income

(n = 501)

High income

(n= 458)

6 months 14.72 (2.86) 64.5 5.6 18.3 29.2 34.2

15 months 14.48 (2.89) 68.0 7.0 18.3 29.5 42.2

24 months 13.85 (2.79) 69.6 9.3 18.1 32.1 42.3

36 months 19.38 (3.17) 60.5 9.9 20.0 34.2 41.7

54 months 11.80 (2.13) 79.2 16.8 32.2 35.1 46.8

Total episodes inhigher quality care

Low income(n = 405)

Middle income(n = 501)

High income(n= 458)

0 episodesc 36.2 25.5 18.2

1 episode 34.1 27.6 22.8

2 episodes 18.5 21.7 20.9

3+ episodes 11.2 25.2 38.1

aIn this column, we present the percentage of children who were still in the same child-care arrangement as reported in phoneinterviews 6 months later (i.e., at 12, 21, 30, 42, and 60 months, respectively). bThe percentages and sample sizes in these three columnswere based on averages across the five complete data sets after using multiple imputation to replace missing values. Quality scoresabove the sample median were classified as higher quality. cThe percentages of children in maternal care only (i.e., no higher orlower quality child care) were 12.3% for low income, 11.8% for middle income, and 9.0% for high income.


8/13/2019 Dearing Et Al 2009

8/21

Middle-childhood achievement and cognitive abil-ity. In the SECCYD, a variety of achievement andcognitive ability scales from the WoodcockJohnsonPsycho-educational BatteryRevised (WJR; Wood-cock & Johnson, 1989) were administered to chil-dren. The reliability and validity of the WJR hasbeen well documented. At third and fifth grades,broad mathematics and broad reading scores wereobtained as general measures of achievement inthese domains. On the WJR, broad mathematicsscores are computed from two achievement scales:calculations, a scale used to access childrens abilityto add, subtract, multiply, and divide, and appliedproblems, which requires children to solve practicalmathematical problems. Broad reading scores arecomputed from two scales: letterword identifica-tion, which requires children to match pictographicrepresentations of words with pictures of objects,

and passage completion, which requires childrento select the most appropriate word to complete apassage of text.

Four scales from the WJR were assessed threeor more times, allowing patterns of growth to beexamined. Two of the achievement tests includedin the broad math and reading scores (i.e., appliedproblems and letterword identification) were alsoassessed at 54 months and fifth grade, providingfour total assessments. In addition, two cognitiveability scales were assessed at three and four timepoints, respectively. Memory for sentences, a mea-sure of childrens ability to remember and repeatwords, phrases, and sentences, was assessed at54 months, first grade, and third grade. Picturevocabulary, a measure of childrens ability to recog-nize or name objects in pictures, was assessed at54 months, first grade, third grade, and fifth grade.Descriptive statistics for these scales are displayedin Table 3.

Multiple Imputation (MI) for Missing Values

Of the SECCYD participants, approximately 75%had complete data on all predictors and at least

one achievement assessment during middle child-hood. Moreover, we found very little evidence thatrate of missing data varied by family economicstatus. On the fifth grade achievement outcomesand child-care quality variables, for example, allcorrelations between missing versus complete dataand income-to-needs were smaller than r= .08.Nonetheless MI is generally a preferred method forhandling missing data compared with discardingparticipants or observations (Schafer & Graham,2002; Widaman, 2006). Thus, we conducted our

analyses using MI to include all 1,364 children andfamilies.

Specifically, we used MI by chained equations(i.e., MICE; Royston, 2004), computing five completedata sets that combined observed and imputedvalues. We then used MI estimation techniques(e.g., mifit in STATA) to run our analyses and com-bine estimates from these five data sets according toRubins rules (Rubin, 1987; Schafer & Graham,2002). Consistent with the SECCYD protocol, qualityscores were imputed only if nonmissing or imputedhours data indicated that children were in care 10 ormore hr per week. Based on averages across thesefive complete data sets, the percentages of childrenin higher quality child care at each assessment pointand across the study are presented in Table 2 forfamilies with: (a) income-to-needs 2.00 (i.e., lowincome), (b) income-to-needs between 2.01 and 4.00

(i.e., middle income), and (c) income-to-needs > 4.00(i.e., high income). Family income-to-needs wasmoderately correlated (average r = .27 across thefive data sets) with number of episodes in higherquality care but was not associated with the numberof episodes in lower quality care (average r= ).01across the five data sets).

Addressing Potential Selection Bias Using Covariatesand Generalized Propensity Scores

A concern for the present study is that selectionprocesses rather than child-care quality per se maymoderate the effects of family income on childachievement. Following McCartney, Bub, and Bur-chinals (2006) advice for addressing concerns overselection, we used three different methodologicalapproaches to study questions and assessed therobustness of findings across these methods. First,we included a large set of child and parent charac-teristics as statistical covariates in our regressionand multilevel models. The covariates that weincluded (listed in Table 1) have proven to be asso-ciated with selection into higher quality child carein past research (e.g., NICHD Early Child Care

Research Network & Duncan, 2003), although eventhe most thorough set of covariates leaves open thepotential for omitted variable bias (Duncan, Mag-nuson, & Ludwig, 2004).

As a second approach, we used propensityscores to help control for potential selection bias(Imbens, 2000; Rosenbaum & Rubin, 1984). Propen-sity score matching is a three-step process that canreduce as much as 90% of selection bias in non-experimental data (Leon & Hedeker, 2006). In thefirst step, the treatment of interest is regressed


8/13/2019 Dearing Et Al 2009

9/21

on a set of covariates thought to be associated withselection; generally, logistic regression is used for

this first step. Next, estimated probabilities forreceiving the treatment, taken from the logisticregression analysis in the first step, are used tomatch participants who did receive the treatmentwith those who did not receive the treatment.Finally, in the third step, the effects of the treatmentare then estimated for these matched participants.

Although propensity score matching has gener-ally been reserved for studies comparing one treat-ment group with one comparison group, Imbens(2000) demonstrated that this method can beapplied to studies of ordered doseresponse datausing a weighting scheme. To form the weights,dosage levels for the treatment of interest areregressed on a set of covariates using ordinal (ormultinomial) logistic regression. Then, the weightsare computed using the inverse of the predictedprobabilities corresponding to the dosage of treat-ment that each participant received. Imbens refersto these weights as generalized propensity scores.To compute and use generalized propensity scoresin the present study, we first regressed the numberof episodes in which children were in higher qual-ity child care on the child and maternal characteris-tics displayed in Table 1. Next, we used the

resulting generalized propensity scores as weightsin our analyses of the moderating effects of higherquality child care.

As a third methodological approach for address-ing potential selection bias, we used propensityscores as variables in our analyses. Specifically, weincluded both the main effect of childrens propen-sities to receive a high dosage of higher qualitychild care and the interaction of these propensityscores with family income-to-needs. Thus, we wereable to examine whether the propensity to receive a

high dosage of higher quality child care moderatedthe effects of income-to-needs on child achievement

and, most importantly, whether higher qualitychild care per se moderated income above andbeyond the potential moderating effects of selectionpropensity. We present results for the moderatingeffects of selection based on childrens propensitiesto be in five episodes of higher quality child care,although we also used a variety of other specifica-tion strategies (e.g., propensity to be in more thanone episode or more than three episodes) to exam-ine the potential moderating effects of selection.Across all strategies, the moderating effects ofhigher quality care were substantively identical.

It is important to note that propensity scores ulti-mately rely on measured covariates, and thismethod is not a substitute for randomization.Although randomization can be used to balanceparticipants on both observed and unobserved vari-ables, propensity scores are a means of balancingparticipants on observed variables. The longitudinalSECCYD design, however, allowed us to balanceparticipants on characteristics that were observedprior to collecting the child outcome data, animportant advantage when using propensity scores(Rubin, 2007). Moreover, using multiple approachesto control for potential selection bias provided an

indication of the robustness of our results beyondrelying solely on traditional covariate methods.

Results

Moderating Effects of Higher Quality Child Care forMath and Reading Achievement

As a first step in our analyses, we regressed chil-drens math and reading achievement scores on

Table 3

Descriptive Statistics for WoodcockJohnson Subscales

Assessment time

M(SD) Missing (%) M(SD) Missing (%) M(SD) Missing (%)

Broad reading Broad math Applied problems

54 months 424.72 (19.27) 22.9First grade 470.05 (15.54) 25.0

Third grade 494.60 (15.75) 25.9 493.50 (12.76) 25.8 497.33 (13.19) 25.8

Fifth grade 507.62 (14.12) 27.2 510.53 (13.04) 27.2 509.83 (12.85) 27.2

Letterword identification Memory for sentences Picture vocabulary

54 months 369.36 (21.41) 22.6 457.15 (18.29) 22.7 459.54 (14.09) 22.3

First grade 452.59 (23.99) 24.8 481.32 (14.76) 24.8 483.94 (12.27) 25.2

Third grade 493.86 (18.73) 25.7 494.85 (14.48) 25.7 496.94 (11.51) 25.7

Fifth grade 510.12 (17.52) 27.2 505.85 (12.08) 27.3


8/13/2019 Dearing Et Al 2009

10/21

family income-to-needs, number of episodes inhigher quality child care, number of episodes inlower quality child care, and the interaction of thetwo child-care quality variables with income-to-needs. To help control for potential child-careselection effects, we also included a variety ofdemographic, attitude, and psychological character-istics of children, mothers, and families as covari-ates, namely: (a) child gender, ethnicity, and birthorder; (b) maternal age, years of education, averagepartner status, childrearing values, sensitivity,separation anxiety, personality (i.e., neuroticism,extraversion, and agreeableness), attitudes towardemployment, and verbal intelligence; and (c) familyhousehold size and quality of the home environ-ment.

For the broad math and reading outcomes, weestimated random intercept models, allowing us to

examine average math and reading achievementacross third and fifth grade as well as variations bygrade in the estimated associations. For the fourindividual achievement scales that were assessed atthree or more time points between 54 months andfifth grade (i.e., applied problems, letterword iden-tification, memory for sentences, and picture vocab-ulary), we estimated multilevel growth modelswith random intercepts and random slopes. For theapplied problems, letterword identification, andpicture vocabulary scales, unconditional growthmodels indicated significant (p< .01) linear andquadratic change over time. For the memory forsentences scale, there was significant (p < .01) linearchange over time.

In the multilevel growth models, we centeredlinear and quadratic achievement slopes on theirgrand means so that intercepts in the first levels ofthe models provided estimates of average achieve-ment across middle childhood. Thus, for appliedproblems, letterword identification, and picturevocabulary, predictors of interest and covariateswere specified for three growth parameters: (a)average achievement between 54 months and fifthgrade, (b) linear change in achievement between

54 months and fifth grade, and (c) quadratic changein achievement between 54 months and fifth grade.For memory for sentences, a similar model wasestimated with the exception that average achieve-ment was estimated between 54 months and thirdgrade, and there was no quadratic growth parame-ter in the model. A summary of the results fromthese models is presented in Table 4. For therandom intercept models, we present only theresults for average broad math and readingachievement in this table, because none of the Ta

ble4

TheModeratingEffectsofHigherQua

lityChildCareandLowerQualityChildCareforAssociationsBetweenFamilyIncome-to-NeedsandAchievementDuringMiddleChildhood

Broadmath:

G35

Broad

reading:G35

Appliedproblems:54m

G5

Letterwordidentification:

54mG5

Memoryfor

sentences:54mG3

Picturevo

cabulary:54mG5

Intercept

Age

Age

2

Intercept

Age

Age

2

Intercept

Age

Intercept

Age

Age

2

Income-to-needs

.89**(.29)

.80**(.31)

1.09***(.29)

).22(.16)

.02(.02)

1.41***(.38)

.26(.30)

).06

(.04)

.64

(.35)

).05(.13)

.51*(.23)

).09(.14)

.01(.02)

Higherquality

episodes

.42

(.29)

.13

(.31)

.39

(.32)

.11(.18)

).01(.03)

.32

(.37)

.03(.22)

).01

(.03)

.22

(.36)

).10(.09)

.22

(.22)

).21(.15)

.03(.02)

Lowerquality

episodes

).15

(.27)

).50

(.34)

).26

(.26)

.16(.15)

).01(.02)

).66

(.36)

.29(.21)

).03

(.03)

).38

(.27)

).10(.09)

).46*(.21)

).10(.11)

.02(.02)

Income-to-Needs

HigherQuality

Episodes

).19*(.08)

[).18*]

).18*(.09)

[).14

]

).20*

(.09)

[).17*]

.05(.05)

[.03]

).01(.01)

[.00]

).29*

(.12)

[).24*]

).16(.09)

[).13]

.03

(.01)

[.02]

).06

(.10)

[).08]

.01(.04)

[.02]

).12

(.06)

[).10

]

).01(.05)

[).01]

.00(.01)

[.00]

Income-to-Needs

LowerQuality

Episodes

).04

(.10)

).06

(.11)

).07

(.08)

.05(.06)

).01(.01)

).15

(.11)

).10(.07)

.02

(.01)

.06

(.10)

).04(.03)

).07

(.07)

).01(.04)

.00(.01)

Note.Coefficientsforinterceptsare

associationsbetweenpredictorsandaver

ageachievementacrosstime.Coefficients

forAgeandAge

2

areassociationswithlinearandquadratic

changesinachievement.Standarderrorsareinparentheses.Inbracketsarehigherqualityinteractioncoefficientsfrom

modelsinwhichtheinteractionforlowe

rqualitychildcare

wasexcluded.Allmodelsincludedcovariatesfor:childbirthorder,ethnic

ity,andgender;maternalage,education

,partnerstatus,childrearingvalues,employmentattitudes,

sensitivity,separationanxiety,neuroticism,extraversion,agreeableness,and

verbalintelligence;andhouseholdsizean

dHomeObservationoftheEnvironment

score.

p

dearing et al 2009

Documents