Sex Differences in Early Childhood Education Intervention Impacts on Cognitive Outcomes

Abstract

Several influential studies reported sex differences in early care and education (ECE) treatment on young adult IQ and academic outcomes. This paper extends that work by asking whether sex differences in impacts of the Carolina Abecedarian Project emerged during the treatment period or subsequently and whether sex differences were maintained into middle adulthood. The randomized clinical trial (98% Black, 51% female) followed 104 infants 5 to 45 years of age. Longitudinal analyses estimated treatment and sex-by-treatment differences at 5 years, from 5 to 21 years, and at 21 and 45 years. Results revealed treated children entered school with higher IQ and reading skills than control children. Treatment impacts on IQ and math increased over time for females and decreased for males yielding sex differences in treatment impacts at 21 and 45 years. These findings suggest that, while the ECE treatment similarly benefited boys and girls in the short term, the long-term impacts likely depended on subsequent experiences.

Considerable evidence suggests that early childhood is a critical period for development and that high-quality early care and education (ECE) can change developmental trajectories – especially for children experiencing early life stressors such as poverty and systemic inequities (National Academy of Science, Engineering, and Medicine [NASEM], 2019). ECE programs implemented over 50 years ago show improved adult educational, employment, and health outcomes (Campbell et al., 2012; Garcia et al., 2018), with larger impacts on educational outcomes for females (Garcia et al., 2018). Despite documenting important gender differences in treatment impacts, these studies did not examine whether these sex differences emerged during the treatment period (i.e., early childhood) or after children entered school and whether the sex differences in impacts are maintained into middle adulthood. Identifying when the sex differences emerge and whether they persist has important policy implications. Such differences likely reflect differential responses to treatment if they emerge during the treatment period, but sex differences could reflect differences in how society in general and schools in specific support the male and female participants if differences emerge after children left ECE for kindergarten. Efforts to develop ECE interventions that maintain their impacts for both sexes over time would be different depending on when those sex differences in impacts emerge. The purpose of this paper is to ask three questions of a seminal ECE intervention study about sex differences in treatment impacts on IQ and academic skills in the Carolina Abecedarian Project: 1) do impacts differ by sex at the end of the treatment period; 2) do rates of change in treatment impacts differ by sex during the school years and early adulthood?; and 3) are treatment impacts on IQ maintained to 45 years of age and do they differ by sex?

Theoretical Background - Early Education as a Protective Factor

The broad socio-ecological and transactional frameworks describe how children’s own characteristics, their interactions with caregivers, and their environment shape development (Bronfenbrenner & Morris, 2006; Garcia Coll et al., 1996; Sameroff, 2009). According to these models, the interactions young children experience with their parents and ECE caregivers promote the emotional, social, and cognitive skills that support subsequent development. That is, according to these models, young children thrive when caregivers and parents provide frequent warm, responsive interactions and enriching materials and experiences.

The interactions between young children and their caregivers are especially important in early childhood because rapid brain development during this time lays the groundwork for later cognitive and behavioral skills (Shonkoff & Phillips, 2000). Given this rapid development, children are particularly vulnerable to stressors stemming from poverty and structural racism during the early years (NASEM, 2019). However, interventions that occur during this sensitive period of development are believed to buffer children from some of these adversities, setting off a positive developmental “cascade” in which the improved skills then lead to higher skills as development advances (Masten & Cicchetti, 2010). Economic theories of building human capital (Cunha & Heckman, 2007; Heckman, 2007) suggest a similar concept that “skill begets skill,” or early childhood programs lay the foundation on which children can continue to build skills during formal schooling. The positive impacts that ECE programs have on adult outcomes (e.g., Campbell et al., 2012; Deming, 2009; Gray-Lobe et al., 2023; Heckman et al., 2010; Lazar et al., 1982) provide further evidence of the importance of early life experiences on the development of the brain and, thus, on the cognitive and social development of the child (Shonkoff & Phillips, 2000). Two of the most successful ECE programs, Perry Preschool Project (PPP) and this study, the Abecedarian project (ABC), were developed based on the social-ecological models – focusing on frequent sensitive and stimulating interactions between the teachers and children and providing children with scaffolded hands-on opportunities to learn (Ramey et al., 2012; Weikart, 2004). While both projects documented treatment impacts on IQ at the end of the treatment period, neither examined academic skill at the end of the treatment period. To date, only PPP has demonstrated impacts into middle age (Garcia et al., 2018). For these reasons, this study of the ABC project tests for treatment differences on IQ, reading, and math skills at entry to kindergarten and on IQ in middle age.

Early ECE programs from the 1960s and 70s often recruited families and children experiencing poverty, many of whom were Black, into their intervention studies. Thus, it is important when examining early studies to also consider the context and unique experiences of Black children growing up in the United States during a time of racial and civic unrest. During the time of the early ECE studies, many Black communities were seeking to dismantle legacies of oppression like Jim Crow laws and segregation. García Coll and colleagues’ (1996) model of the development of minority children in the United States emphasizes that the developmental pathways of minority children are impacted by their “unique ecological circumstances,” including the interplay between social position, oppression, and segregation due to systemic racism. They also argue that the educational context is particularly important for shaping the development of racially minoritized children. Education can be viewed both as a contributor to development and as an outcome. Racially minoritized children often do not have the same educational opportunities as their White peers and face racism from educational institutions, teachers, and peers in many forms, such as fewer resources, lower expectations, and biased curriculum (Gardner-Neblett et al., 2021). However, ECE programs that are safe, racially affirming, and enriching are likely to be a protective factor in the participants’ lives even as they continue to face racism and discrimination in other facets of their lives (James & Iruka, 2018; Osborne et al., 2021; Shonkoff et al., 2021).

Furthermore, multiple aspects of children’s identities influence their experiences in schools. Research suggests that sex can play an important role in the way children are perceived and treated (Jensen & Tisak, 2020; Shivers et al., 2021). The interrelated nature of race and sex can increase the likelihood that children will experience different types of bias and shape the way that bias is experienced (APA, 2012). For Black children, sex can play a role in the resources they are able to access, the types of social interactions they experience, and the expectations others have for them (Bryan, 2020; García Coll et al., 1996). Thus, a second aim of this paper is to consider the role of ECE as a protective factor for racially minoritized children and the moderating role of sex.

Early Education Intervention Projects with Adult Outcomes

Long-term follow-up studies have been conducted for several early ECE programs. Perhaps the two most influential early ECE programs are ABC and PPP, which primarily served Black children from poor households and communities. Both programs involved highly qualified teachers who provided child-centered activities, strong teacher-child relationships, and frequent conversations between children and teachers (Ramey et al., 2012; Weikart, 2004). Both PPP and ABC reported treatment impacts on cognitive skills during the treatment period, on academic skills during the school years, and on education, income, and health risk factors during adulthood (Campbell et al., 2014; Garcia et al., 2020; Heckman et al., 2010). The PPP impacts varied by sex with large ECE impacts on educational outcomes for females and on income for males (Heckman et al, 2010), but ABC impacts were not reliably different by sex in cross-sectional tests with limited power (Campbell et al., 2002). These programs are estimated to generate rates of return¹, the rate of benefits of the program to society, over and above program costs of 13.7% for ABC (García et al., 2018) and 7 to 10% for PPP (Heckman et al., 2010). The impact of these studies has been touted by advocates to justify the continued and expanded funding of ECE programs such as Head Start, pre-kindergarten programs, and quality rating and improvement systems (for review see Burchinal et al., 2015; Shonkoff & Phillips, 2000).

Three other long-term follow-up evaluations of preschool ECE programs provide additional evidence of long-term impacts, with some sex differences found in two studies. Lazar and colleagues (1982) examined long-term impacts of 11 ECE programs conducted during the 1960s and 1970s, finding attendees were less likely than nonattenders to be retained or referred to special education and more likely to graduate from high school, but did not find program impacts varied by sex. Deming (2009) compared siblings in the National Longitudinal Survey of Youth who did and did not attend Head Start, finding that Head Start attendees were more likely to graduate from high school and have higher scores on an index of positive adult outcomes. He reported differences in Head Start impacts by sex for some outcomes, with a larger impact on college attendance for females and a larger impact on employment for males (Deming, 2009). Gray-Lobe and colleagues (2023) examined adult outcomes for children who were and were not selected to attend the Boston Pre-Kindergarten Program (BPKP), finding that compared to non-attendees the attendees were more likely to take and score higher on the SAT, graduate from high school, and enroll in any or a 4-year college, and attendees were less likely to receive school-based disciplinary actions. They also reported sex differences, with larger BPKP impacts for males than females on fewer disciplinary actions and higher levels of high school graduation, 4-year college enrollment, and college graduation.

Two studies asked whether there were sex differences in ECE impacts across studies. Anderson (2008) examined ABC, PPP, and the Early Training Project to test whether these ECE programs had impacts separately by sex for IQ, educational status, employment, and criminal justice assessments during the preteen years (5–12 years), teen years (13–19 years), and adult years (20–27 years). He designed the analyses to address concerns about small sample size and multiple inference in the studies’ many publications. Anderson limited the number of outcomes examined, employed statistical methods not dependent on asymptotic properties that are questionable with small sample sizes, and accounted for testing program effectiveness across multiple outcomes. He reported that these three programs had a significant impact on educational outcomes and that intensive intervention early in life may positively affect later-life outcomes, at least for Black females from underserved communities, but perhaps not for Black males.

Magnuson and colleagues (2016) conducted a meta-analysis of preschool ECE programs that focused on normally developing children, employed rigorous designs, and had at least 10 children in both the treatment and comparison groups. They examined 23 ECE programs such as Head Start, summer Head Start, state/local pre-kindergarten, Bright Beginnings, Montessori, and Early Reading First. Their meta-analyses indicated that ECE programs are more effective for females when considering cognitive and academic, behavior, and mental health outcomes, but are more effective for males for other school outcomes such as grade retention and special education (Magnuson et al., 2016). They replicated Anderson’s results and extended them with follow-up analyses suggesting that Anderson’s conclusions might be limited to higher quality programs or programs implemented before 1976.

In summary, considerable evidence supports the existence of sex differences in ECE impacts but further careful examination within highly influential studies might be warranted. Anderson (2008) reported differential treatment impacts by sex. However, analyses did not include adult IQ or any academic achievement outcomes and did not address whether sex differences in treatment impacts emerged during or after the ECE treatment. Magnuson and colleagues (2016) conducted careful meta-analyses across studies, replicating Anderson’s findings for the three studies he examined. As a meta-analysis, data were not analyzed separately by study. Furthermore, both Anderson and Magnuson assumed that sex differences in treatment impacts were due to differences in response to the ECE program, not to subsequent experiences. Therefore, these studies did not address the question of whether ABC finds similar sex differences on adult outcomes as reported by PPP and, if so, whether those differences were detected at the end of the ECE program or emerged over time.

Research Questions

Through conducting longitudinal analyses of IQ, reading, and math skills, we asked three questions regarding comparing treated and control ABC participants: 1) how large were treatment impacts and did they vary by sex at entry to kindergarten?; 2) how did treatment impacts vary between the end of treatment (5 years-of-age) and early adulthood (21 years-of-age) and did change in those impacts vary by sex?; and 3) how large were treatment impacts on IQ in middle adulthood and did they vary by sex?

Methods

Participants

The ABC Project was an intensive infant and preschool program serving four cohorts of predominantly African American (98%) children from the local community between 1972 and 1983 (Ramey et al., 1979). Parents were recruited shortly before or after giving birth if they had risk factors such as being single or teen parents or having low levels of education or income (Ramey et al., 1977). The parents were young, with an average age of 20 years (SD = 5.9) among the mothers and 23 years (SD=5.9) among the fathers. Many had less than a high school education, with an average of 10^th grade education (SD = 1.8) among the mothers and 10^th grade education (SD =1.7) among the fathers. Most parents were not married (80%), and often the parents were not living together (72%) when the child was born. About half of the children were male. The 111 children were randomly assigned to treatment or control groups, with no significant differences between the groups on these demographic characteristics. Attrition included 1 treatment child due to a medical diagnosis as a preschooler, 1 control child whose parents withdrew as a preschooler, and 7 deaths (early childhood: 1 treatment, 3 control; young adulthood: 2 treatment, 1 control). Failure to follow 10 children into school for the first three years of school led to lower sample sizes for those years, but these children were included in subsequent follow-up studies. Participation rates for follow-up, relative to the original enrolled study sample (N = 111) were 84% at ages 6–8, 92% at ages 12 and 15, 94% at age 21, 92% at age 30, and 87% at age 45. Further, participation rates are even higher when computed relative to the eligible sample (i.e., alive and not formally withdrawn), ranging from a low of 93% at age 45 years up to 99% at 21 years. There were no significant differences on baseline demographic characteristics of the attritted individuals and those involved in either the 5-, 21-, or 45-year-old follow-up studies. The current study was conducted with IRB approval from the University of Virginia, Protocol Number 4022.

Both the treatment and control groups were provided access to a social worker and to medical care, but only the treated group received free full-time year-round child care at the university child care center from infancy to kindergarten entry (Ramey et al., 1977). The ABC treatment was designed to meet professional standards of quality ECE, and to focus on scaffolded experiential learning and frequent teacher-child verbal interactions (Ramey et al., 2012).

Measures

This paper focused on IQ and academic outcomes because they showed sex-by-treatment interactions in other studies and were measured from the end of treatment into adulthood in this study. Trained data collectors individually administered the assessments in schools when the participants were 5, 6.5, and 8 years-of-age and in the participant’s home or child development research institute when the participants were 12, 15, 21, and 45 years of age. Table 1 lists IQ, reading, and math scores over time by treatment group and sex and Figure 1 plots these means by time, treatment group, and sex.

Table 1.

Outcomes by Treatment Group and Sex

	Females						Males
	Control			ECE Treatment			Control			ECE Treatment
	N	Mean	Std	N	Mean	Std	N	Mean	Std	N	Mean	Std
Cognitive Skills
WPPSI 5y IQ	23	95.91	13.93	23	101.9	11.71	21	91.19	13.25	26	101.0	10.46
WISC-R 6.5y IQ	22	92.41	10.70	23	97.57	13.40	21	92.10	13.05	24	99.29	11.26
WISC-R 8y IQ	23	92.61	11.26	22	97.18	12.83	19	94.21	13.95	26	98.38	11.01
WISC-R 12y IQ	23	86.70	10.10	24	95.21	10.24	21	90.38	13.62	24	93.00	8.37
WISCR 15y IQ	25	88.40	13.70	24	93.92	13.27	22	93.05	11.81	26	96.88	9.16
WAIS-R 21y IQ	25	84.52	9.35	24	91.17	11.37	22	86.91	8.22	26	87.23	8.63
WASI-2 45y IQ	23	89.00	16.31	23	98.39	12.69	18	91.78	17.31	24	92.17	12.17
Reading Skills
PIAT Reading Recognition 5y	22	113.1	20.69	21	123.4	17.35	20	110.2	18.70	23	117.6	17.03
PIAT Reading Recognition 6.5y	22	99.91	11.06	22	105.2	11.75	21	99.57	12.63	25	97.64	11.65
WJ Broad Reading 8y	21	85.14	11.91	21	97.33	11.70	19	83.89	11.93	26	91.31	14.03
WJ Broad Reading 12y	23	84.17	10.45	24	92.63	13.73	21	85.33	12.26	24	88.75	14.53
WJ Broad Reading 15y	25	87.84	10.79	24	96.75	14.31	22	89.23	10.93	26	91.31	11.18
WJ Broad Reading 21y	25	85.80	14.31	24	95.83	16.15	22	89.32	12.98	26	89.38	17.24
Math Skills
PIAT Math 5y	23	104.2	12.83	23	103.7	13.83	20	98.05	14.72	23	105.2	11.75
PIAT Math 6.5y	22	102.0	11.12	22	102.5	8.93	21	100.2	14.59	25	101.8	10.33
WJ Broad Math 8y	21	91.33	13.54	22	97.50	12.28	19	92.84	15.55	26	97.92	12.63
WJ Broad Math 12y	23	84.17	13.80	24	94.08	13.97	21	87.86	15.69	24	89.42	10.64
WJ Broad Math 15y	25	85.20	13.59	24	95.13	15.89	22	89.00	12.63	26	90.85	11.12
WJ Broad Math 21y	25	81.96	10.95	24	89.50	12.04	22	87.27	11.80	26	87.23	9.86

Note: WPPSI = Wechsler Preschool and Primary Intelligence Scale; WISC-R = Wechsler Intelligence Scale for Children-Revised; WAIS = Wechsler Adult Intelligence Scale; WASI-2 = Wechsler Abbreviated Intelligence Scales; PIAT = Peabody Individual Achievement Test; WJ = Woodcock Johnson Academic Achievement Battery

Figure 1. Plots of Sex x Treatment Group Means over Time.

Note: The sex x treatment group mean scores on repeated assessments of standardized IQ and achievement tests are plotted. In these plots, F=female and M=male.

IQ Skills

Participants were tested with Wechsler intelligence tests. The Wechsler Preschool and Primary Scale of Intelligence (WPPSI; Wechsler, 1967) was administered at 5 years of age, the Wechsler Intelligence Scale for Children-Revised (WISC-R, Wechsler, 1974) was administered at 6.5, 8, 12, and 15 years of age, the Wechsler Adult Intelligence Scales – Revised (WAIS-R; Wechsler, 1981) were administered at 21 years of age, and the Wechsler Abbreviated Intelligence Scales (WASI-2; Wechsler & Zhou, 2011) were administered at 45 years of age. The Wechsler scales are a widely used measure of cognitive skill level and have standardized scores with a mean of 100 and standard deviation of 15 in the norming population.

Academic Skills

Academic skills were assessed using the Peabody Individual Achievement Test (PIAT; Dunn & Markwardt, 1970) when students were 5 and 6.5 years of age and using the Woodcock Johnson Academic Achievement Battery (WJ; Woodcock & Johnson, 1977) when students were 8, 12, 15, and 21 years of age. Reading skills were measured with the PIAT Reading Recognition scales and the WJ Broad Reading Scale that combined the Letter-Word Identification and Passage Comprehension Subscales. Math skills were measured with the PIAT Math Scales and WJ Broad Math Scale that combined the Applied Problems and Calculations Subscales. The PIAT scales were and the WJ scales are widely used measures of academic skills, and both have standardized scores with a mean of 100 and standard deviation of 15.

Data Analysis

To address questions of power in prior analyses (Campbell et al., 2001), longitudinal analyses that utilized all the information from the repeated assessments instead of focusing on a single age were implemented. Longitudinal hierarchical linear models (HLM) estimated individual trajectories in separate analyses of the seven repeated measures of IQ and six repeated measures of reading and math skills and tested them for treatment and sex differences. Perhaps due to the long follow-up period, a cubic polynomial model was needed to describe change over time (see Figure 1 for plots of the sample means over time to show the higher-order pattern of change over time for these outcomes). Fitting such a high-order polynomial model is problematic because such longitudinal models are difficult to interpret and may be overfitted given the number of repeated measures. For this reason, various transformations were applied to the outcomes, and the log transformation was selected. Model fitting involved fitting an unconditional model with quadratic random effects and cubic fixed effects. The time-varying covariates were also included to account for bias the long gap of time between testing in adulthood and changing academic tests over time. Those models were simplified by omitting nonsignificant higher-order terms. Analyses were conducted using Proc Mixed in SAS v9.4 (SAS, 2004).

The models estimated individual intercepts and slopes, with the intercept set at 5 years and allowed for up to cubic patterns of change overall from 5- to 21-years. The analyses of math and reading included a time-varying variable that indicated whether the PIAT or WJ was administered at that age to account for differences in the two tests. Analyses of IQ included a time-varying variable indicating whether the assessment was from the 45-year assessment or prior assessments. Sex and treatment status were included as main effects and crossed with each other and time. The models included all interactions involving sex or treatment that were statistically significant. The treatment and treatment x sex effect sizes were computed for each assessment point when significant treatment or treatment x sex differences in main effects or rates of change were detected. The outcome data were standardized, so effect sizes could be computed directly from the models. The model fitting process for the analysis of each outcome is described in the results section below.

Results

The descriptive statistics for the outcome measures are shown in Table 1 and the sex-by-treatment group means are plotted in Figure 1.

Longitudinal Analyses of IQ

The longitudinal HLM for analyzing the natural log of IQ scores was identified through a series of analyses (see Table S1 in the online supplemental files). We first fit an unconditional model that included quadratic random and cubic fixed effects for time and the time-varying variable indicating the 45 year data. None of the quadratic random variances or covariances were statistically significant (see column 1) and were dropped from the model. The model that dropped the nonsignificant quadratic random variances and covariances (see column 2) indicated statistically significant linear random effects and cubic fixed effects and served as the basis for our analytic model. Treatment and sex were added to the model shown in column 2 and fully crossed with the time fixed effects (see column 3). Results indicated statistically significant linear random effects and cubic change over time with interactions among sex, treatment, and linear time fixed effects (see column 4). The resulting model is shown below.

$$\begin{gathered} {\mathrm{Y}}_{\mathrm{i}\mathrm{j}}={\mathrm{B}}_{0\mathrm{i}+}+{\mathrm{B}}_{1\mathrm{i}}{\text{Age}}_{\mathrm{i}\mathrm{j}+}+{\mathrm{B}}_{2\mathrm{i}}{{\text{Age}}^2}_{\mathrm{i}\mathrm{j}}+{\mathrm{B}}_{3\mathrm{i}}{{\text{Age}}^3}_{\mathrm{i}\mathrm{j}}+{\mathrm{B}}_{4\mathrm{i}}{\text{Adult}}_{\mathrm{i}\mathrm{j}}+{\mathrm{e}}_{\mathrm{i}\mathrm{k}} \\ {\mathrm{B}}_{0\mathrm{i}}={\mathrm{\delta }}_{00}+{\mathrm{\delta }}_{01}{\text{Sex}}_{\mathrm{i}}+{\mathrm{\delta }}_{02}{\text{Treatment}}_{\mathrm{i}}+{\mathrm{\delta }}_{03}{\text{Sex}}_{\mathrm{i}}\text{x}{\text{Treatment}}_{\mathrm{i}}+{\mathrm{\lambda }}_{0\mathrm{i}} \end{gathered}$$

$$\begin{array}{l}{\mathrm{B}}_{1\mathrm{i}}={\mathrm{\delta }}_{11}+{\mathrm{\delta }}_{11}{\text{Sex}}_{\mathrm{i}}+{\mathrm{\delta }}_{12}{\text{Treatment}}_{\mathrm{i}}+{\mathrm{\delta }}_{13}{\text{Sex}}_{\mathrm{i}}\text{x}{\text{Treatment}}_{\mathrm{i}}\\ {\mathrm{B}}_{2\mathrm{i}}={\mathrm{\delta }}_{22};{\mathrm{B}}_{3\mathrm{i}}={\mathrm{\delta }}_{33};\\ {\mathrm{B}}_{4\mathrm{i}\mathrm{k}}={\mathrm{\delta }}_{44}+{\mathrm{\delta }}_{41}{\text{Sex}}_{\mathrm{i}}+{\mathrm{\delta }}_{42}{\text{Treatment}}_{\mathrm{i}}+{\mathrm{\delta }}_{43}{\text{Sex}}_{\mathrm{i}}{\text{x}\text{Treatment}}_{\mathrm{i}}\end{array}$$

The results of this final longitudinal analysis of log-transformed IQ are shown in Table 2. Results indicated significant treatment differences at the end of treatment ($B=0.57$), but the treatment impacts did not significantly vary by sex $(B=-0.20)$. The participants showed significant linear $(B=-0.11)$, quadratic $(B=0.01)$, and cubic ($B=0.001$) patterns of change over time from 5 to 21 years of age. Rates of linear change over time were significantly different by sex $(B=0.09)$, and these sex differences were moderated by treatment $(B=-0.04)$. The difference in the rate of linear change-by-treatment status was not significantly different for the females ($B=0.01,se=.01$), but was significantly negative for the males $(B=-0.04,se=.01,p<.05$). No evidence emerged that indicated sex or treatment differences in quadratic or cubic rates of change over time.

Table 2.

Testing for Sex × Treatment Differences on Adult Cognitive and Academic Outcomes

	Ln(IQ)	Ln(Reading)	Ln(Math)
Fixed Effect	β(se)	β(se)	β(se)
Intercept	0.367***(0.095)	0.897*** (0.129)	0.117(0.087)
Administered test was Piat		0.205* (0.088)	0.509*** (0.071)
Treatment	0.574**(0.179)	0.358** (0.127)	0.261(0.142)
Male	−0.204(0.185)		−0.12(0.142)
Treatment × male	0.175(0.357)		0.195(0.283)
Time	−0.112**(0.034)	−0.599*** (0.047)	−0.039*** (0.007)
Treatment × time	−0.013(0.01)		0.012 (0.012)
Male × time	0.093***(0.026)		0.013 (0.012)
Treatment × male × time	−0.044*(0.02)		−0.057* (0.024)
Time2	0.012*(0.005)	0.077***(0.006)
Time3	−0.005***(0.002)	−0.003***(0.000)
Adult	0.356***(0.067)
Adult × male	0.116(0.125)
Adult × treatment	0.019(0.134)
Adult × male × treatment	−0.046(0.25)
Random Effects	σ2 (se)	σ2 (se)	σ2 (se)
Intercept variance	0.7166** (0.1152)	0.4769*** (.0863)	0.3659*** (.0731)
Intercept, linear slope covariance	0.0013*** (0.0003)	−0.0396** (.0144)	0.0005 (.0046)
Linear slope variance	−0.0102* (.0046)	0.0135*** (.0037)	0.0017*** (.0006)
Intercept, quadratic slope covariance		0.0021** (.0008)
Linear slope, quadratic slope covariance		−0.0006 ** (.00002)
Quadratic slope variance		0.00003** (.00001)
Residual	0.2155*** (0.0142)	0.1615*** (.0139)	0.3080*** (.0226))

Note:

^*p<.05

^**p< .01

^***p < .001

IQ measured by Wechsler Preschool and Primary Intelligence Scale, Wechsler Intelligence Scale for Children-Revised, and Wechsler Adult Intelligence Scale, Reading and math measured by Peabody Individual Achievement Test and Woodcock Johnson Academic Achievement Battery

Finally, the discontinuous longitudinal model compared treatment and sex differences from the 21- and 45-year follow-up studies. Results indicated significant differences in overall scores $(B=0.36)$ but not in sex $(B=0.12)$, treatment $(B=0.02)$, or sex-by-treatment $(B=-0.05)$ differences.

Figure 2 illustrates the findings from this longitudinal analysis by showing the effect sizes over time overall and by sex. As shown in the first set of bars in this figure, the overall treatment effect size ranged from .57 at 5 years to .44 at 45 years, but male and female treatment effect sizes showed very different patterns of change over time. The time × sex × treatment interaction is reflected in the declining treatment impacts over time for the males and the increasing treatment impacts over time for the females.

Figure 2. Treatment Effect Sizes by Age.

Note. Figure shows the standardized mean differences between treatment and control participants on IQ, reading, and math assessments.

Longitudinal Analyses of Reading Skills

The same model building process was used to identify the HLM used to analyze the natural log of reading. Table S2 of the online supplemental files detail the process. First, the quadratic random and cubic fixed effect model was fitted, including the time-varying variable indicating which test was used for each data collection period. The model indicated no further trimming was warranted. Sex and treatment were added to this unconditional model and crossed with time (see column 2). The interactions among time, sex, and linear, quadratic, and cubic time were not statistically significant, and the interactions with quadratic and cubic time were dropped. The next model tested whether time and sex interacted with linear changes in the absence of the higher order interactions. They continued to be statistically nonsignificant and were dropped. The final model is shown below and the results in the final column of the online Table S2.

$$\begin{gathered} {\mathrm{Y}}_{\mathrm{i}\mathrm{j}}={\mathrm{B}}_{0\mathrm{i}+}{\mathrm{B}}_{1\mathrm{i}}{\text{Age}}_{\mathrm{i}\mathrm{j}}+{\mathrm{B}}_{2\mathrm{i}}{{\text{Age}}^2}_{\mathrm{i}\mathrm{j}}+{\mathrm{B}}_{3\mathrm{i}}{{\text{Age}}^3}_{\mathrm{i}\mathrm{j}}+{\mathrm{e}}_{\mathrm{i}\mathrm{k}} \\ {\mathrm{B}}_{0\mathrm{i}}={\mathrm{\delta }}_{00}+{\mathrm{\delta }}_{01}{\text{Sex}}_{\mathrm{i}}+{\mathrm{\delta }}_{02}{\text{Treatment}}_{\mathrm{i}}+{\mathrm{\delta }}_{04}{\mathrm{P}\mathrm{I}\mathrm{A}\mathrm{T}}_{\mathrm{i}}+{\mathrm{\lambda }}_{0\mathrm{i}} \\ {\mathrm{B}}_{1\mathrm{i}}={\mathrm{\delta }}_{10}+{\mathrm{\lambda }}_{1\mathrm{i}},{\mathrm{B}}_{2\mathrm{i}}={\mathrm{\delta }}_{20}+{\mathrm{\lambda }}_{2\mathrm{i},}{\mathrm{B}}_{3\mathrm{i}}={\mathrm{\delta }}_{30} \end{gathered}$$

The second column of Table 2 shows the results of the longitudinal analyses of the natural log of PIAT and WJ reading standardized scores with this final model. Results indicate significant treatment impacts $(B=0.36)$ at entry to school without evidence that those impacts varied over time. Children, on average, scored higher on the older PIAT test ($B=.21$) than the WJ test. After accounting for the changing tests, the participants showed significant linear ($B=-0.60)$, quadratic $(B=0.08)$, and cubic $(B=-0.003)$ patterns of change over time from 5 to 21 years of age.

Longitudinal Analysis of Math Skills

Table S3 in the supplemental files describes the model building process for analyzing the natural log of the math scores over time. The model with quadratic random and cubic fixed effects, and the time-varying variable indicating test yielded a random covariance matrix that was not positive definite. Accordingly, the quadratic random effect was dropped, and the model yielded statistically significant linear random and fixed effects. Sex and treatment were added to this model and crossed with linear time. The resulting model is shown below.

$$\begin{gathered} {\mathrm{Y}}_{\mathrm{i}\mathrm{j}}={\mathrm{B}}_{0\mathrm{i}}+{\mathrm{B}}_{1\mathrm{i}}{\text{Age}}_{\mathrm{i}\mathrm{j}}+{\mathrm{e}}_{\mathrm{i}\mathrm{k}} \\ {\mathrm{B}}_{0\mathrm{i}}={\mathrm{\delta }}_{00}+{\mathrm{\delta }}_{01}{\text{Sex}}_{\mathrm{i}}+{\mathrm{\delta }}_{02}{\text{Treatment}}_{\mathrm{i}}+{\mathrm{\delta }}_{03}{\text{Sex}}_{\mathrm{i}}\text{x}{\text{Treatment}}_{\mathrm{i}}+{\mathrm{\delta }}_{04}{\text{PIAT}}_{\mathrm{i}}+{\mathrm{\lambda }}_{0\mathrm{i}} \\ {\mathrm{B}}_{1\mathrm{i}}={\mathrm{\delta }}_{10}+{\mathrm{\delta }}_{11}{\text{Sex}}_{\mathrm{i}}+{\mathrm{\delta }}_{12}{\text{Treatment}}_{\mathrm{i}}+{\mathrm{\delta }}_{13}{\text{Sex}}_{\mathrm{i}}\text{x}{\text{Treatment}}_{\mathrm{i}}+{\mathrm{\lambda }}_{1\mathrm{i}} \end{gathered}$$

The second column of Table 2 shows the results of fitting this model to the longitudinal assessments of PIAT and WJ math standardized scores. Results indicate nonsignificant treatment impacts $(B=0.26)$ and no significant sex $(B=-0.12)$ or sex-by-treatment $(B=0.20)$ differences at entry to school. Children, on average, scored higher on the older PIAT test $(B=.51)$ than the WJ test. The participants showed significant linear $(B=0.06)$ patterns of change over time from 5 to 21 years of age. By the 8-year assessment, significant treatment differences emerged ($B=.30,se=.14,p<.05)$ and continued to be statistically significant through the 21-year assessment $(B=.46,se=.20,p<.05)$. A significant sex × treatment × time interaction $(B=-.06$) indicated that the treatment impacts changed differently over time for males and females. The difference in rate of linear change by treatment status was significantly positive for the females ($B=0.04,se=.02,p<.05$), but was not significantly different for the males ($B=-.02,se=.02)$.

The estimated overall treatment effect sizes shown in the third set of bars in Figure 2 also illustrate the sex × treatment × time interaction. While the overall treatment effect size increased over time from .23 at 5 years to .46 at 21 years, the effect sizes for the females increased over time and for the males decreased over time. At 5 years of age, the estimated treatment impact for females was .16 and for males was .36. By = 21 years of age, the estimated effect size from this model was about .80 for females and about .09 for males.

Sensitivity Analyses

Follow-up analyses were conducted to test whether model misspecification may have yielded biased estimates of impacts at either end of the polynomial model. The estimated standardized mean differences were computed using the observed data, multiple imputed data, and under the models described above. Results suggest the estimated impacts at 21 years of age may be underestimated under the models, due to model misspecification and overestimated with the observed ages at 5 years of age due to missing data (see Table S4 supplemental materials).

Discussion

This study extends the report of overall treatment differences in IQ and academic skills through 21 years of age in the ABC Program (Campbell et al., 2002) by demonstrating that treatment impacts on IQ extend to middle age and that sex differences in treatment impact on IQ and math skills emerged over time. Results of analyses of IQ and math skills indicated that the treated females maintained or experienced an increase in their advantage over control females over time whereas the treated males experienced a decrease in their advantage over control males over time. These results are consistent with prior analyses and extend them by suggesting that there are different patterns of treatment impacts over time for Black females than males in IQ and academic skills and that those differences emerged after the treatment ended.

These findings are consistent with many prior studies, especially those with large proportions of Black participants. Prior studies reported Black females had larger treatment impacts during the school years for both this study and PPP (Anderson, 2008; Magnuson et al., 2016). Similar findings of larger impacts for females than males during early adulthood were reported for the early cohorts of Head Start that included many Black children (Deming, 2009), but not with the more recent evaluation of the Boston pre-kindergarten program. The Boston study reported males showed more treatment impacts than females in early adulthood in a sample that included about 40% Black individuals (Gray-Lobe et al., 2023).

None of these studies tested whether those sex differences in academic skills were present at the end of the ECE treatment period. The present paper extends findings from prior studies by demonstrating there were no sex differences in treatment impacts at the end of treatment, but they emerged linearly over time during the school years and are maintained through middle age. By demonstrating that sex differences in treatment impacts emerged after the Black children in this study entered elementary school, these findings are most likely explained by sex differences in subsequent experiences rather than in their early childhood experiences.

Possible Explanations for Sex Differences in Treatment Impacts that Emerge during the School Years

Schooling is a major factor in the acquisition of academic and cognitive skills (Hanushek & Woessmann, 2008), and therefore possible differences in how Black males and females are treated at school is a potential explanation for the emergence of sex differences in treatment impacts on these skills during the school years. Black females and males face structural racism throughout their lives, but the form of racial discrimination appears to vary by sex (NASEM, 2019). Evidence shows that schools too often treat Black children more negatively than White children, with higher levels of discipline and suspension (Rocque & Paternoster, 2013), teachers with racially biased negative stereotypes and expectations (Ferguson, 2001), and culturally insensitive expectations regarding “good” and “bad” behavior for students (Silva et al., 2014). Furthermore, the 2008 Task Force on Resilience and Strength in Black Children and Adolescents (APA, 2008) and the 2012 Presidential Task Force on Educational Disparities (APA, 2012) reports from the American Psychological Association note the differential treatment of ethnic and racial minority children based on sex leads to disparities early in children’s education. Black boys are likely to be rated as having higher rates of problem behavior and referred for disciplinary infractions more often than girls; teachers, and some parents, are also likely to have lower educational expectations for boys than girls regardless of their academic performance; and boys are more likely to experience institutional barriers and racial profiling (Epstein et al., 1998; Ferguson, 2001; Roderick, 2003; Strayhorn, 2010). Teachers are more critical of Black than White students and of boys than girls, and are especially critical of Black boys, viewing them as more hostile even when they display similar behaviors as other children (Mulligan & Flanagan, 2006). This implicit bias too often results in academic disengagement among Black boys (Langhout & Mitchell, 2008), and, thus, could explain why ECE program impacts on educational outcomes often attenuate more for the males than the females. This suggests that experiences in school might have altered the academic trajectories of the treated male participants differently than for the treated female participants in ABC.

The social context of the ABC study might further help to explain why these sex differences in impacts emerged during the school years. The racially segregated schools in Chapel Hill, N.C. desegregated by 1966 (Greene, 2005). According to firsthand accounts of Chapel Hill residents, although many in the Black and White communities supported desegregation, the Black community was upset about how the desegregation occured. Many teachers from the Black schools were demoted or let go, thus many teachers and administrators in the Chapel Hill schools were White and female. In addition, the identities and traditions of the Black schools were not preserved during the desegregation process (Greene, 2005).

Given the composition of ABC at the time, the Black children in the ABC treatment group likely began elementary school with considerably higher verbal and academic skills than other Black children and with an expectation that teachers would actively engage with them and scaffold learning (Haskins, 1985). Both their higher skill levels and an expectation of active support from teachers, likely challenged their White teachers’ expectations regarding the academic and social skills of Black children as being less advanced academically and more passive in interacting with the teacher (Ferguson, 2001; Silva et al., 2014). These contradictory expectations regarding appropriate social behavior of Black children in the classroom of the teachers and of the treated children likely explain the early negative teacher ratings (Haskins, 1988). Teachers in kindergarten reported that treated children were more advanced cognitively, but also more hostile than the control children (Haskins, 1985). Teachers maintained this view of greater hostility among treated boys, but not girls, through second grade. While supposition, it is possible that the girls in the treatment group adapted their classroom behaviors to match their teachers’ expectations whereas the boys in the treatment group did not, or that the teachers, who were mostly females, saw the girls’ behaviors through a female lens that favored girls but did not give this same benefit to the boys (Brophy & Good, 1973).

Implications for Policy & Practice

The ABC and PPP studies have been influential in demonstrating the impact of ECE programs, but studies of these interventions reveal that even these intensive research intervention ECE projects did not inoculate children from subsequent school and life experiences. In this context, it is important to note that results from this study and PPP have been extremely influential in promoting early childhood education policies for all children despite their focus on Black children and the evidence that only females experience those long-term impacts on cognitive and academic skills. The finding of differential treatment effects favoring females on educational outcomes is not new (Anderson, 2008; Magnuson et al., 2016). The sex differences in school-age and adult outcomes have been presented previously by economists (e.g., Anderson, 2008; Magnuson et al., 2016; Heckman et al., 2010), but the differences in long-term impacts by sex in these studies is not often acknowledged by developmentalists and policymakers. We hope that publishing in the developmental literature and focusing on developmental trajectories will increase attention to this issue beyond econometric literature. Generalizing the findings from this study to today’s social context is limited because of large changes in social context in general (Duncan & Magnuson, 2013). Nevertheless, we believe that the emergence of sex differences in treatments impacts during the school years in this study of Black children is relevant in today because issues of differential structural racism by sex in the schools persist and likely interfere with the maintenance of ECE program impacts over time for Black children.

Limitations

There are a few limitations of the present study that should be acknowledged. First, the sample size is small, reducing power to detect significant sex differences in the effect of the treatment. This study was conducted by researchers 50 years ago and involved many components that make it difficult to replicate such as free transportation, daily monitoring of classrooms, child-centered activities, weekly coaching with a focus on scaffolded conversations and warm teacher-child relationships, and on-site medical care (Ramey et al., 2012), so it likely differed from many of today’s large-scale early childhood education programs. Third, we do not present any evidence linking our findings of sex differences in treatment impacts on developmental trajectories to schooling experiences. Our interpretation of the findings is based, in some part, on prior analyses of teacher ratings of the participants in this study, and in large part on the growing literature of implicit racism, especially of Black boys. Fourth, our polynomial models have the advantage of using all data from all individuals on that outcome but make strong assumptions about the shape of change over time. Findings, especially regarding treatment impacts at the youngest and oldest ages, are biased to the extent the model is incorrectly specified. Finally, some critics argue that the ABC sample is outdated because the ECE landscape has changed substantially since the 1970s (Duncan & Magnuson, 2013).

Despite these limitations, we believe this paper contributes to literature. The long-term nature of the study allows us to examine adult outcomes, which is not possible in more recently conducted studies. Furthermore, Black children today continue to face systemic racism and biased experiences in their daily lives including in school settings (Gardner-Neblett et al., 2021).

Conclusion

Overall, the results of the present study support the importance of considering sex when examining the impacts of ABC participation over time. Findings of sex differences in the maintenance of IQ and math treatment impacts during the school years suggest there might be different societal factors that maintain longer-term impacts for Black females but disrupt longer-term impacts for Black males. At a minimum, these findings suggest that Black children’s experiences once they enter the formal school environment likely play a substantial role in the maintenance of ECE impacts.

Public Significance:

Findings from this paper provide further evidence that ECE can improve educational outcomes for low-income Black children, but that subsequent experiences may erode those impacts especially for low-income Black males.