A Multicomponent, Preschool to Third Grade Preventive Intervention and Educational Attainment at 35 Years of Age

This cohort study examines the association between a preschool to third grade intervention with educational attainment at midlife and differences by program duration, sex, and parental educational level.

Note. 1. Family risk indicators. Characteristics were measured from administrative records (e.g., birth records) from primarily ages 0 to 5, parent reports up to age 12, and for home environment problems and adverse child experiences retrospective reports by participants. Demographics were measured from school records. Ns for comparison group males and females were 287 and 256, respectively. 7 comparison group and 1 program group participant had missing gender information. *95% CI does not include zero. Note. Adjusted with IPW attrition and selection. Gender, race, and sociodemographic factors are included as covariates. Child welfare history by age 4 is not included in the models of Bachelor s' degree and Masters' degree or higher because it predicted the outcomes perfectly. There are 57 cases (4%) report Master degree or higher. When examined by the 6 groups, the "no prek had any follow on group" predicted the outcome perfectly, so it was not examined. The correlation is the unadjusted biserial correlation between the extent of intervention in years and the outcome, and it adjusts for the dichotomous outcomes.
Groups are as follows: (1)Prek to 3 rd grade, n=160 (2) Prek to 2 nd grade n=351 (3) Prek to 1st grade n=116 (4) Prek/Prek + K n=277 1 (5) No Prek had anyfollow on n=149 (6)   Note. Gender, race, and sociodemographic factors are included as covariates. Child welfare history by age 4 is not included in the models of Bachelors' degree because it predicted the outcomes perfectly. There are 57 cases (4%) report Masters' degree or higher. The adjusted means by the 6 groups are too small to report. When examined by the 4 groups, the "no prek and any school-age group" predicted the outcome perfectly, so that group was not included in the analysis. 1. Reference group for all comparisons. *95% CI does not include zero.

Study Background
The Chicago Longitudinal Study (CLS) is a prospective cohort investigation of early childhood experiences and well-being over the life course (1)(2)(3). The study sample of 1,539 children of the same age attended early childhood programs in the Chicago Public Schools over 1983-1986, and have been followed to midlife. Nearly two thirds (N = 989) attended the Child-Parent Center (CPC) Education Program in preschool and kindergarten and one third (N = 550) attended alternative kindergarten programs in the Chicago Public Schools. All children completed kindergarten in the spring of 1986, with those continuing in CPC up to 2 nd or 3 rd grade. CLS participants were born in 1979 and 1980 and grew up in high-poverty neighborhoods in Chicago. Matching the racial and ethnic composition of the neighborhoods, 92.9% of the cohort is African American and 7.1% are Hispanic. Nearly 80% of participants resided in areas of concentrated poverty, defined as 60% or more of individuals at/below 185% of the federal poverty line. The high degree of economic disadvantage faced by study participants is fully described elsewhere (4,5).
In this article, data are included from birth to age 35, up to 32 years after program enrollment at ages 3-4 years. Data have been collected from many sources and provide a full accounting of not only of educational attainment but predictors and antecedents. These include birth records, K-12 school records, parent and teacher surveys, participant surveys, involvement in the criminal justice system, employment and earnings, and school enrollment and graduation from high school through postsecondary education. The National Student Clearinghouse provides the most complete records of college degree completion, and these records are supplemented with participant reports, and data directly from institutions.
The CLS is currently in the early midlife phase and investigators are examining the links between early childhood experiences and well-being up to age 35. Previous phases of the study have investigated the individual, family, intervention and school predictors of well-being up to age 10, 15 to 18, 18 to 24, and age 26 to 28 (6)(7)(8). Samples sizes have ranged from 1,233 to 1,473, which reflect the combination of survey response and administrative records. The sample size in this article is 1,398, and includes individuals with valid data on educational attainment up to age 35. The main sources are the National Student Clearinghouse, Illinois Shared Enrollment and Graduation Consortium (which was later incorporated into the Clearinghouse), school administrative records, and self-reports.

Design and Validity of Program Contrasts
The CPC program group is a complete cohort of 989 children who attended preschool and kindergarten in all 20 CPCs. The comparison group of 550 children attended full-day kindergarten programs in five randomly selected schools participating in the Chicago Effective Schools Project (CESP; N = 374) or in CPC-affiliated schools (N = 176). Both interventions were for children at risk of school underachievement due to poverty and related factors. 15% of the comparison group attended Head Start preschool (3,8). Consequently, the CLS is a matched group, alternative-intervention quasi-experimental design in which the performance of CPC participants is compared to children in demographically-similar neighborhoods who enrolled in the usual early childhood interventions available to vulnerable families in the Chicago Public School District. Schoolage services from first to third grades were available to all children who enrolled in the CPC-affiliated schools regardless of preschool or kindergarten participation. eTable 1 shows the characteristics of the original sample and by attrition status by age 35.
The comparison group matched the program group on age, eligibility and participation in intervention, and neighborhood and family poverty. Eligibility criteria for CPC and CESP enrollment were similar and included (a) residence in a school area receiving federal education aid from Title I of the Elementary and Secondary Education Act of 1965, (b) demonstration of high educational need, and (c) parents agree to participate in the program to support children's learning. For CPC, this was specified as being involved up to one-half day per week in the program.
Because the CPC program was a larger-scale established intervention operating in the school district since 1967, had positive evidence of effectiveness on student achievement and school success, and was reserved for those most in need, random assignment to groups was impossible. Not only would it have been unethical, as program benefits were known and not uncertain, but contamination (noncompliance) would have most surely occurred. Children and families randomly assigned to the comparison condition in this school-wide intervention could not be required to take the assignment, especially if they lived in a CPC attendance area. Nor would they have given the known benefits and the legal requirement that the program serve those most in educational need. Cross-contamination has occurred for other multi-level preschool to third grade interventions. The accumulated evidence in CLS demonstrates that findings are interpretable as program effects, as estimates are consistent across many different model specifications, comparison group contrasts, and analytic techniques. These techniques include latent-variable structural modeling, propensityscore weighting, matching, and stratification, and alternative covariate specifications. Causal mechanisms of change have also been shown to be consistent with the theory of intervention. These findings and interpretations are fully documented elsewhere (9-12).

Adult Follow-Up and Group Equivalence
At an average age of 35.1 years, 90.5% (N = 1,398) of the original CLS cohort had data for educational attainment by May 31, 2015. Sample recovery rates for the program and comparison groups were, respectively, 91.4% and 89.8%. Approximately 60% of the sample resides in Illinois (based on the age 35 survey), with many others (15%) remaining in the Midwest. The study sample represents the original sample well on nearly all attributions and no evidence of selective attrition has been found in prior studies. Data that have been missing for education, health, and behavioral outcomes are accounted for measured covariates and predictors. Thus, the assumption of missing at random has been commonly found. eTables 1 and 2 show the characteristics of age 35 follow-up and attrition sample as well as the program and comparison groups at the beginning of the project. Child and family characteristics were measured from birth records, school and health administrative records, and family surveys. The latter assessed baseline characteristics when children were primarily between 7 and 12 years; and they supplemented records data. The p-values show the significance of group differences at the beginning of the study and at follow up. For early home environment (e.g., adverse child experiences), retrospective reports from participants were used. As shown in eTable 1, the characteristics of the original and age 35 study samples are generally similar, including for the number of family risks (4.52 vs. 4.49, respectively), CPC preschool (64.3% vs. 64.7%, respectively) and extended program participation (35.9% vs. 36.8%, respectively). However, the attrition sample had a higher proportion of males compared to the follow-up sample (71.4% vs. 47.6%) and was more likely to be disadvantaged. Since the attrition sample represented only 9.2% of the original cohort, the impact of these differences was small.
As shown in eTable 2, the two consistent differences that have been found are that program participants are from higher poverty neighborhoods and mothers have higher rates of high school completion as reported in birth records. These differences are reduced once other socio-demographic factors are taken into account. However, the advantage in parent education is specific to program females, as rates of high school graduation are nearly identical for program and comparison males. Indeed, no differences between program and comparison males have been found for any baseline measure or covariate-which number over 30 indicators (eTable 2).

Child-Parent Center (CPC) Education Program
The CPC program is fully described in many reports (4,5,13). It began in 1967 in four new centers on the city's westside. This was the result of the landmark Elementary and Secondary Education Act of 1965 for which federal funding from Title I of the Act was used by the school district to open the centers. The Chicago Public School District was the first to use Title I for preschool and thereby established CPC as the second oldest (after Head Start) federally-funded preschool. Although CPC began as a comprehensive preschool program, children received continuing services in kindergarten and the early grades the following year, resulting in the preschool to 3 rd grade program that it is today. The program was modified as a school reform model in 2012 as part of expansion in and outside of Chicago funded by the U. S. Department of Education. Six core elements are implemented; effective learning experiences, collaborative leadership, aligned curriculum, parent involvement and engagement, professional development, and continuity and stability. A system of resources is now available to plan, monitor, and evaluate progress for wide-scale use and sustainability. This includes a program manual for implementation (14).
The program was developed in response to three major problems facing Chicago schools: low rates of attendance, family disengagement with schools, and low student achievement. The conceptual foundation is that well-being is a product of proximal and distal influences at multiple levels of contexts (individual, family, school, community) experienced during the entire early childhood period (ages 3 to 9). The program's focus on the quality and continuity of learning environments indicates that optimal development can be promoted through enriched experiences and settings during early childhood and the transition to school. Due to discontinuities in instructional support and philosophy between early childhood and school age settings, improvements in the integration and alignment of services during this important ecological transition improves children's levels of readiness for kindergarten that are sustained over the elementary grades.
CPC provides comprehensive education and family support services. Under the direction of a leadership team at each site and in collaboration with the Principal, CPC enhances school readiness skills, increases early school achievement, and promotes parent involvement. It is a stand-alone school or center in which all children receive services. Health and nutrition as well as referrals to social services also are provided. The Head Teacher (HT) or Director works under the leadership of the elementary school Principal. HTs are the administrative leads for the program and manage implementation, provide coaching and supervision to staff, and help establish expectations of performance. The Parent Resource Teacher (PRT) directs the CPC's parent resource room and family services, and outreach activities are organized by the School-Community Representative (SCR). Health services are coordinated between the preschool and elementary grades.
After preschool participation at ages 3 and/or 4 in small classes with student-teacher ratios of 17:2, the K-3rd component provides reduced class sizes (maximum of 25), teacher aides for each class, continued parent involvement opportunities, and enriched classroom environments for strengthening language and literacy, math, science, and social-emotional skills. The key program elements are described in further detail (4,5).  For high school completion status, 114 participants reported postsecondary education attendance, but they are missing on whether they completed high school via diploma or GED. Their types (high school graduation or GED) of high school completion were estimated based on available information from other data sources, including ISEG, NSC, CPS, and DCFS. For years of education, 2 participants are missing last grade they completed before they dropped out of school.

Educational Attainment Measures by Age 35
Across all sources of information, 72% of the age 35 follow-up sample had available administrative records from one or more sources, with the remainder from survey reports. These percentages were similar by intervention status: preschool (73%), comparison (70%), extended intervention to 2 nd or 3 rd grade (74%), and less extended participation (0 to 3 years; 70%).
The sample size for educational outcomes at age 35 is 1,398. Among the 1,539 participants in the original sample, we have educational information on 1,473 individuals. Participants fit the following criteria to be included in the age 35 sample of educational outcomes: Findings are reported as adjusted rates (percentages) or means between groups after accounting for potential biases in estimation. The covariates included 15 child and family attributes. With the exception of public aid, subsidized lunches, and home environment, the covariates were measured primarily from birth to age 5 from administrative records (e.g., birth records, public aid, child welfare, neighborhood poverty) or family surveys. A dummy code for missing data on the covariates also was included in the models to assess if estimates based on multiple imputation were similar for those with and without missing data. Preprogram measures and covariates from administrative records came from the Illinois Department of Child and Family Services, the Illinois Department of Public Health, the Illinois Department of Health and Family Services, and the Chicago Public Schools. School and neighborhood poverty status were from U. S. Census records. Four models were estimated for each outcome. The unadjusted model is also included as point of reference.
Model 1 (covariate-adjusted). This is the standard probit-regression model that includes 15 covariates, including earlier (preschool) and later (school-age) intervention entered simultaneously. PK-3 intervention impacts were estimated separately. Prior studies demonstrate the consistency of this model with latentvariable, propensity score weighting, and propensity score matching (5)(6)(7)(8)(9). The model provides a comparison with the others Model 2. (attrition correction). This specification is Model 1 estimated as a weighted regression, whereby the weight is the Inverse Probability Weighting (IPW) for attrition. In this approach (20,21), the predicted probability of being in the recovery sample (R = 1; otherwise 0) is estimated for each measure of educational attainment via probit regression conditioned on the predictors (X) hypothesized to influence sample recovery. The propensity score was estimated as: (1) P 1i = Pr(R = 1|X) + e The model had 26 variables, including program participation and the covariates as predictors (see eTable 4). The inverse of this predicted probability (1 / P 1i ) was used as a weight variable for the recovery sample (1 / (1 -P 1i ) for the attrition sample) in all outcome models after verifying that data were missing at random (controlling for Xs) and that propensity distributions between groups overlapped. The IPW approach generally produces estimates with the lowest variances and standard errors in large samples (22,23). Unlike propensity score matching, IPW uses all available data and can be combined with other propensity scores to estimate more complex models (23). eTables 5 to 12 and eFigures 1 to 4 show further results and supporting evidence.

Model 3.
(program selection correction). Findings were also reported for models using IPW weights (P 2i ) for selection into each of the intervention components. The program selection scores were estimated as follows, in which T denotes participation in the program (T = 1; otherwise 0) and X a set of covariates: (2) P 2i = Pr(T = 1|X) + e The estimated model for program selection (preschool, school-age, and extended intervention) was based on 17 preprogram predictors with weights 1/ P 2i for the program group and 1/(1 -P 2i ) for the comparison group (see eTable 3). Estimates based on propensity score matching and regression-based propensity score adjustments yielded similar results.

Model 4.
(attrition and program selection correction). As the main model of the study, this specification is the IPW weighted regression with the multiplicative weight term (Equation 3: P 3i = P 2i * P 1i + e) as the weight variable. This model estimates the adjusted group differences accounting for attrition from the sample and selection into the program. The models were estimated separately for preschool/school-age and extended intervention (PK-3). See eTables 5 to 12. Subgroup results are shown in eTables 7 to 10 and eFigures 3 and 4.
The validity of each model in producing estimates interpretable as program (causal) effects is based on three major assumptions that have been met in this and prior studies (4)(5)(6)(7)(8)(15)(16)(17)(18). First, participation and selection into the program is unaffected by the services other participants receive. This absence of treatment contamination or cross-over between program and control conditions is called the stable-unit-treatment-value assumption. Second, participation or selection in the program (T) is conditionally independent (given covariate Xs) of the response or outcome (Y) of intervention (i.e., Pr(T/X, Y) = Pr(T/X). That is, after accounting for the observed (measured) predictors of participation, the estimated propensity score for each individual is uncorrelated with one or more outcomes. Thus there is no specification error or omitted variables that are jointly correlated with program participation and outcome. A corollary assumption is that the distribution of propensity scores for each program group shows common support in that they are predominately overlapping (e.g., have similar propensities for participation or attrition). Supporting evidence for these three assumptions indicate that strongly ignorable treatment assignment was present, which is a functionally similar to groups being randomly assigned to the program (24,25).
To enhance interpretability, coefficients from probit regression were transformed to marginal effects. 95% confidence intervals were reported. Subgroup effects were highlighted for outcomes showing overall main effects. See eTables 5 to 12 and eFigures 1 to 4 for results. Clustering by 25 program and control sites at the beginning of the intervention made no difference in estimates (intra-class corrections averaged .05 based on a two-level unadjusted variance components model). This is not surprising given the length of time between participation and outcome measurement.