ABSTRACT
Do early childhood investments for low-income children narrow the academic achievement gap in elementary school? I study this question in the context of Head Start by using a new variation in federal funding expansions across counties in the 1990s. Using student-level data from Texas, I find that exposure to more generous Head Start funding during childhood significantly improved test scores, particularly for low-income Hispanic students. Hispanics benefited from funding expansions through increased access to Head Start and improvements in program inputs. These advances enhanced their language proficiency and reduced the likelihood of their special education needs during elementary school.
I. Introduction
Racial, ethnic, and socioeconomic disparities in academic achievement have been ongoing issues in the U.S. education system (Reardon, Robinson, and Weathers 2008). These disparities persist over time, raising serious concerns about children’s life prospects and the state of social mobility in the United States (Heckman 2006; Duncan and Murnane 2011).
Do early childhood investments play a role in narrowing the academic achievement gap? Early childhood is a period of physical, cognitive, socioemotional, and language development that a child’s environment can influence. There is strong theoretical support in the fields of economics, neuroscience, and child development that points to early childhood as a critical time to invest in. Despite the strong theoretical support, regardless of the program’s scale, many early childhood interventions targeting skill development demonstrate initially promising results that fade out when students reach the third grade (Bailey et al. 2017). On the other hand, prior research provides evidence on the long-term effects of such programs on children’s socioeconomic and health outcomes. Given the promising initial results and long-term gains, the fade-out pattern in medium-run outcomes is puzzling.
I study whether early childhood investments close the achievement gap in elementary school in the context of the largest federal early childhood program in the United States, Head Start. The program began in 1965 to provide education, health, and social services to low-income children. Although Head Start is shown to improve children’s long-term outcomes,1 critiques have indicated that the program has not fulfilled its mission of closing the achievement gap across socioeconomic groups when children are in elementary school (Barnett 2011). The best available evidence is based on a random assignment of 4,442 children to a national sample of Head Start centers in the early 2000s. It reports small impacts on test scores at the end of the program participation year that fade out by the third grade (Puma et al. 2012). Studying the period of the 1990s, when federal Head Start funding grew substantially, I show that the program reduced the achievement gap between Hispanic and White students in the third grade in Texas.
The 1990s were a golden age for Head Start due to immense expansion and the passing of two important acts to improve the program’s capacity and quality. I investigate two research questions related to this time period. First, to what extent did federal Head Start funding expansions during the 1990s affect student performance in Texas? Second, how does the effect of public investments relate to the way funds are spent? The investigation of the first question provides new insight into the old debate. The second question presents new evidence on whether public investments were spent as planned, by linking administrative data on Head Start program characteristics and budgets to funding expansions. The evidence concerning the second question has been limited, primarily due to data availability and quality (Currie and Neidell 2007). Furthermore, which program quality measures in early childhood education matter most for child outcomes is still an open question (Blau and Currie 2006; Walters 2015).
My empirical strategy uses a panel fixed effects model that leverages variation of cohort exposure across counties in the timing and scale of Head Start funding expansions in the 1990s. The main identification assumption is that funding expansions are exogenous to other underlying geographic-level trends in test scores. I employ various analyses to check the reliability and robustness of the estimates and show that the results are not driven by the expansion of alternative early childhood programs during this time, selection bias, or other endogenous factors.
My analyses use a large, demographically, and socioeconomically diverse student population in Texas. Student-level data on student characteristics and achievement are provided by the Texas Education Agency (TEA) between 1994 and 1999. To construct a unique county-by-year data set on Head Start spending per age-eligible child for 1988–1994, I match grantee-level spending data from the Consolidated Federal Funds Reports (CFFR 2019a, 2019b) with administrative data that describe the serving counties for each grantee. Head Start spending years are restricted to 1988 and 1994, when the third-grade students would be age-eligible to attend Head Start. Combining student-level data with information on Head Start funding generosity, I estimate the effect of exposure to the program’s funding at age four on third-grade standardized test scores in math and reading (at ages nine and ten) for children born between 1984 and 1990.
My main finding is that exposure to Head Start funding expansions significantly improved academic performance in the third grade. For free/reduced lunch (FRL) certified students, a $500 increase in Head Start funding per child led to a 0.03 standard deviation (σ) increase in average test scores in math and reading combined.2 Using several sources on program characteristics and budgets, I show that additional Head Start funding led to significant increases in program participation rates, number of teachers hired, and full-time enrollment. Moreover, federal funding expansions are associated with more generous spending on education, nutrition, social services, and parent involvement in the Head Start program. This analysis suggests that program capacity and quality improvements are essential pathways for the ultimate effect on test scores.
Estimates by race and ethnicity show that improvements among Hispanic students are the primary driver of these results. Overall, among low-income students, the funding increase during the 1990s led to a 13 percent reduction in the test score gap between Hispanics and Whites in math and reading combined. These gains persist through the fifth grade. There are at least three testable channels through which Head Start could be beneficial for Hispanics. First, Hispanics have higher participation rates in Head Start in Texas compared to Whites and Blacks. I find that additional funding induced more Hispanic children to participate in the program. Second, the program participation could improve their language skills by exposing them to English at an early age, which could positively affect their academic performance. My analysis shows that Head Start funding exposure improved their language proficiency. Third, the program could help special needs children during childhood; Head Start centers must reserve 10 percent of their enrollment for children with disabilities. I show that additional funding reduced the likelihood of having special education status for Hispanics.
This work contributes to the economics literature in a number of ways. First, my analyses provide new evidence on the effects of Head Start on academic achievement by taking advantage of a different policy lever (spending, not participation), a different time period, and a unique population. Using spending increases in the 1990s in Texas, I find that Head Start expansions led to improvements in medium-run outcomes, especially for low-income Hispanic students.
Second, my findings align well with a growing number of papers that find public investments and education policies in early childhood or elementary school are more effective for Hispanics and children with low baseline scores (for example, Currie and Thomas 1999; Bitler, Hoynes, and Domina 2014; Gibbs 2014; Figlio and Ozek 2019). My study explores the use of public funds in early childhood for low-income children during the 1990s, which has been understudied.
Third, I contribute to the growing literature that examines the effect of public spending on academic success. Recent causal evidence shows that public spending during the formal school year improves student outcomes (Jackson 2018). What is less known is whether public spending during early childhood makes a difference in academic performance. My paper provides new evidence on this question in the context of Head Start.
In the following, I provide background information on the Head Start expansions and review the prior research in Section II. Section III describes data sources, followed by an overview of the methodology in Section IV. I report my results in Section Vand discuss potential mechanisms in Section VI. I then present robustness checks in Section VII and conclude in Section VIII.
II. Background and Prior Literature
A. The Head Start Program
Head Start is a federally funded early childhood education program that provides education, health, nutrition, and other services to economically disadvantaged children and their families. The program is designed to reduce disparities in school readiness, health, and other social aspects between low-income children and their more advantaged peers. The eligibility criteria are that children must be between ages three and five and come from families with income at or below the poverty level. Also, at least 10 percent of the children served in each center must have some type of disability, regardless of the income eligibility.
Head Start began as a part of the “War on Poverty” initiative in 1965. Between 1990 and 2000, enrollment increased by about 60 percent, and federal funding per enrolled child doubled (see Online Appendix Figure A.1). The 1990s expansion resulted from a significant effort by the Bush and Clinton administrations to improve capacity and quality constraints. Additional funding was used to increase enrollment, improve teacher salaries and training, expand services for families of children attending the program, and help local Head Start agencies purchase facilities. As a result of these policy efforts, there was a substantial ramp-up in federal funding per child, providing a natural experiment to study the program.
Head Start is a federal–local matching grant program. The federal government determines the program’s funding annually, as a component of the federal budget and allocates it to states based on the relative number of public assistance recipients, unemployed persons, and children from families below the poverty line (H.R.14449—93rd Congress 1973–1974). To receive funding, local agencies must submit their grant proposals directly to the Head Start Bureau.3 Grantees must provide at least 20 percent of the financing, which may include in-kind contributions through community partnerships. These grantees are heterogeneous in several dimensions, such as costs of personnel and space (depending on the geographic location, for example) and type of sponsoring agency (school system or private nonprofit) (Currie and Neidell 2007).4 As a result, there is geographic variation in funding levels, which provides part of the identification in this paper.
As stated by Currie and Neidell (2007), a local grantee can obtain additional funding in three main ways: (i) the federal government allocates more funding in a given year, (ii) program directors write more convincing grant proposals and attract more funds, or (iii) grantees attract more considerable local funds from the state or other local community agencies based on need or better connections. Because part of the funding variation stems from the grantees’ qualifications (and is not due to exogenous policy changes), careful consideration is needed to isolate exogenous variation to identify the program’s effectiveness. Section IV addresses some of these issues in detail.
B. Prior Literature
Head Start has been evaluated extensively over its existence.5 Prior research reports initial effects on cognitive outcomes that fade out when students reach the third grade. In the long run, Head Start is shown to be effective in improving socioeconomic and health outcomes.
The best available evidence on the short-term impacts of Head Start on children comes from the Head Start Impact Study (HSIS), an independent national random assignment study of Head Start that took place in the early 2000s and followed the children up to the third grade. The reports by Puma et al. (2005, 2010, 2012) show that (i) one year of Head Start improved cognitive skills by a small amount; (ii) by the end of the first grade, the effects mostly faded out; and (iii) on the third-grade follow-up, the impact of Head Start had disappeared. More recent papers, which address the flaws in the HSIS, find significant positive effects of Head Start for short-term test score impacts up to the first grade (Feller et al. 2016; Zhai, Brooks-Gunn, and Waldfogel 2014; Kline and Walters 2016). Additionally, Bitler, Hoynes, and Domina (2014) find larger short-term impacts at low quantiles of the test score distribution and persistent effects through the first grade for Spanish speakers at the bottom of the test score distribution. However, these recent papers do not provide evidence on test score effects in the third grade.
The best available evidence on the longer-term effects of Head Start comes from quasi-experimental studies. Previous literature has exploited within-family comparisons of siblings who have and have not participated in the program (for example, Currie and Thomas 1995, 1999; Garces, Thomas, and Currie 2002; Deming 2009; Bauer and Schanzenbach 2016). Other studies use discontinuities due to program funding and eligibility rules (Ludwig and Miller 2007; Carneiro and Ginja 2014). A few recent papers use variation in Head Start funding expansions in its early introduction (Thompson 2018; Barr and Gibbs 2022; Johnson and Jackson 2019; Bailey, Sun, and Timpe 2020). These papers show that Head Start has been effective in improving long-term socioeconomic and health outcomes.
Similar to Thompson (2018) and Johnson and Jackson (2019), who use Head Start funding expansions across counties over time in the 1960s, my study uses a new variation that stems from the 1990s program expansion to analyze the program’s impact on academic performance. Additionally, I use student-level administrative data in my study, which is advantageous to overcome statistical inference issues. Exploring Head Start’s impact on the low-income student population in Texas, which is predominantly Hispanic, I show that the program led to significant gains in academic achievement in third grade through fifth grade.
III. Data Construction and Summary Statistics
A. Data Construction
I combine several data sets on student test scores and demographics, Head Start spending, school quality, and economic conditions to analyze the effect of the program’s expansion on academic performance. This section provides an overview of the data sources; Online Appendix II includes detailed descriptions.
Administrative student-level data include the universe of exam takers in Texas from the third grade through the eighth grade between 1994 and 1999. The majority of the analysis focuses on the third-grade students. These data are from the Texas Education Agency (TEA) and include test scores monitored through the Texas Academic Assessment System (TAAS). Relevant to my paper, these data contain information on the year of birth, sex, ethnicity, FRL status, language proficiency, and special education status of each student at each school district. From the school district information, I assign the county of residence, which serves as a proxy for the child’s county of residence at age four. Along with the year of birth, this determines each student’s exposure to Head Start funding.
I conduct the following sample restrictions. Starting with the universe of students who took the TAAS test between 1994 and 1999, I first drop observations with missing demographic information, missing test scores, exempt testing status, or nonstandard test administration. Next, using the administrative data description, I restrict my analysis to students who are certified for FRL or are identified as economically disadvantaged based on their families’ welfare eligibility because they are more likely to be eligible for Head Start. From here, I will refer to this sample as FRL certified.
I develop a measure of Head Start funding per age-eligible child in a county using several sources. Head Start spending data are from the CFFR, which include information on local appropriations for federally funded programs for each grantee. A Head Start grantee could oversee one county or a group of counties. This factor is considered when constructing the county-level spending by incorporating information from the Head Start budgets (PCCOST) (U.S. Department of Health and Human Services (HHS), Administration for Children and Families, Head Start Bureau 1999). Head Start spending years are restricted to 1988 and 1994, when the third-grade students would be age-eligible to attend Head Start. To construct the population denominator of children ages three and four years old,6 county-level population counts for each age group are extracted from the Surveillance, Epidemiology, and End Results Program (SEER) between 1988 and 1994 (National Cancer Institute 2017). Together with these three data sources, I construct the Head Start funding per child variable at the county–year level.
I augment these data with additional data sources to bring information on (i) Head Start enrollment and program budget items, (ii) school quality and other pre-kindergarten (pre-K) alternatives in Texas, and (iii) economic conditions. First, I compiled data from the Program Information Reports (PIR) between 1988 and 1994 (Office of Head Start 2019). Starting in 1988, the Office of Head Start has collected comprehensive data from all grantees and delegates on the services, staff, children, and families served by the program. These data provide information on the number of funded enrollment, staff, demographic composition of children and staff, and director qualifications. Second, I source the school-level pupil-to-teacher ratio between 1994 and 1999 and information on school-level state-funded pre-K (outside of Head Start) enrollment from the Common Core of Data (CCD) (U.S. Department of Education, National Center for Education Statistics 2019) between 1984 and 1990.7
Next, I compiled data on county-level economic conditions as controls that are assigned at the time of birth. These include per capita income, per capita transfer payments for cash income support, medical benefits, food assistance, retirement, and disability programs from the Regional Economic Information System (REIS) (U.S. Department of Commerce, Bureau of Economic Analysis 2019) between 1984 and 1990. I also control for percentage urban population, percentage Black, percentage Hispanic, percentage single parent, percentage of 0- to 18-year-olds living in poverty from the 1983 City and County Data Book (United States, Bureau of the Census 1983), and unemployment rate from the Bureau of Labor Statistics.
B. Summary Statistics
My analysis centers on children exposed to Head Start funding in Texas between 1988 and 1994. Figure 1 presents the average Head Start spending per child in the 15 most populous counties in Texas between 1980 and 1995.8 This figure shows substantial variation in Head Start spending per child across counties and within a county over time. For more insight into the geographic variation in Head Start funding, Figure 2 presents maps of (i) funding levels in 1988 and (ii) the growth from 1988 to 1994, respectively.9 These maps show a great deal of variation across counties in both levels and growth in funding. My basic identification strategy uses this geographic and time variation to identify the effect of Head Start on academic achievement.
Head Start Funding per Child in the 15 Most Populous Counties in Texas
Notes: Head Start spending (in 2014$) data are from the CFFR, coupled with the population counts for three- and four-year-olds at the county-level from the SEER. For more details about data construction, see Section III. The 15 most populous counties include Bend, Bexar, Brazoria, Cameron, Collin, Dallas, Denton, El Paso, Fort, Harris, Hidalgo, Montgomery, Nueces, Tarrant, Travis, and Williamson. Vertical lines (1988–1994) indicate the period of this study.
Geographic Variation in Head Start Funding per Child
Notes: Head Start spending (in 2014$) data are from the CFFR, coupled with the population counts for three-and four-year-olds at the county-level from the SEER. Growth measure is calculated using 1988 as the base period, and the shades in the map on the right are determined based on the terciles of the growth distribution. For more details about data construction, see Section III.
Table 1 presents summary statistics for third-grade students, first for the full sample, economically advantaged students (FRL noncertified) and economically disadvantaged students (FRL certified). The first three columns show a large discrepancy in average test scores by economic status.10 Economically advantaged students have substantially higher test scores relative to FRL-certified students. Importantly, Hispanics make up more than 50 percent of these FRL-certified students. On average, they have the lowest test scores relative to other economically disadvantaged students and tend to live in counties with higher funding for Head Start per child.
Sample Characteristics of Third-Grade Students in Texas
Figure 3 shows the evolution of standardized test scores across birth cohorts by their FRL status. The most striking observation is a considerable achievement gap among racial and ethnic groups, even among FRL-noncertified students. For students who were noncertified (Figure 3A) and certified (Figure 3B) for FRL, the academic performance of Hispanic and Black students seems to be improving across cohorts. Overall, these observations suggest there is value in analyzing the potential effects of early childhood investments separately by race and ethnicity.
Average Standardized Test Scores across Birth Cohorts, by Free or Reduced Lunch Status
Notes: The bars indicate average test scores for Black, Hispanic, and White students separately. Test scores are math and reading composite scores, standardized using all third-grade students who took the standardized test in Texas between 1994 and 1999, from the TEA. The sample is divided into two groups: Panel A includes students who are not identified as economically disadvantaged (FRL noncertified) and Panel B includes students who are FRL certified or who are identified as economically disadvantaged based on their families’ welfare eligibility.
To demonstrate the relationship between average test scores and Head Start funding, I restrict the sample to FRL-certified children and collapse the data on test scores to county-level averages. Online Appendix Figure A.3 shows a positive correlation between average test scores and Head Start spending per child for FRL-certified children.
IV. Empirical Strategy
To study Head Start’s effect on academic performance, I exploit variation in county-level Head Start funding per child in the first half of the 1990s. Following Ludwig and Miller (2007) and Sanders (2012), I assign Head Start funding exposure based on each student’s county of residence and birth year.11 My empirical strategy uses a county and birth year panel fixed-effects approach, which relies on variation in Head Start funding per child within counties and over time, conditional on observables. Formally, I estimate the following equation for the sample of FRL-certified students using a newly assembled data set on county-level Head Start spending per child:
1
where Yiscbt denotes the outcome variable (standardized test scores) for student i in school s in county c in birth year b, and in test year t. HSfundingc(b+4) represents Head Start funding per child in county c when student i was four years old ( = b + 4). Xiscbt is a vector of individual-level demographic controls, including sex, ethnicity, and an indicator for bilingual and English as a second language. Zct has three sets of county-level controls, including per capita income transfers, average characteristics of students collapsed from the main data, and additional controls, which include share of zero- to five-year-old population by ethnicity, log population, and unemployment rate. Wc(b+4) includes county-level controls assigned at age four, such as income per capita, share of pre-K enrollment, and income transfers per capita. Finally, θs, ξb, ηt are school, birth year, and test year fixed effects, respectively. πc * b is a county-specific linear trend. Standard errors are clustered at the county level. The coefficient of interest is β, which is interpreted as the conditional change in the outcome variable from a unit ($500) increase in exposure to federal Head Start funding per child at age four.
For this research design to be valid, funding expansions must be exogenous to other underlying geographic-level trends in test scores. Threats to identification are any differential trends across counties correlated with spending changes, which may also influence student outcomes. I use several methods to probe the validity of the key identification assumption.
First, I take county-level characteristics measured in 1980, before the expansions occurred, and use them to predict the levels and changes in Head Start spending per child from 1988 to 1994 (similar to Hoynes and Schanzenbach 2012). For this analysis, I collapse the data at the county level. The independent variables include the percentage of the 1980 population living in an urban area, Black, Hispanic, single parent, less than age five, age 65 or older, total population, percentage of 0- to 18-year-olds living in poverty, as well as income, education, and welfare spending per capita (in 2014 dollars).
The results are presented in Online Appendix Tables A.2 and A.3.12 Simple correlations imply that counties with a larger Hispanic population, a more extensive farmland, and a higher share of single mothers, poor, very young, or elderly have more Head Start funding. In contrast, counties with a larger Black population tend to have less funding, which could be because Blacks in Texas live in more urban areas with high population density. Additionally, counties with more social spending have higher funding for Head Start, and counties with higher per capita income have lower funding. The determinants of the change in Head Start funding from 1988 to 1994 are also similar (see Online Appendix Table A.3).
After controlling for the characteristics described above, the last columns of Online Appendix Tables A.2 and A.3 show that the percentage of Blacks and percentage of 0- to 18-year- old population living under poverty are significant determinants of Head Start expansions, and together all variables explain around 20 percent of the overall variation (R-squared ≈ 0.20). Nevertheless, I include county-specific time trends in my analyses to control for possible differences in trends across counties. The results are robust to the exclusion of these trends.
Second, director quality could be a possible confounding factor if directors who can obtain more funds may also run better programs in other aspects. For example, a bias will occur if better-qualified directors write better grants to receive additional funds and then operate higher-quality programs. In this case, funding levels could be correlated with child outcomes (Frisvold 2006; Currie and Neidell 2007). To rule out this possibility, I show in Online Appendix Table A.4 that federal funding increases are not significantly associated with directors’ qualifications.13 A bias could also occur if some communities devote local resources toward children and if these resources during childhood also help children succeed in their school years. Addition of school fixed effects controls for this possibility, as well as for other neighborhood characteristics that could be correlated with student success.
Next, Head Start may have been introduced or expanded with other local policies that affect children’s outcomes, such as other War on Poverty programs. For instance, the timing of Head Start’s introduction corresponds to the introduction or expansion of other government programs, including Medicaid, Medicare, Food Stamps, and the Supplemental Nutrition Program for Women, Infants, and Children (WIC). To address the concerns regarding contemporaneous policy changes that targeted four-year-old children, I directly control for county-level spending for other social programs.14 Moreover, including county-specific linear time trends also accounts for the fact that some communities may improve over time.
Finally, critical to my identification strategy is that Head Start spending expansions should not systematically change a particular cohort’s composition within a school. For example, if low-income families decided to move based on Head Start’s funding generosity and more generous communities had better schools, such nonrandom selection would misattribute higher student performance to Head Start exposure. In Online Appendix Table A.1, I formally test this and other types of self-selection by examining whether student characteristics such as sex, race/ethnicity, and county-level income per capita are correlated with funding after conditioning on school fixed effects. I show that Head Start funding variation does not predict changes in the composition within a school, suggesting the estimates are not biased by families’ self-selection into a particular cohort within a school.
A potential threat to my identification strategy is the existence of state-provided preschool alternatives to Head Start, as all preschools provide similar services to improve children’s school readiness. Texas has offered a half-day public pre-K program since the 1985–1986 academic year. State-provided pre-K aims to enhance children’s academic performance by providing early childhood education for four-year-old children identified as at risk.15 Although it is mandatory for any district with at least 15 eligible children to offer a half-day education-based program for four-year-old children, attendance is voluntary (Texas Education Code §29.1532).16 Thus, starting in 1985, an eligible child in Texas could attend a public pre-K as an alternative to Head Start. Although funding for pre-K is allocated directly to school districts by the state of Texas, districts are encouraged to partner with licensed childcare centers and Head Start programs to provide preschool services (Barnett et al. 2011).17 This raises the possibility that Head Start and pre-K might have operated jointly to some degree during the 1990s.
Online Appendix Figure A.5 plots the number and share of children enrolled in Head Start and pre-K in Texas between 1988 and 1994. Relative to Head Start, which had a 6 percent share of age-eligible enrollment in 1988, pre-K was more extensive, with a 14 percent share. This figure also shows that the timing of the expansions of both programs coincide. This is a concern, as the existence of a large-scale pre-K is a potential threat to identification as a confounding factor. Hence, it is essential to control for the availability of other preschool alternatives, as failing to control for it could misattribute the effects of other preschools to Head Start.
To diagnose and address this concern, I directly analyze the relationship between Head Start spending and pre-K expansions between 1988 and 1994. Table 2, Column 5 shows that, after controlling for county and year fixed effects, there is no significant relationship between a share of age- and income-eligible children enrolled in pre-K and Head Start funding per child. If anything, this analysis suggests that Head Start program expansions may have crowded out the enrollment for pre-K for low-income children in Texas.
Effect of Head Start Funding per Child on Head Start and Pre-K Enrollment
Although this analysis confirms that the variation in Head Start per child does not significantly predict pre-K enrollment, there is still a possibility that, in terms of facility usage and program operations, pre-K and Head Start cooperated in the 1990s.18 To take this possibility into account, I control for the share of age-eligible children enrolled in pre-K in a given county during the time of exposure to Head Start and show that the results are not sensitive to adding this control.
V. Results
Having documented the plausible exogeneity of the Head Start funding variation, I now present the results of Equation 1. As noted above, the main sample is restricted to FRL-certified students, and Head Start exposure is assigned at the time when a child was four years old.19 To simplify the interpretation of the coefficient of interest, Head Start spending per child is scaled by $500, the equivalence of mean spending over the study’s period. Thus, the coefficients should be interpreted as the effect of exposure to a $500 increase in Head Start spending per child. All monetary values are converted to 2014 dollars, and test scores are standardized using the entire student population.
A. Federal Head Start Funding and Test Scores
Panel A of Table 3 shows the effect of Head Start spending per child on combined test scores in math and reading. Column 1 reports the results for the main sample. The coefficients in Panel A indicate that a $500 increase in Head Start spending per child leads to a statistically significant 0.031 standard deviation (σ) increase in third-grade test scores.20
Baseline Estimates of the Effect of Head Start Funding Exposure on Student Outcomes in Third Grade
Previous studies show that the returns to public education investments are higher for groups at the lower end of the skill distribution (for example, Bitler, Hoynes, and Domina 2014). Given that male students tend to be lower-achieving relative to females, Head Start may yield larger returns for males than females. In the next two columns of Table 3, I analyze whether exposure to more generous Head Start funding improves third-grade test scores differentially by sex. In Columns 2 and 3, additional Head Start funding exposure is associated with improvements in test scores for males and females, with slightly larger point estimates for males. However, the estimated coefficients are not statistically significant.
In the rest of Table 3, I present the estimates for non-Hispanic Whites, Blacks, and Hispanics, respectively. The race and ethnicity breakdown reveals that improvements for Hispanics are the main driver of the results. In particular, I find that a $500 increase in Head Start spending per child leads to a 0.051 σ increase in test scores in math and reading (p < 0.01). These results suggest that around a $500 increase in federal Head Start funding exposure closes more than 13 percent of the gap relative to the raw mean difference in test scores ( = 0.051/0.38).21 This estimate is similar to Currie and Thomas (1999), who find that Head Start participation closes at least 25 percent of the test score gap between Hispanics and Whites.
Although the results for Hispanics are positive, large, and statistically significant, the analogous estimates for Whites suggest that these cohorts did not experience improvements in test scores with exposure to additional Head Start funding. Also, the effects for Blacks are positive and economically significant but statistically imprecise.22
As noted, Hispanic students lag behind both Black and White students academically. In contrast to Black children, who are historically underprivileged compared to Whites, Hispanics often live in immigrant, Spanish-speaking families and communities (Currie and Thomas 1999). During my study period, Hispanics comprise more than 50 percent of FRL-certified third-grade students in Texas, and 37 percent have limited language proficiency. Table 1 reports that Hispanics were living in areas with more generous Head Start funding compared to Whites and Blacks. For example, average Head Start funding exposure for Hispanics was $687, with a standard deviation of 1178; by contrast, funding exposure for Blacks was approximately $300, with a standard deviation of 211. Additionally, counties with a high fraction of Hispanics experienced more funding expansions relative to communities with a high fraction of Blacks, as shown in Online Appendix Table A.3.
There are at least four channels by which Head Start could be beneficial for Hispanics. First, considering that Hispanics have higher participation in Head Start in Texas (11.4 percent Hispanic enrollment as a share of all poor children) compared to Whites (5 percent) and Blacks (4.7 percent), additional funding may induce more Hispanic children into the program. Table 2 presents the enrollment effects of Head Start funding expansions among Whites, Blacks, and Hispanics separately. Indeed, the enrollment effects of additional Head Start funding are larger for Hispanics (6.5 percentage point increase, p < 0.01) relative to Whites (2.8 percentage point increase, p < 0.01) and Blacks (1.5 percentage point increase, p < 0.01).
Second, Head Start could increase their exposure to English and develop their language skills early, which could influence their educational performance in both reading and math. Panel B of Table 3 reports that additional Head Start funding increases the likelihood of their language proficiency (0.01 percentage point increase, p < 0.05). One might expect that improved language proficiency would improve reading skills more than math skills. Given that the results on math scores (0.070σ) are larger than reading (0.037σ), this channel alone does not explain the result’s overall pattern (see Online Appendix Table A.5).
Third, Head Start could help special needs children during early childhood by providing services to identify and potentially treat some types of special education placements, such as behavioral disorders or learning disabilities. This channel is plausible because the program has been required by federal law to reserve 10 percent of its enrollment for children with disabilities. Panel C of Table 3 shows that for Hispanic students, additional funding reduces the likelihood of having special education status in the third grade (0.01 percentage point reduction, p < 0.1).23
Finally, although not statistically testable, Head Start could promote cultural assimilation for Hispanics that may help children adapt to school more easily (Currie and Thomas 1999; Bitler, Hoynes, and Domina 2014).
B. Do Test Score Effects Persist over Time?
Test score fade-out has been at the heart of the Head Start policy debate. As evidence against the program’s effectiveness, critics point out that test score gains at the time of school entry dissipate quickly. In this section, I provide suggestive evidence on the effect of the Head Start program on test scores from the third through eighth grade. The main limitation here is that this analysis relies on cross-sectional data because I observe students in each grade who took standardized tests in Texas between 1994 and 1999.24 This limits the scope of the findings, mainly because I am unable to distinguish between the nonlinearity of the effects in test scores over the years (the impact in the third grade is expected to decrease as students age) and the reduction in exposure to funding generosity across the years (older cohorts experience less exposure to Head Start relative to younger cohorts).
With this limitation in mind, Figure 4 presents estimated coefficients and their 95 percent confidence intervals from separate regressions with the standardized test scores combined, in math and reading for each grade as the outcome variables, and exposure to Head Start funding as the right-hand-side variable.25 Panel A of Figure 4 shows the evolution of the results when the sample includes all FRL-certified students, and Panel B shows the estimated coefficients when the sample is restricted to Hispanic students who are FRL certified.
Effect of Head Start Exposure on Test Scores for Each Grade
Notes: This figure plots coefficients and their 95 percent confidence intervals obtained when the dependent variable is standardized test scores combined, in math, and in reading separately for each grade and the independent variable is real federal Head Start spending per child (in 2014$) when the child was four years old. Head Start spending per child is scaled by $500 (average spending during the period of study); thus, the coefficients should be interpreted as the effect of a $500 increase in funding. All regressions include controls for demographics and county-level characteristics, school, test year, and birth year fixed effects, along with county-specific linear trends. The sample consists of students from third through eighth grade who are FRL certified or who are identified as economically disadvantaged based on the description by the TEA. Panel A reports the results for the sample of all FRL-certified students, and Panel B reports the results for the sample of Hispanic students who are FRL certified. Student data are from the TEA and include information on year of birth, ethnicity, economic disadvantage indicators, and test scores conducted between 1994 and 1999. Head Start spending data are from the CFFR and include years 1983–1994. Standard errors are clustered at the county level.
Overall, the results suggest that the effects on test scores appear to be positive throughout, but the effect sizes decline after the fourth grade. Although not conclusive, this pattern in Panel A is consistent with the results of previous literature (Currie and Thomas 1995; Krueger and Whitmore 2001; Deming 2009).
In Panel B, for Hispanic students, the group with the highest returns, there is suggestive evidence of more persistence of test score gains in comparison to Panel A, especially in reading. The estimated effect declines from 0.051σ in the third grade to 0.036σ in the fifth grade in combined test scores. Although the point estimates in combined test scores are smaller, they are still sizable and statistically significant in the fifth grade. In reading test scores, the effect size remains around 0.04σ through sixth grade. There could be two explanations for this pattern. As discussed, there are several reasons why Hispanics benefit from early childhood investments, which may explain more persistent effects on test scores among this group. Prior literature also shows supporting evidence that early childhood investments are more effective for Hispanics (Gibbs 2014). Finally, the results for Hispanics may take a longer time to fade out because the initial impact size for this group is larger.
In sum, this analysis suggests that for Hispanics, Head Start investments improved test scores from the third through fifth grade, reading test scores persist through the sixth grade, and the fade-out pattern takes a longer time to occur.
C. Interpretation and Magnitude of Estimates
The estimates reported in Table 3 represent intention-to-treat (ITT) effects, which can be interpreted as the average effect of Head Start funding exposure on economically disadvantaged children. My main sample consists of FRL-certified students. These students would have been eligible for Head Start as children if their families’ socioeconomic status had stayed approximately the same from early childhood to age nine.26
To make accurate comparisons with the literature, I attempt to convert the ITT estimate of Head Start funding on test scores to the treatment effect on the treated (TOT). This conversion requires having a “first stage” that provides an estimate of exposure to Head Start funding on participating in Head Start for the sample. Unfortunately, Head Start participation is not observed in the main data set. As a proxy, I use the estimated effect from Column 1 of Table 2, which reports a $500 increase in funding is associated with an 11.2 percentage point increase in Head Start enrollment at the county level for low-income children in Texas.27
Scaling up the test score impact of a 0.031σ by 0.112, the implied effect of Head Start enrollment at the county level corresponds to a 0.28σ increase in test scores. While imperfect, this estimate provides a comparison point with the literature. However, this should not be interpreted as the “true” effect of participating in Head Start, given the issues mentioned above.
Comparing the test score impact of Head Start from my paper with Deming (2009) and the HSIS (Puma et al. 2012), my study reports a larger and statistically significant impact on third-grade test scores. My analysis and the other two papers differ in terms of the study period, the sample, and the identification strategy, potentially explaining the larger findings here. While Deming (2009) studies the Head Start program in the 1980s (pre-1990), and Puma et al. (2012) examine the random assignment in the 2000s, my work focuses on the 1990s, during which Head Start evolved significantly in terms of capacity and quality. Moreover, the other two papers analyze nationally representative samples, whereas my sample consists of low-income children in Texas who are predominantly Hispanic. Last, my identification strategy differs from the other two papers by taking advantage of the funding expansions across counties and over time.
VI. Discussion of Mechanisms
One channel by which Head Start funding increases could improve test scores is by serving more children. As expressed above, Table 2 reports the results that show that additional funding led to significant increases in Head Start enrollment for economically disadvantaged children in Texas.
A second potential channel is improvements in existing program characteristics that could lead to better academic outcomes. Prior research has shown that reductions in student-to-teacher ratios benefit students, particularly children from disadvantaged backgrounds (Krueger and Whitmore 2001). To my knowledge, there is only one paper that examines the effect of program inputs on cognitive and noncognitive skills in the Head Start literature. Examining the impact of different inputs in Head Start centers using the HSIS, Walters (2015) finds that teacher education, teacher certification, and class size are not associated with test scores improvements. He shows that the key input that improves children’s cognitive skills is the provision of full-time services at the center level.
To explore this, I examine the relationship between federal funding increases and program inputs, such as the number of teachers per child, enrollment per child, full-time enrollment per child, child-to-teacher ratio, and director’ssalary,28 as well as per child spending for education, health, nutrition, social services, and parent involvement in the programs. For this analysis, I employ data on program characteristics from the PIRs (available for 1988–1994), director’s salary from the PIRs (available for 1992–1994), and program budgets for various types of spending available in an administrative data set on Head Start budgets called PCCOST (available for 1992–1994 for some programs).
Table 4, Panel A reports the findings on program inputs. The first three columns show that a $500 increase in Head Start funding per child is associated with statistically significant increases in the number of teachers per child, enrollment per child, and full-time enrollment per child. This finding is consistent with Walters (2015). In Column 4, the coefficient on the child-to-teacher ratio is negative as expected; however, it is statistically insignificant. Finally, the results in Column 5 suggest that part of the increase in funding translated into increases in directors’ salaries.
Mechanisms: The Relationship between Head Start Funding and Head Start Program Characteristics and Budget
Panel B of Table 4 shows that a $500 increase in Head Start funding per child is associated with increases in spending on education, health, nutrition, social services, and parent involvement. Given that budget data are only available for a subset of programs, these results are suggestive at best. Overall, there is suggestive evidence that, on the margin, federal funding partially went to spending for services that might improve education and health development for children.
Since the 1990s, Head Start has aimed to increase teacher qualifications by setting aside funds to improve teachers’ education. Although not testable, quality improvements may have been partly driven by the advances in teacher qualifications.29
VII. Robustness
In this section, I conduct various robustness exercises to address potential concerns with the estimation strategy. Table 5 presents multiple sensitivity checks to the main specification. As a point of comparison, I include the baseline estimates in Column 1 that show the effect of Head Start exposure on third-grade test scores. In Column 2, I omit county-specific linear trends, which leads to a small increase in the effect size. Column 3 presents results with no pre-K controls, which slightly decreases the magnitude of the effect.
Effect of Head Start Exposure on Third-Grade Standardized Test Scores Sensitivity of Results to Alternative Specifications
In Column 4, I exclude the controls for average county-level income per capita at the time of birth, at age four, and test year. The estimated coefficient remains significant with this restriction, but the magnitude decreases slightly. In the next column, I omit controls for the county-level measures of per capita transfer payments for cash income support, Food Stamps, medical care, retirement, and disability programs and find that the estimate’s magnitude decreases without these controls. This is expected because Head Start funding expansions are positively correlated with the generosity of transfer payments. Excluding these critical controls biases the main effect downward.
In the last column, I include school-specific linear trends to control for possible improvements in students’ neighborhoods. Adding these trends leads to smaller coefficients with similar standard errors, which means that school trends may be correlated with Head Start funding changes. Because the overall effect is relatively sensitive to adding school trends, I replicate the main results by adding school trends. Online Appendix Table A.7 presents these estimates, which show that the effect for Hispanics remains statistically significant, though the effect size is slightly smaller.
Next, I analyze whether the results are sensitive to two additional Head Start exposure measures constructed using alternative population denominators and find that the results are consistent across all three measures. Column 1 of Online Appendix Table A.8 reports the estimates using my preferred measure, Head Start per child, and in the next two columns, I present results obtained with Head Start per capita (Column 2) and Head Start per poor child (Column 3).30 Regardless of the differences in the measures, this table shows that an increase in Head Start funding, on average, leads to similar test score gains.31
As a falsification exercise, I analyze the effect of exposure to Head Start on test scores for the sample of children who are not FRL certified or living in poverty; hence, they are not likely to benefit from Head Start. I present these results in Online Appendix Table A.10 and show that Head Start exposure does not affect test scores for this group.
To check for potential selection effects, I test whether increases in federal Head Start spending affect the composition of a particular grade within a school (similar to Carrell and Hoekstra 2010). If the variation in Head Start spending is not correlated with selection into the sample, I would expect to find no correlation. Online Appendix Table A.1 presents regression results obtained by regressing exogenous student characteristics on Head Start exposure conditional on birth year, test year, and school fixed effects, as well as other controls included in the main specification. The results suggest no evidence that a particular cohort’s composition (probability of being a specific race or sex, likelihood of being FRL certified, and county’s income composition) is correlated with Head Start funding. In the last column, I test whether Head Start funding increases are associated with the “predicted test scores” using students’ observable characteristics and find no association between them. Overall, these tests suggest that the main results are not driven by selection bias.
Because Head Start serves children between the ages of three and five, one would not expect exposure to Head Start that occurred during other ages to be associated with improvements in outcomes (Thompson 2018). As a falsification test, I estimate models where I analyze Head Start exposure at different ages.32 Table 6 reports results from specifications that use the sample of third-graders with the assignment of exposure to Head Start changing from age three through age eight as the right-hand-side variable.33 These findings suggest that Head Start exposure at ages three and four matters the most, with the largest impact on test scores coming from exposure at age four. This is expected, considering that four-year-old children make up around 50 percent of the total children served in Head Start (Office of Head Start 2019).
Falsification Test: The Effect of Head Start Funding Exposure on Third-Grade Combined Standardized Test Scores—Differential Effects by Age of Exposure, 3—8
In sum, this section presents evidence that the results are robust to excluding county-specific linear trends, do not appear to be the result of the expansion of the alternative early childhood programs during this time period, and are not driven by selection bias. The results also show that Head Start funding exposure at age four leads to the largest improvements in third-grade test scores, and exposure at ages older than age five does not affect test scores.
VIII. Conclusion
This paper provides new evidence on the old debate of whether early childhood investments for economically disadvantaged children narrow the academic achievement gap in elementary school. Many early childhood interventions are effective in improving children’s skill development initially, but the effects disappear when students reach the third grade (Bailey et al. 2017). Using new variation in the federal funding expansions for the Head Start program in the 1990s and student-level administrative data from Texas, I find that exposure to more generous Head Start funding during childhood led to substantial improvements in test scores.
The gains of low-income Hispanic students drive the overall improvements in test scores. In particular, a $500 increase in Head Start funding per child led to a 13 percent reduction in the test score gap between Hispanics and Whites in math and reading combined. Hispanics benefited from funding expansions through increased access to Head Start and improvements in program inputs. These advances enhanced their language proficiency and reduced their likelihood of special education needs during elementary school.
Head Start has served low-income children for more than 50 years to reduce education and health disparities across socioeconomic groups. During the 1990s, the federal government expanded Head Start funding to improve its capacity and quality. The substantial national funding increase within a relatively short time created a natural experiment that resulted in a large variation of the adoption of funding expansions across counties and over time. Combining several sources on program characteristics and budgets, my findings provide new evidence on how public funds are spent and suggest that both program capacity and quality improvements are essential pathways for the ultimate effects on test scores.
Early childhood investments have been receiving significant political attention. Therefore, a comprehensive understanding of the potential benefits and costs of such investments is essential for informed policymaking. My findings indicate that Head Start passes the cost-benefit test by a wide margin and provides valuable insight for policymakers considering future public investments in early childhood education.
Footnotes
The author thanks Nicholas J. Sanders, Elira Kuka, Na’ama Shenhav, Hilary Hoynes, Chris Walters, Chris Herbst, Gaetano Basso, Cassie Hart, Vasco Yasenov, Michel Grosz, Ariel Pihl, and brown bag participants at UC Davis for providing input. She is also grateful to Matthew Neidell and David Frisvold for generously providing data. An earlier version of this paper circulated under the title “Local Area Spending Exposure to Head Start and Academic Performance: Evidence from Texas.” All errors are those of the author. Student-level data are from the Texas Education Agency. These data are currently administered by the University of Texas Education Research Centers in Austin and Dallas and are available for use to approved researchers.
Supplementary materials are available online at: https://jhr.uwpress.org.
↵1. The following studies show positive long-term effects of Head Start: Currie and Thomas (1995, 1999); Garces, Thomas, and Currie (2002); Deming (2009); Bauer and Schanzenbach (2016); Thompson (2018); Bailey, Sun, and Timpe (2020); Barr and Gibbs (2022); Johnson and Jackson (2019).
↵2. This intention-to-treat (ITT) estimate of a 0.03σ corresponds to a treatment effect on the treated (TOT) of an approximately 0.3σ increase in test scores.
↵3. Grants are issued through a competitive process, with priority given to agencies that can demonstrate the most cost-effective operation, and existing programs have priority when reapplying (Currie and Neidell 2007).
↵4. However, each center must comply with publicly known standards, which are described in the Head Start Act §641a, 42 USC §9801, Pub Law No. 110-134 (2007).
↵5. Several studies have reviewed the literature on Head Start’s effectiveness (Barnett 1995; Currie 2001; Barnett and Hustedt 2005; Ludwig and Phillips 2008; Shager et al. 2013; Duncan and Magnuson 2013; Gibbs, Ludwig, and Miller 2013).
↵6. Three- and four-year-old children comprise approximately 90 percent of Head Start enrollment (U.S. Administration for Children Youth and Families 1990).
↵7. One caveat of using the CCD is that it only reports the enrollment in public pre-K programs that take place in public elementary schools, excluding public pre-K programs that take place in private preschools or daycare. Therefore, it provides an incomplete picture of the preschool scene.
↵8. There are 254 counties in Texas. I chose the 15 biggest counties for clear visualization of the variation. These counties make up around 60 percent of the student population in Texas.
↵9. The growth measure is calculated using 1988 as the base period, and the shades are determined based on the terciles of the growth distribution.
↵10. The distribution of student test scores in math and in reading by FRL status is shown in Online Appendix Figure A.2.
↵11. Because individuals are not linked across the years, I rely on the county of residence at each grade to assign treatment exposure. If low-income families make migration decisions based on the availability of the services provided by Head Start, the assignment using county of residence would lead to bias in my estimates. I test this directly in Online Appendix Table A.1 and find no evidence that the Head Start spending affects the student composition within school over time.
↵12. See Online Appendix Figure A.4 for a visual presentation of the correlations.
↵13. One caveat to this analysis is that data on directors’ characteristics are available from the PIRs starting in 1992; therefore, the sample is restricted to the years 1992–1994.
↵14. See Section III for a detailed discussion about which controls are added.
↵15. The at-risk population includes the following: children unable to speak and comprehend the English language; children certified for the FRL program; children who are homeless as defined by federal law; a child whose parents are either on active military duty, in an activated reserve unit, or who were killed or wounded while serving on active duty; and children in the Texas foster care system (Texas Education Code §29.1531).
↵16. Andrews, Jargowsky, and Kuhne (2012) evaluate this program and show it has been effective at improving math and reading test scores, reducing the likelihood of being retained in a grade, and decreasing the probability of receiving special education services.
↵17. State funds eligible children through the Foundation School Program based on average daily attendance (Texas Education Agency 2014).
↵18. Starting in 2003, the state law requires the new pre-K establishments to coordinate and cooperate with Head Start (Texas Education Code §29.1533; Texas Education Code §29.158).
↵19. Additional results are presented with Head Start exposure at ages three through eight in Section VII.
↵20. I present the results on combined test scores in math and reading for simplicity. To provide additional insight, in Online Appendix Table A.5, I examine the effect of Head Start in math and reading test scores separately and show that the estimated coefficients are larger and more statistically precise for math scores (a 0.045σ increase, p < 0.05) relative to reading scores (a 0.023σ increase, p < 0.1).
↵21. Using Table 3, the average standardized test score for Hispanics is −0.449 and −0.069 for Whites. The raw difference is 0.38.
↵22. It is worth noting that the magnitude of the estimated effect on Blacks is not negligible. Considering the enrollment increases for Blacks, the estimated TOT is larger than the TOT for Hispanics.
↵23. Due to the lack of detailed data, I am unable to explore the effect of Head Start separately across categories of disability. With the detailed data from North Carolina (NC), Muschkin, Ladd, and Dodge (2015) examine the effects of two early childhood education (ECE) initiatives, which occurred in the 1990s and early 2000s in NC, on the probability that children are placed into special education by the end of the third grade (the outcome I explore in my paper) and are placed in specific disability categories. Taking advantage of the rollout of two ECE programs’ introductions across counties and over time, they show that the programs’ introduction and expansion reduced special education replacement in the third grade (similar to my findings). They also find that one of the two programs reduced placements for educable handicaps, specific learning disabilities, and other health impairments, providing some evidence that supports the main argument.
↵24. Unfortunately, I do not have access to panel data that would allow me to follow the same students across years.
↵25. Online Appendix Table A.6 presents these results in table form.
↵26. However, eligibility does not imply participation. Head Start only serves a proportion of the eligible children in a given year due to capacity constraints. There is also an issue with incomplete take-up, meaning that not all eligible children actually enroll in the program.
↵27. This exercise requires a strong assumption that funding increases translated into “only” enrollment expansion. However, in the 1990s, additional funding was appropriated toward quality improvements. As a result, TOT here could be interpreted as an upper bound of the effect of Head Start participation.
↵28. Online Appendix Figure A.6 plots the average of some of these inputs between 1988 and 1995.
↵29. PIR data have information on the number of teachers with AA or BA degrees starting in 1999.
↵30. To construct age-specific poor population counts at the county-year level, I use the Small Area Income and Poverty Estimates (SAIPE) of the U.S. Census Bureau (1989, 1993, 1995, 1997, 1998, and 1999). These data report estimated counts for the number of children aged 0–17 and children aged 5–17. Taking the difference between these two variables, I construct the number of children younger than age five. The number of poor children age three or four is two-fifth of the number of children younger than age five (Frisvold 2006). For more details about the data construction, see Online Appendix II.
↵31. In Online Appendix Table A.9, I show the results using test scores by sex and by ethnicity using Head Start per capita in Panel A and Head Start per poor child in Panel B. These results closely follow the patterns of the estimates in Table 3.
↵32. Here exposure to Head Start is assigned based on ages from three through eight separately. For example, for a child who is born in 1985, their age of exposure as a three-year-old would be in 1988, a four-year-old in 1989, and a five-year-old in 1990.
↵33. Online Appendix Figure A.7 visually represents the coefficients for each age of exposure with 95 percent confidence intervals.
- Received April 2019.
- Accepted July 2021.
References
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.