Does Head Start Make a Difference?

The impact of participation in Head Start is investigated using a national sample of children. Comparisons are drawn between siblings to control for selection. Head Start is associated with large and significant gains in test scores among both whites and African-Americans. However, among African-Americans, these gains are quickly lost. Head Start significantly reduces the probability that a white child will repeat a grade, but it has no effect on grade repetition among African-American children. Both whites and African-Americans who attend Head Start, or other preschools, gain greater access to preventive health services. (JEL I38, H43)

Head Start is a federal matching grant program that aims to improve the learning skills, social skills, and health status of poor children so that they can begin schooling on an equal footing with their more advantaged peers. Begun in 1964, as part of the "War on Poverty," Head Start has enjoyed great public and bipartisan support. Presidents George Bush and Bill Clinton both pledged to increase federal funding so that all eligible' children could be served. Today 622,000 children, roughly 28 percent of eligible 3–5-year-olds, are served at a cost of $2.2 billion per year, or approximately $3,500 per child, per year (Anne Stewart, 1992).

Policyrnakers and the general public appear to believe that the benefits of Head Start are well known and well documented. However, a careful reading of the literature reveals that credible studies that demonstrate lasting effects of Head Start are limited. The studies that do exist are typically restricted to small geographic regions and specific racial groups.

In this study we use a national sample of data from the National Longitudinal Survey of Youth (NLSY) and the National Longitudinal Survey’s Child-Mother file (NLSCM) to reexamine the impact of Head Start on school performance, cognitive attainment, preventive medical care, and health and nutritional status. Although our study is no substitute for a national randomized trial, we do take some novel steps to sort out the effects of the Head Start program from possible nonrandom selection into the program. First, we contrast children who have been enrolled in the Head Start program with their siblings who have not, in order to control for family background effects on cognitive and health outcomes. Second, using the same sibling contrasts, we compare the impact of Head Start relative to "no preschool" with the impact of participation in other preschools relative to "no preschool." These "difference-in-difference" estimates further control for possible biases in the estimates due to child-specific determinants of participation in Head Start.

When selection is controlled in this way, Head Start has positive and persistent effects on the test scores and schooling attainment of white children, relative to participation in either other preschools or no preschool. In contrast, while the test scores of African-American children also increase with participation in Head Start, these gains are quickly lost, and there appear to be no positive effects on schooling attainment.

Relative to "no preschool," participation in either Head Start or preschool is associated with improved utilization of preventive medical care, as proxied by immunization rates, among whites and African-Americans. In contrast, there is no evidence that Head Start has any effect on child height-for-age, a longer-run indicator of health and nutritional status.

The rest of the paper is laid out as follows. The first section contains a brief overview of the previous literature. In the second, the methods are discussed. The third section provides a description of the data and our child outcome measures. The estimated effects of Head Start are presented in the fourth section. We conclude with a crude assessment of the possible long-term benefits of the program and weigh these against its cost.

Most previous studies of Head Start have focused only on assessing gains to IQ, despite the broad goals of the Head Start program. For example, although Head Start provides "a comprehensive health services program which includes a broad range of medical services" (Head Start Bureau, 1992), a recent review of 210 studies conducted by the US Department of Health and Human Services (Ruth McKey et al.. 1985) cites only 34 studies that have examined effects on health. These studies provide useful qualitative information about the health effects of the program, but very few of them attempt to quantify, the effects in any way. McKey et al. also note that very few studies have examined the impact of Head Start on schooling attainment.

The most convincing studies of the IQ effects of Head Start utilize a treatment and control design with random assignment. These studies typically find that there are initial gains to Head Start which fade over time and become insignificant by the third grade. However, Steven Barnett (1992) notes that experimental evaluations of the longer-term effects on IQ may be biased by attrition because children who move are likely to be lost from the experiment (although the direction of any bias is not obvious). A second limitation is that existing experimental evaluations have not been based on national samples of children in representative Head Start programs. Many studies, for example, focus exclusively on African-American children.

Head Start is also said to be associated with reductions in grade repetition, high-school dropout rates, and teen pregnancies, and with improvements in children’s medical care and health status (cf. Children's Defense Fund, 1992). The most widely cited evidence in support of these longer-term benefits of Head Start actually comes from experimental studies of model preschool programs such as the Perry Preschool Project or the Tennessee Early Training Project. These programs were funded at higher levels, involved more intensive interventions, and had better-trained staff than the typical Head Start program. For example, the Perry Preschool Project was funded at a rate of about $6,000 per child (almost twice that of the average Head Start program). Twenty years after the program, researchers found that the "treatments" were more likely to graduate from high school, had fewer pregnancies per female child, and had lower crime rates. However, the study involves a very small sample of 58 treatments and 65 controls, and many differences (such as the rate of teen pregnancy and the rate of violent crime) are not statistically significant (John R. Berrueta-Clement et al., 1984).

In summary, despite literally hundreds of studies, the jury is still out on the question of whether participation in Head Start has any lasting beneficial effects.

II Methods

The key empirical problem facing us is that, as we will see below, children are not randomly selected into the Head Start program. The program guidelines require that 90 percent of participants must be from families living below the federal poverty line although, in practice, 95 percent of children served in 1992 were poor (U.S. Department of Health and Human Services, 1993). In addition to being poor, Head Start children may also be disadvantaged in other observable ways. Estimates that do not take account of these differences are likely to underestimate the beneficial effects of the program. We will, therefore, examine the impact of Head Start on child well-being conditional on an array of observable mother and child characteristics.

The economic model of the family (Gary Becker, 1981) suggests that families choose whether or not to make the effort necessary to enroll their children in Head Start or other preschools on the basis of the expected returns from that investment. Families who find this investment worthwhile may make other unobserved investments in the child's human capital. In this case, studies that do not take account of unobserved differences between families may overestimate the beneficial effects of Head Start.

At many sites, there are fewer places than child applicants, and so participant selection will also reflect the choices made by program administrators. There are over 1,300 Head Start programs (Cheryl Hayes et al. 1990), all administered at the community level, and there is a good deal of heterogeneity, in their management and quality and in the interpretation of the federal guidelines (U.S. Department of Health and Human Services, 1993). Remarkably little is known about the selection practices used by administrators, although Ronald Haskins (1989) cites evidence that local staff tend to select the most disadvantaged children to participate in Head Start. Similar evidence on selection procedures is suggested by Lee et al. (1990). Unlike most adult training programs, evaluation is not based on child performance in the program, and so there is little incentive to cream off the more able applicants. In any case, whatever the mechanism underlying participant selection by administrators, estimates of the effects of Head Start that do not take this process into account may be biased.

In order to control for unobserved characteristics correlated with selection into the program we estimate models with fixed effects for each household. These models control for constant characteristics of households, including permanent income, maternal education, and other measures of (unobserved) family background and tastes. If it is primarily these constant factors that determine participation in Head Start, then fixed-effects models will provide unbiased estimates of the true program effects.

However, there may also be child-specific factors that affect participation. If, for example, parents wished to maximize the sum of their offspring's lifetime utility, then they might choose to enroll more able children in Head Start. On the other hand, if they seek to equalize outcomes, they might enroll the least able child. In the first case, fixed-effects estimates would provide an overestimate of the impact of Head Start, while in the latter case, they would yield an underestimate.

There are two other reasons why the inclusion of household fixed effects could bias estimated program effects toward zero. First, it is well known that in the presence of measurement error, differencing can result in "throwing the baby out with the bath water," since much of the true "signal" may be discarded while the "noise" remains.

Second, in the fixed-effects models the effects of Head Start are identified using the subset of households in which some children attended Head Start while others did not. If there are any spillover effects of Head Start from one sibling to the other, then the difference between the two siblings will be an underestimate of the true pro-gram effect. Spillover effects may be important because a child teaches his or her sibling something learned in Head Start, because the parent gains access to a service that is of benefit to both children, or because the parent makes compensating investments in the non-Head Start child.

In order to gain an understanding of the importance of the potential biases in the fixed-effects estimates due to child-specific factors, and spillover effects, we compare fixed-effects estimates of the effects of participation in Head Start to fixed-effects estimates of the effects of enrollment in other preschools. The decision to enroll a child in some other kind of preschool is also properly treated as a choice. As is the case for Head Start, fixed-effects estimates of the impact of other preschools will be unbiased if there are no unobserved child-specific characteristics that affect this choice, and no spillovers.

If the child-specific factors or spillovers bias the estimated coefficients on Head Start and on preschool in the same way, then the difference between the estimated coefficients will be accurately estimated, even if the individual coefficients are not. For example, suppose that parents send favored children either to Head Start or to preschool, depending on their means, and keep other children at home. In this case the fixed-effects estimates of Head Start and other preschools will both be biased upward. But the estimated difference between the effects of Head Start relative to no preschool and the effects of other preschools relative to no preschool will be subject to less bias.

We show below that, for several of our outcome measures, the fixed-effects estimates of the effects of Head Start exceed those of enrollment in other preschools. Still, there are two possible ways in which these results could be driven by the biases discussed above. First, it could be the case that children who attend either kind of preschool are systematically more favored or more able than their siblings and that the gap in ability between Head Start children and their stay-at-home siblings is greater than the gap between other preschool children and their siblings. Second, spillover effects could be greater within families in which a subset of children attend other preschools than within families with a subset of children attending Head Start.

It is difficult to rule out the possibility that the degree of parental favoritism is greater in households with some children who attend Head Start than in households in which some children attend preschool. However, we do not find any evidence consistent with the view that Head Start children are favored. For example, relative to their siblings, they are no more likely to be taken to the doctor in the first three months of life, and they score no higher on the "recognition of body parts" test, a test that was administered to sample children before they were age-eligible to attend Head Start. Moreover, we will discuss evidence below which suggests that preschool children may actually be more favored relative to their siblings than Head Start children, in which case the difference between the estimated effects of Head Start and preschool in the fixed-effects models provides a lower bound on the true difference.

Finally, the potential for spillover effects may be greatest in the most disadvantaged households and among children in programs like Head Start that make explicit attempts to improve parenting skills. In this case, Head Start effects will be underestimated relative to the effects of other preschools in the fixed-effects models. Spillovers are also likely to accrue to younger siblings, and we explicitly investigate this issue.

III. Data Description

The National Longitudinal Survey of Youth (NLSY) began in 1979 with 6,283 young women who have been surveyed annually ever since. As of 1990, these women were aged 25-32 and had given birth to over 8,500 children. In 1986, the NLS began a separate survey of the children of the NLSY, the National Longitudinal Survey’s Child-Mother file or NLSCM. The second and third waves of the NLSCM were undertaken in 1988 and 1990. In these two waves, mothers were asked whether their children had ever participated in Head Start. For this study, data on children and their mothers from all three waves of the NLSCM have been combined with information about the mother drawn from each wave of the NLSY. Attention is restricted to children aged 3 and older, and since the fixed-effects estimates are based on sibling comparisons, the sample includes only children who have at least one sibling over three years old. These rules result in a sample of nearly 5,000 children.

It is important to note that the original NLSY oversampled the poor, and so a relatively large proportion of the sample children-about one-fifth-participated in Head Start. In addition, due to oversampling there are large enough numbers of African-Americans to allow separate examination of this group.

A. Child Outcomes

We focus on four measures of child outcomes. The first pair are indicators of academic performance: the Picture Peabody Vocabulary Test (PPVT) score and whether the child has progressed through school without repeating a grade. The second pair of outcomes are related to child health: whether the child has been immunized for measles, and height standardized by age and gender using national norms (height-for-age). Table 1 provides details about the coding of these variables. Each row shows the measure, the age group for whom the measure was recorded, and some additional comments.

The relationship between test scores and future wages has received considerable attention from economists. In his summary of this literature, Eric Hanushek (1986 p. 1152) concludes that, in most studies, "years of schooling and measures of cognitive ability exhibit independent effects on earnings." Unfortunately, the majority of these studies focus on the scores of high-school students rather than on those of young children. However, Richard Murnane et al. (1993) find that a high-school senior's mastery of skills taught no later than the 8th grade (as measured by achievement on standardized tests) is an important determinant of future wages.

While there is some evidence that test scores predict future schooling and labor-market outcomes, the relationship is certainly not one-to-one. For example, developmental psychologists emphasize that a positive self-image and appropriate socialization may also contribute to scholastic success. Thus, the absence of grade repetition is examined as a second, more direct measure of academic performance.

Academic performance in early grades has been shown to be a significant predictor of eventual high-school completion (Atlee L. Stroup and Lee N. Robins, 1972; Dee N. Lloyd, 1978; Byron Barrington and Bryan Hendricks, 1989; Robert Cairns et al.. 1989; James Grissom and Lorrie Shepard, 1989; Margaret Ensminger and Anita Slusarcick. 1992). The relationship between high-school completion and wages is well-established: most studies find that an additional year of high school is associated with an 8-percent increase in lifetime wages (see Joshua Angrist [1990] for a recent estimate). High-school graduates are also less likely to be unemployed (James Markey. 1988). Educational attainment has also been shown to be associated with improvements in health (Michael Grossman, 1973) and job satisfaction (Robert Michael, 1982; Robert Haveman and Barbara Wolfe, 1984). These results suggest that by improving performance in early grades. Head Start participation could translate into a significant increase in the probability of graduating from high school with attendant improvements in future wages and employment probabilities.

As discussed above, in addition to early childhood education, the Head Start program provides a broad range of health-care services. Specifically, Head Start guidelines require that each child be given a physical examination; an assessment of immunization status; a growth assessment; vision, hearing, and speech tests; a hemoglobin or hematocrit test (for anemia); and a tuberculin skin test. Head Start centers are also required to screen for sickle-cell anemia, lead poisoning, and parasitic infection, if these problems are common in the community. The NLSCM data only allow us to assess immunization status, and growth (as discussed below), but given the guidelines, it is not unreasonable to suppose that children who gain access to immunization services are also more likely to gain access to at least some of the other required health services. In this case, immunization can be viewed as a marker for access to a bundle of important health services.

Head Start program performance standards also state that "every child in a part-day program will receive a quantity of food in meals... and snacks which provides at least 1/3 of daily nutritional needs... (Head Start Bureau, 1992 p. 40). Poor children are at much greater risk of nutritional deficiencies than other children. For example, 21 percent of 1-2-year-old children in low-income households suffer iron anemia compared to 7 percent of 1-2-year-olds from higher-income households (Barbara Devancy et al., 1989). These deficiencies have been linked to short attention spans, lethargy, impaired immune status, and growth retardation.

With our second measure of child health, we place the spotlight on nutrition. Height-for-age is an indicator of both nutritional status and health, and it captures the effects of longer-term deprivation. It has been profitably used in the economic history and development literatures (see for example Robert Fogel [1986], Reynaldo Martorell and Jean-Pierre Habicht [1986], and the review in John Strauss and Thomas [1995]). Many readers may be surprised to find that even in as rich a society as the contemporary United States, poor children are at risk of stunting, defined as low height-for-age. Data from the second National Health and Nutrition Survey (National Center for Health Statistics, 1981) indicate that 15 percent of poor female children 2-5 years old are below the fifth percentile of height-for-age. The corresponding figure for males is 11 percent.

Since child growth varies systematically with age and gender, height is standardized following guidelines from the National Center for Health Statistics (1976). Each child in the sample is compared with the median child in a population of well-nourished white children of the same age and gender in the United States, and the sample height-for-age expressed as a percentage of this median. However, given evidence of systematic deviations from the standards in populations of poor children, we use the measure of height taken closest to the child's fifth birthday in order to compare siblings of approximately similar ages.

B. Characteristics of Head Start and Other Children

The characteristics of Head Start children, other preschoolers, and all other children are presented in Table 2, distinguishing whites from African-Americans. Neither Head Start participants nor enrollees in other preschools are random samples of children: the probability of attending Head Start declines with income, whereas the probability of attending other preschools rises with permanent income. For example, among all children living in house holds in the bottom quartile of the permanent-income distribution, nearly 30 percent have attended Head Start, whereas only 15 percent attended other preschools. In the top quartile, 40 percent of children attend other preschools and 4 percent attend Head Start. Slightly over half the children in the sample never attend any preschool, and that fraction is essentially constant across the income distribution. This suggests that the mechanism governing selection to Head Start is quite different from that underlying selection into other preschools, or even into no preschool.

Table 2 shows that, in addition to lower average levels of permanent income, Head Start children are disadvantaged in most other observable respects. Relative to children who attended other preschools, children who attended Head Start have mothers and grandmothers who are less educated, and who had lower scores on the Armed Forces Qualification Test (AFQT), a measure of human capital. These differences between Head Start and other preschool children are all statistically significant for both whites and African-Americans, although the gaps are substantially larger among whites. For example, the difference in maternal education between white children in Head Start and white children in other preschools is 1.6 years, while the difference is only 0.8 years among African-Americans. The major exception to this generalization is that the mothers of African-American Head Start children are as tall as the mothers of other African-American children, while white mothers of Head Start children are shorter than other white mothers. White Head Start children also tend to be disadvantaged relative to children who attended no preschool, though the gaps are smaller than those between the Head Start and preschool groups. Among African-Americans, however, the only significant difference is in income: in all other observable respects, Head Start children are no worse off than their peers who attended no preschool.

Finally, Table 2 shows that, relative to whites, and controlling for preschool status, African-American mothers of Head Start children are actually better educated that comparable white mothers, although they tend to live in lower-income households. However, the AFQT scores of African-American women are much lower than those of whites, a fact that is true throughout the income distribution and suggests that AFQT measures more that native "ability."

C. Parental Favoritism? Evidence from Within-Family Income Differences

 As discussed above, the fixed-effects models estimated below are identified using the subset of families with at least one child who attended Head Start and at least one who did not. Similarly the effects of preschool attendance are identified using the subset of children in which at least one child attended preschool and at least one did not. Table 3 focuses on the within-family income changes that are associated with participation in Head Start and other preschools.

Panel A of Table 3 reports, for children who attended Head Start, other preschools, or no preschool (in the columns), the percentage with siblings who attended Head Start, other preschools, or no preschool (in the rows). For example, the entry in the upper left corner of the

table indicates that 41 percent of white children who attended Head Start had a sibling who also attended Head Start, and therefore, 59 percent had a sibling who did not. In the fixed-effects models, only the latter group is used to identify the effects of Head Start.

Of these 59 percent, the vast majority (about three-quarters) did not attend any preschool. Thus, fixed-effects estimates of the impact of Head Start will be based largely on within-family comparisons of children in Head Start with siblings who did not attend any preschool. The converse is also true: families with at least one child in preschool and at least one child not in preschool were unlikely ever to have had a child in Head Start. Estimates of the effects of Head Start and other preschools are therefore based on largely non-overlapping samples of families. This result is important because it facilitates the comparison of Head Start effects to the estimated effects of attending other preschools.

Panel B of Table 3 presents the means and standard errors of two measures of income for each type of sibling pair. Permanent income (which is family-specific) is reported in the first column, while income at the time the child was three years old is reported in the second. Income at age 3 is relevant since this is the time when most children would enter Head Start or some other preschool. Rows 1-3 confirm that, relative to children who attended other preschools or no preschool, Head Start children are disadvantaged both in terms of permanent income and income at a point in time.

A second fact, which is apparent from row 4 of Table 3, is that there is little within-family difference in household income at the time the child was age 3 between Head Start children and those who never went to preschool, In contrast, rows 5 and 6 indicate that transitory income is associated with within-family movements between other preschool and no preschool, and also between Head Start and other preschool The within-family gap between preschool and no-preschool children is about $6,000 among whites and $8,000 among African-Americans. Similarly, the within-family gaps between other-preschool and Head Start children are $5,000 and $3,000 for whites and African-Americans, respectively.

These results show that, when family income rises, parents are more likely to send age-eligible children to preschool. Assuming that parents want to do what is best for their children, but are constrained by income, this finding suggests that a favored child would be more likely to be sent to preschool, other things being equal. We do not find any similar pattern for Head Start. Hence, there is some evidence consistent with the view that preschool children are actually more favored relative to their stay-at-home siblings than Head Start children, which implies that the difference between the estimated effects of Head Start and of preschool in the fixed-effects models discussed below may be an underestimate of the true Head Start premium.

 Tables 4 and 5 present regression estimates of the effects of participation in Head Start and other preschools on the four child outcomes. In order to highlight the importance of controlling for observed and unobserved family-specific effects, three sets of estimates are presented in each case. "Unadjusted" ordinary least-squares (OLS) estimates [in columns (i)-(iii)] do not control for any observable covariates: this baseline shows the sample means. "Adjusted" OLS estimates [in columns (iv)-(vi)] do control for mother- and child-specific observables. Fixed-effects estimates [in columns (vii)-(ix)] also control for all unobserved time-invariant mother-specific effects in addition to child-specific observables.

All the regressions are estimated separately for whites and African-Americans; to facilitate comparisons between the two groups, difference between the estimated coefficients are reported in the third column in each panel. In each regression, the excluded category is children who did not attend preschool. The F statistic for the test that the estimated 'difference-in-difference" between Head Start and other preschool children is zero is reported just below each panel of estimates (along with the associated p value).

The observables in the "adjusted" OLS regressions include child age, gender, and whether the child was the first born, (log) household permanent income, the mother's education, her AFQT score, her height, the number if siblings in the mother's household when she was age 14, and the education of the maternal grandmother. The fixed-effects models include child age, gender, and whether the child is the first born, as sell as household income at the time the child was age 3.

A. Measurers of Academic Performance

The first three columns of panel A in Table 4 indicate that the PPVT scores of white children are, o average, about twice those of African-American children. In part, this is a reflection of the fact that whites live in higher-income households than African-Americans. But that is only part of the story since nonparametric estimates indicate that white children have higher PPVT scores at all income levels (Currie and Thomas, 1993).

Within racial groups, white children who attended other preschools or no preschool tend to score better, on average, than Head Start children. For example, white Head Start children score an average of 5 percentile points lower on the PPVT than white children who did not attend preschool and 15 percentile points lower than whites who attended other preschools. Both of these differences are statistically significant. In contrast, there are no statistically significant differences among African-Americans.

Moving across the columns in panel A in Table 4 shows the importance of controlling adequately for all observed and unobserved family characteristics associated with selection into Head Start. Column (iv) suggests that, among whites, the difference between the PPVT scores of Head Start and other children disappears when observables are controlled.

However, column (vii) demonstrates that when unobserved differences between families are controlled, using mother fixed effects, participation in Head Start is actually associated with a significant 6-percentile-point increase in the PPVT score relative to no preschool, while participation in other preschools has no statistically significant effect on test scores. The gap between the effects of Head Start and other preschools is statistically significant. The difference between columns (iv) and (vii) indicates that, consistent with Haskin's (1989) observations, it is the most disadvantaged white children in terms of unobservables who are selected into the Head Start program. On the other hand, controlling for unobservables has little effect on the estimated coefficient for other preschools, once observable characteristics are included in the model.

The results for African-Americans indicate that selection may be less important for them: there are no statistically significant effects of Head Start or preschool in any of the three specifications. Column (ix) shows that the difference between the Head Start effects for whites and African-Americans is large – nearly 6 points – and statistically significant.

We turn nest to our second measure of academic performance: absence of grade repetition. The first three columns of panel B in Table 4 show that about one-third of white and nearly half of African-American sample children age 10 or older are reported to have repeated a grade. Although white Head Start children are about 20 percent more likely to have repeated a grade than white children who attended other preschools, this difference is not statistically significant. Among African-Americans, the gaps between the different groups of children are even smaller. The OLS estimates in columns (iv)-(vi) also indicate that there are no statistically significant effects of type of preschool on the probability of grade repetition.

However, the fixed-effects estimates, shown in columns (vii)-(ix) indicate that whites who attended Head Start are 47 percent less likely to repeat a grade, relative to their siblings who did not attend preschool. Those who attended another type of preschool are no less likely to have repeated a grade than their siblings who stayed at home. The "difference in differences," that is, the gap between the effect of Head Start and the effect of preschool, is also large (40 percent) and statistically significant (p value = 0.01).

In contrast, attendance at either type of preschool has no statistically significant effect on the probability of grade repetition among African-Americans (although the point estimate of the coefficient on the other preschools is large). Once again, the racial difference in the impact of Head Start is statistically significant.

In sum, after controlling for mother-specific observables and unobservables we find that, for whites, the academic performance of Head Start children is significantly better than that of siblings who stayed at home. In addition, the estimated effects of Head Start are much greater than those of attending other preschools once both observable and unobservable characteristics of families are controlled. Among whites, this difference-in-difference estimate is statistically significant both for PPVT scores and for grade repetition. Among African-Americans, however, the tale is more dismal: neither Head Start nor other preschools is associated with enhanced academic performance.

B. Measurers of Health Status

 Table 5 presents the estimated effects of participation in Head Start and other preschools on two measures of health status: immunization probabilities and height-for-age. The first three columns of panel A suggest that both whites and African-Americans are about 15-percent more likely to have had a measles shot if they attended Head Start rather than another preschool. These gaps are statistically significant. There is little difference in these means between the other-preschool and no-preschool children, which is surprising in light of the differences in family background between these two groups. For both racial groups, the difference in differences between Head Start and other preschool children is statistically significant.

Column (iv) shows that, among whites, controlling for observables reduces the effects of Head Start to zero, while the effect of attending other preschools increases slightly and becomes statistically significant. Among African-Americans, the inclusion of observables reduces the Head Start advantage by over half, but it remains significant.

When fixed effects are included [in columns (vii) and (viii)], we find that Head Start is associated with an 8–9-percent higher probability of being immunized among both white and African-American children. Attendance at other preschools is also associated with a higher probability of being immunized. While the estimated coefficient on preschools is greater than the estimated effect of Head Start among whites, the difference is not statistically significant. Among African-Americans, the effect of other preschools is not significantly different from zero, but it is not significantly different from the coefficient on Head Start either. Relative to other preschools then, there is not health-care "premium" associated with Head Start.

The relationship between type of preschool and child height-for-age is presented in panel B of Table 5. The unadjusted OLS estimates [in columns (i) and (ii)] show that white children who attend preschools are significantly taller than other white children, but that African-American children who attend Head Start are taller still. The coefficient on preschool in column (ii) is not statistically significant. However, the hypothesis that Head Start and preschool have the same effect on the height-for-age of African-Americans cannot be rejected with any confidence.

When observables are controlled in column (iv) and(v), the preschool effect among whites is somewhat weaker, but it remains significant. A good part of the difference between columns (i) and (iv) is accounted for by the influence of maternal height, although other measures of maternal human capital (her education) are also statistically significant. This result suggests that height is influenced both by genetic factors and by parental investments in the health and human capital of children. The fixed-effects estimates for whites, in column (vii), eliminate the influence of all shared genetic characteristics as well as all other fixed maternal characteristics; this results in a further weakening of the relationship between preschool and child height, although it remains positive and significant, albeit at a 7-percent level.

Among African-Americans, the inclusion of observable maternal and child characteristics [in column (v)] cuts the positive correlation between Head Start and child height by more than half. It also becomes statistically insignificant. Similarly, column (viii) shows that we do not find any statistically significant effect of either Head Start or preschool when fixed effects are included in the model.

These results suggest that the positive correlation between Head Start and height-for-age among African-Americans that is noted in column (ii) reflects the selection of taller African-American children into the program. This impression was confirmed by estimating regressions of birth weight on participation in the program. Birth weight is highly correlated with future child height-for-age, but it could not possibly be influenced by future participation in Head Start. We found that African-American children who attended Head Start were heavier at birth than African-American children who did not. For whites, however, we did not find any correlation between birth weight and enrollment in Head Start or preschool, so the positive effect of preschool on height-for-age appears to be a genuine program effect.

Thus, in spite of positive effects of attendance at Head Start or other preschools on the utilization of preventive health care, the large nutritional component of the Head Start program, and the fact that other preschools appear to have positive effects on growth of some children, we find not evidence that participation in Head Start has an effect on nutritional and health status as measured by height-for age.

C. Differences in the Effect of Head Start Among Whites and African-Americans

 The cognitive effects of Head Start appear to vary dramatically by race, even when selection into the programs is taken into account: Head Start has a smaller effect on the test scores and schooling attainment of African-Americans than on the test scores and academic achievement of whites. Why does race matter?

One hypothesis is that there is heterogeneity in the Head Start programs that serve children of different races. While most programs are in compliance with most standards, slightly over 11 percent of Head Start operators monitored in 1993 were found to be out of compliance with 50 or more of 222 items reviewed, while another 18 percent needed improvement in 26 – 50 areas (U.S. Department of Health and Human Services, 1993). It is possible that African-American children are more likely to be served by inferior programs. Unfortunately, this hypothesis cannot be tested directly, as we have no information about individual programs.

An alternative hypothesis is that the benefits of compensatory education depend both on the program itself and on the child’s home background, including, for example, the level of resources at home, as well as the type and quality of school attended after Head Start. To the extent that African-American children come disproportionately from more disadvantaged homes, located in poorer communities, and attend troubled schools, one might expect Head Start to have either smaller initial effects or effects that dissipate more quickly over time.

We begin to address these issues by estimating models that allow the effects of Head Start and other preschool attendance to vary with maternal AFQT and child age. These results are shown in Table 6. All of the models included fixed effects. We do not show results for height-for-age, since there were no significant effects of Head Start (or significant racial differences) to be explained.

Maternal AFQT can be regarded as an index of maternal background or of human capital. It is highly correlated with years of education, as shown in Figure 1, but has the advantage of being a continuous rather that discrete variable. If children from better backgrounds gain more from Head Start or preschool, then the interactions between AFQT and Head Start or preschool will be positive.

The results in columns (i) and (ii) of panel A indicate that the positive effects of Head Start on PPVT increase with AFQT among both whites and African-Americans. However, neither interaction is statistically significant. The interactions between AFQT and preschool are also insignificant. Turning to the absence of grade repetition, column (iv) shows that, among whites, there is a large and statistically significant interaction between Head Start and AFQT: a 10-point increase in the normalized maternal AFQT score reduces the probability of failure among Head Start Children by 8 percent. We do not find any similar effect among African-Americans [column(v)]. Moreover, the differences between whites and African-Americans in the AFQT X Head Start interaction is significant (at the 8 percent level) [column (vi)]. We do not find any significant interactions between preschool attendance and AFQT for either race.

Finally, the results shown in columns (vii)-(ix) indicate that, in the regressions for immunization probabilities, interactions between Head Start and AFQT and between other preschools and AFQT are all positive but not statistically significant. In sum, there is weak evidence that children from better backgrounds, as measured by maternal AFQT, gain more from Head Start, but the interaction is only statistically significant in the regressions for absence of grade repetition among whites.

Interactions between the type of preschool and child age allow us to address the question of whether the effects of Head Start and other preschools persist as the child grows older. These estimates are reported in panel B of Table 6. Columns (i) and (ii) contain one of our most interesting results. Not only is the direct effect of Head Start large, positive, and significant for both whites and African-Americans, but the effect (of nearly 7 percentile points) is essentially identical for both racial groups.

This finding stands in sharp contrast with the results discussed above. In Table 4 we found that Head Start was associated with higher PPVT scores among whites but that African-American children did not enjoy similar benefits. The difference lies in the age interactions while the interactions are always negative, for whites they are small and statistically insignificant, while for African-Americans they are large and significant. Thus, for example, by age 10 African-American children have lost any benefits they gained from Head Start, while 10-year-old white children retain a gain of 5 percentile points. There is no evidence of a similar interaction effect among children who attend preschool.

Our results for African-Americans are thus consistent with those of earlier studies (which tended to be dominated by African-American subjects). When we focus on only young African-American Children, we find clear benefits of Head Start. However, in a sample of African-American children of all ages there is no effect of Head Start. This is because the benefits die out very quickly. In contrast white children experience the same initial gains from Head Start but they retain these benefits for a much longer period.

It is also possible to ask whether the rate at which the benefits of Head Start dissipate among African-Americans depends on the environment at Home. To do this, we have estimated models (not shown) that include "triple interactions" among age, Head Start and maternal AFQT. If children from better backgrounds retain the gains from Head Start longer, then this triple interaction will be positive (offsetting the fact that the beneficial effect declines with age). We