The relatively large number of counties with substantial numbers of respondents despite the modest overall sample size is due to clustering in the NLSY79 sampling design. Additionally, obser- vations in the working sample are split relatively evenly between those who were assigned positive Head Start exposure using the method described above (59 percent), and thosewhowere assigned no exposure because they were too old for Head Start when a local program was launched (41 percent), which generates reasonably large samples of both treated and untreated observations within counties. Figure 2 displays a kernel density plot of county level Head Start spending per child
ages three to six for the counties represented in the NSLY79 sample over the same period (in 2012 dollars). Mean spending, indicated with the dashed line, was $170 in this period. More importantly, Figure 2 indicates substantial heterogeneity in early Head Start funding intensity across counties, with several counties spending in excess of $400 per child ages three to six. As noted above, this variation likely reflects the total en- rollment of local Head Start programs, as well as the prevalence of summer versus full- year programs, and allows me to analyze whether children exposed to more highly funded Head Start programs have better outcomes than children exposed to programs with lower spending levels.
C. Outcome Measures
A major advantage of the utilized data and approach is that I am able to estimate the effects of Head Start on a wider variety of outcomes and over a longer portion of the life cycle than previous work in this area. I study outcomes from three broad areas where Head Start was designed to have positive impacts: educational attainment, labor market outcomes, and health status.
14. The youngest NLSY79 respondents were born in 1964 and therefore turned six in calendar years 1969 and 1970.
F ig u re
1 H ea d S ta rt In tr o d u ct io n b y C o u n ty , 1 9 6 6 – 1 9 7 0
N o te s: D at a fr o m
N at io n al A rc h iv es
an d R ec o rd s A d m in is tr at io n .
1110 The Journal of Human Resources
I measure educational attainment using each respondent’s total years of completed education, as well as indicators of whether they were awarded a high school diploma or a four-year college degree. To ensure that I observe final educational attainment, I define these measures using the most recently available survey wave completed after age 30, which in most cases occurred when respondents were in their mid 40s. With respect to labor market outcomes, I begin with two income measures. First is the
mean of all individual wage and salary observations occurring between ages 30 and 48 (measured in annual 2012 dollars), which I refer to as “own income.” This individual-level measure conflates labor supply decisions with earning power, which may be especially problematic for females, and it also omits some relatively common forms of nonwage income. As such, I also construct an income measure that includes wage and salary income for both the individual and their resident spouse (if present), as well as unemployment insurance, child support, and investment income for both the individual and their resident spouse. I again convert annual observations to 2012 dollars, then take the mean of all observations from ages 30–48, and I refer to this measure as “family income.”15 In addition to income, an important aspect of labor market well-being is employment status. Each NLSY79 wave collected information on weeks unemployed in the past year, and I use this information to construct avariable measuring the proportion ofobservationsfromages 30– 48 that each individual was unemployed for two or more weeks. As noted above, health and nutrition have always been a major component of Head
Start programing. While the NLSY79 did not collect comprehensive health data in each
Figure 2 County Head Start Funding Density Notes: Figure shows kernel density plot of county level Head Start funding per child ages three to six in 2012 dollars. Estimated with Epanechnikov kernel and bandwidth of 50. Dashed line indicates sample mean of $170. Counties with over $1,000 in spending per child ages three to six, representing approximately 2 percent of observations, are omitted from the figure but used in the calculation of the sample mean.
15. For both income measures only respondents with five or more valid annual observations are used.
wave, as respondents turned 40 they completed a detailed “40 and over health module.” Using information from this module, I first assess two global health measures: Each individual’s self-rated health (on a 1–5 scale) and an indicator of whether health limits their ability to perform moderately strenuous activities, their ability towork, or their ability to engage in social activities. Respondents also report whether they suffer from various chronic conditions, and I construct a variable measuring how many of the following conditions the individual reports: a heart condition, severe tooth or gum problems, asthma, and high cholesterol. These conditions were selected because they are rela- tively common and are potentially sensitive to the health related services that were typically included the studied Head Start programs. While the availability of many outcome measures is in general a strength, using such a
large set of dependent variables also presents some estimation-related issues. The most important of these is multiple inference: With nine separate outcomes, as well as various subsamples and specifications, I test dozens of hypotheses, which increases the risk of false rejection (type 1 error). Additionally, many of the utilized outcome measures are closely related. For instance, educational attainment is strongly correlated with both own income and family income, and both income and education are strongly correlated with health outcomes. These correlations across measures make it difficult to ascertain how much new information is contained in results for each specific outcome. A final issue is measurement error. All of the utilized outcomes can reasonably beviewed as components of a single underlying index of socioeconomic well-being, but each specific outcome is likely measured with substantial error, which can destabilize the corresponding estimates. To address these issues I follow O’Brien (1984), Carneiro and Ginja (2014), Deming
(2009), and others and construct a summary index of the nine outcomes described above. Specifically, I first standardize each measure to have a mean of zero and a standard deviation of one and equalize signs across outcomes, so that positive values corre- spond to more desirable outcomes. I thentake theweighted average of thesestandardized measures using weights equal to the inverse of the sample covariance matrix, which accounts for dependence across outcomes.16