Profiling the plight of disconnected youth in America

Abstract

In a successful transition from youth to adulthood, individuals pass through a sequence of roles involving school, work, and family formation that culminate in their becoming self-sufficient adults. However, some “disconnected” youth spend extended periods of time outside of any role that constitutes an element of the pathway towards adult independence. Assisting these youth requires a systematic understanding of what “disconnection” means, how many disconnected youth there are, who these youth are, and how the scale of the problem has evolved over time. Using the National Longitudinal Surveys of Youth for 1997 and 1979, we address these issues by creating concrete definitions of “disconnection spells” using rich data on youths’ enrollment, work, and personal histories. We estimate a multi-state duration model to account for right censoring and to understand differences across salient sub-groups. Our estimates imply that in the early 2000s, almost 19% and 25% of young men and young women, respectively, experienced a disconnection spell by age 23 using our basic definition. These rates are substantially higher for certain sub-groups defined by race/ethnicity, parental education, and government aid receipt, rising as high as 30+% by age 23. Approximately 60% of youth with a disconnection spell have it last longer than a year, and close to 10% have it last longer than 4 years. However, once reconnected, a majority of youth go at least three years without a re-disconnection spell. Patterns of initial disconnection changed markedly from the 1980s to the 2000s, as young women saw a 12 percentage point decline over time. Moreover, the Black-White gap in disconnection has fallen for women, but increased for men. Our profile of disconnection experiences provides a starting point for government agencies aiming to understand where, how, and with whom to intervene to prevent lengthy disconnection spells.

1. Introduction

In a successful transition from youth to adulthood, individuals pass through a sequence of roles involving school, work, and family formation that culminate in their becoming self-sufficient adults. Successful paths comprise numerous orderings and timings of these roles. Many youth undergo a misstep in one or more of these areas (e.g. dropping out of school, lack of steady employment, involvement with the justice system, unplanned parenthood) without being seriously detoured from the path to independent adulthood. However, in spite of the billions of federal, state, local and private dollars spent on a wide range of programs aimed at helping disadvantaged youth, many youth still experience “disconnection” – extended periods of time outside of any role that constitutes an element of the pathway towards adult independence. Point-in-time estimates of the prevalence of disconnection depend on the particular definition used, but all studies estimate a substantial population of youth – at least 10% – are disconnected at any given time (see, e.g., Lewis and Gluskin [2018]).

Serious missteps in the transition to adulthood can lead to formidable costs both in personal and social terms. This includes never finishing high school, single motherhood, and greater dependence on welfare [Turner et al. (2006), Rebecca Blank and Kovak (2008a, 2008b)]. Male youth who are out of school and not at work for long periods are more likely to engage in delinquent behavior and in illegal activities to earn a living [e.g. Lochner and Moretti (2004), Ihlanfeldt (2007)]. When the transition to adulthood is eventually made, those who have experienced these problems often continue to pay long-term penalties in the form of lower earnings and fewer weeks employed [see, e.g., Card (2001), Gregg (2001), Pharris-Ciurej, Fernandez and Porter (2018)]. Social costs include higher crime rates, more young children in poverty, and lower economic productivity. Costs to government include higher transfer payments and social support expenses, as well as lost tax revenues. For these reasons, U.S. policy makers have periodically shown interest in remedies, even extending the eligibility age for training under the Workforce Investment Act to target youth who “have become disconnected from both education and the labor market” [Fernandes-Alcantara (2015); see also, Loprest and Nichols (2011), Mendelson (2018)].

A major challenge encountered in implementing any policy intended to assist disconnected youth, however, involves acquiring a systematic understanding of which youths lose their way in the transition to self-sufficient adulthood, and of the full range and combination of roles through which individuals can become connected. By developing profiles depicting the timing and duration of disconnectedness, and the personal and family background characteristics associated with it, policy makers can direct and shape interventions in ways that will maximize their effectiveness. Interventions can be targeted to groups most at risk of experiencing difficulties in the transition process before they become disconnected; and for those who may currently be underserved by existing programs, services can be reoriented to access individuals at the mix of ages and roles when they are most approachable for reestablishing “connectedness”. For these purposes, simple cross-sectional analyses of the number of people disconnected at a point in time or accumulated “disconnected” or “connected” months by a certain age are not sufficient; instead, one requires a longitudinal framework to capture timing of entry into disconnection spells, transition into reconnection spells, and persistence of reconnection spells.

In view of these goals, this paper aims to address four questions:

1. What is a useful way to define a “disconnection spell” and “reconnection spell”?

2. What is the pervasiveness, persistence, and timing of disconnection and reconnection spells in recent decades?

3. Which sub-groups of the population are more likely to enter and remain in disconnection and reconnection spells?

4. How has the nature of disconnection and reconnection spells changed from the early 1980s to early 2000s?

To answer these questions, we utilize the National Longitudinal Surveys of Youth for 1997 (NLSY97) and for 1979 (NLSY79), which permit monthly tracking of schooling, employment, and family support activities characterizing disconnection. In contrast to other rich longitudinal data sources (e.g. Panel Survey of Income Dynamics [PSID], Survey of Income and Program Participation [SIPP]), the NLSY data provide consistent data sources across a 25 year period and provide comprehensive information on individual, social, and economic events during the transition from adolescence to young adulthood. We use the longitudinal data to define concrete notions of disconnection, reconnection, and re-disconnection spells. Using an approach described in more detail in MaCurdy et al. (2010), we estimate a multi-state duration model that detects: (i) the likelihood that a youth with particular characteristics enters a first disconnection spell at various ages, (ii) the length of time a youth spends in this spell if it occurs, (iii) the tendency of a youth to exit a disconnection spell and reconnect, and (iv) the likelihood that a youth disconnects again and enters a re-disconnection spell. The duration model accounts for right censoring (e.g. in the length of reconnection spells) and facilitates analysis of sub-groups that are small in number in NLSY survey samples. This empirical framework not only portrays a comprehensive picture of experiences, it further offers a natural setting for identifying when disconnection spells occur and who makes up the persistent group.

Our answers to the four research questions fill several critical gaps in the literature on disconnected youth. First, we define a youth as entering a disconnection spell if he/she has begun a period of twelve months in which the initial month and at least eight subsequent months are spent not employed, not enrolled in school, and – in a second, stricter definition – not living with a spouse. Similarly, a reconnection spell is defined to begin when a youth begins a period of twelve months in which the initial month and at least three subsequent months are connected in the sense of having employment, enrollment, or residence with a spouse. Disconnection spells are defined to continue until a reconnection spell begins, and reconnection spells continue until a re-disconnection spell begins. Our focus on these notions of spells differs sharply from prior literature. Several studies focus on short-term non-enrollment in school, difficulty in the labor market, or experience with the justice system [e.g. Wertheimer et al. (2002), Wald and Martinez (2003), Wald et al. (2003), Burd-Sharps and Lewis (2017), Meyer and Mittag (2019), National Center for Educational Statistics (2017)], and Brown and Emig (1999) use the NLSY79 to examine the circumstances of youth who spend 26 weeks out of a year not working, not in school, and not married to a connected spouse [see also Brown (1996)]. Other studies focus on stricter criteria. Fernandes-Alcantara (2015) and Edelman et al. (2006) examine the fraction of U.S. youth who spend an entire year without work or school [see also Pfeiffer and Seiberlich (2011) for Germany]. Our approach improves upon these studies by pursuing a middle road. Unlike those with a short-term focus, our definition deliberately avoids capturing youth who may accumulate a number of scattered disconnected months here and there without ever having a sustained problem. Unlike those imposing stricter criteria, our approach captures youth with intense and persistent bouts of non-employment and non-enrollment, even if they have occasional “connected” months.

Second, a large number of youth experience at least one disconnection spell and a significant minority of these youth have that spell last a considerable time. We find the probability of ever experiencing a disconnection spell by age 22 was about 23% for young women and 18% for young men in the NLSY97 cohort. Adding the criterion that a youth was not living with a spouse, these numbers fall only slightly to 19% for young women and are almost unchanged for young men. Of women who become disconnected at least once, we estimate that nearly 30% or more had a disconnection spell lasting more than two years. Due to our use of complete histories on the NLSY97 sample from age 15 through age 23 and the application of our duration model, our estimates of the share of youth who have experienced a disconnection spell by their early 20s and the length of those disconnection spells (i.e. time until reconnection) is as high or higher than estimates common to the prior literature, even in cases where our definitions are stricter. For instance, point-in-time prevalence estimates in Lewis and Gluskin (2018) and Loprest and Nichols (2011) are similar or lower despite the use of looser disconnection definitions, reflecting the fact that we are able to examine the portion of the population that ever experienced disconnection. Hair et al. (2009), who use the NLSY97 and a definition of disconnection based on 26 consecutive weeks out of work, out of school, and not married, find that 19.7 percent of youth ever experienced disconnection. Their sample has significant right censoring that they do not account for; consequently, their estimates of disconnection incidence and disconnection spell length are biased downwards. Loprest and Nichols (2011) complement their point-in-time results with hazard models accounting for right censoring, but cannot address left censoring and length-biased sampling inherent to their SIPP data, which could affect both their disconnection and reconnection estimates. For example, they find that approximately 42% of low-income single mothers with a disconnection spell have a spell last 12 or more months, while we find that more than 60% of all youth with a disconnection spell have the spell last this long. While the difference in results could be related to our differing definitions and populations, a portion of the difference could be explained by Loprest and Nichols (2011) being unable to track the full course of a disconnection spell in their SIPP sample and restricting attention to those for whom they observe the start of the spell.

Third, basic socioeconomic characteristics are powerful predictors of initial disconnection, but not of reconnection and re-disconnection. Parental education, race, and NLSY aptitude test scores all play a practically and statistically significant role in initial disconnection. For example, even though youth whose parents did not complete high school made up only 11% of the total youth population in NLSY97, our results suggest that they made up more than 22% of youth with at least one spell of disconnection by age 23. Male African-American youth made up 8% of the youth population, but 14% of youth with at least one spell. Those in the lowest quartile of NLSY’s test score made up 52% of youth with at least one spell. We use the nationally representative survey to make these and other comparisons across key sub-groups for initial disconnection, reconnection, and re-disconnection. In contrast, studies that examine disconnection spells in a longitudinal framework typically have analyzed only certain subsets of the population or only a portion of the entire disconnection trajectory. For example, a number of studies focus on disconnection spells among single mothers who are not in school and have limited income from work and government transfers [e.g., Blank and Kovak (2009), Loprest and Nichols (2011); see also Rebecca Blank and Kovak (2008a,2008b), ASPE (2011), Hetling et al. (2015), Meyer and Mittag (2019), Turner et al. (2006)]. Similarly, others have examined disconnection spells in select regions of the country. Hok et al. (2016) use longitudinal data from Washington State (1998–2009) to examine the relationship between disconnection spells and family reunification after foster care placement. Sansale et al. (2019) use the NLSY97 to study long spells of employment disconnection, but focus on the role personality may play in determining disconnection duration among 24–35 year olds [see also Almlund et al. (2011); for the broader economic literature on unemployment duration, see e.g. Farber et al. (2015)]. Millett and Kevelson (2018) use the Educational Study of 2002 to examine the effect of demographic characteristics and prior disconnection on renewed disconnection, but do not examine the full trajectory of disconnection experiences.

Fourth, the likelihood of ever disconnecting declined across the NLSY79 and NLSY97 cohorts, marginally for young men but significantly for young women. Using the less strict definition, for example, a young woman’s probability of ever experiencing a disconnection spell by age 20 dropped from 28% to 17%. After conditioning on family background and test scores, we see that the experience of African-Americans has changed markedly with a striking difference by sex: Black women were more likely to ever have a disconnection spell and less likely to reconnect than other women in the NLSY79, but these gaps closed entirely by the early 2000s period in NLSY97. In contrast, Black men were no more likely to ever have a disconnection spell than other men in the NLSY79, but a significant gap emerged in the NLSY97. To the best of our knowledge, our paper is the first to examine changes in disconnection over time, but our results are consistent with existing literature on youth employment and enrollment. Card and Lemieux (2001) show that cohorts in the 1990s are enrolled in school at higher rates than cohorts in the late 1970s and early 1980s. Ingels et al. (2012) show that post-secondary enrollment among youth was 22 percentage points higher in 2006 than in 1974, slightly outpacing the lower employment rates in 2006 than in 1974. Consistent with our evidence on particular improvements for women, Ingels et al. (2012) find that the gap between women and men in post-secondary employment/enrollment closed from 1974 to 2006 [Heckman and LaFontaine (2010) demonstrate graduation rates, as opposed to enrollment, have actually stagnated over time, though there are still more adverse changes for men]. Altonji et al. (2012) find that on dimensions such as parental education, youths have become better positioned for long-term success over time. They show improvements in measured individual ability underlying earnings gains have been higher for women than men.

Our multi-state duration model provides a flexible approach to accounting for right censoring, incorporating non-monotonic duration dependence, and modeling complex sequences of potential outcomes, an approach that is substantially simpler than standard competing risk models with multiple states. We believe this approach is a useful descriptive tool that researchers from a wide array of backgrounds can apply easily and fruitfully to conduct longitudinal analysis. Importantly, we do not use the model to uncover structural parameters from an economic model. A long literature discusses the econometric challenges involved here, in areas such as separate identification of unobservable individual heterogeneity and “true” state dependence [see, e.g., Heckman and Singer (1984a), Heckman and Singer (1984b), Baker, et al. (2000), Kroft et al. (2013), Torgovitsky (2019)]. Nonetheless, various elements of our empirical framework are motivated by the rich literature on children’s school continuation decisions, how they are shaped by family and other characteristics, and the transition from school to work [see, e.g., Cameron and Heckman (1998), Eckstein and Wolpin (1999), Black et al. (2005), Eckstein and Wolpin (1995)]. We hope our empirical results motivate investigation into the extent to which existing economic models are capable of explaining the patterns in timing and persistence of disconnection spells that we describe.

The remainder of this paper consists of six sections. Section 2 discusses the data, defines disconnection and connection spells, and describes our multi-state duration model. The four subsequent sections address each of the four research questions above. Section 3 illustrates the usefulness of our disconnection spell definition. Section 4 describes the pervasiveness, persistence, and timing of disconnection and reconnection spells using the NLSY97 cohort. Section 5 examines differences across important sub-groups in the NLSY97 cohort. Section 6 documents how youths’ disconnection experiences have changed between the time periods covered by the NLSY79 and NLSY97 cohorts. Section 7 summarizes key findings and concludes.

2. Empirical framework for examining disconnection trajectories

Profiling the complete trajectory of youths’ disconnection experiences requires detailed longitudinal data on youth, a clear definition of what one means by disconnection and connection spells, and an empirical framework capable of handling data limitations and research goals. This section first describes the data we rely on and introduces a notion of a disconnection spell. We then explain the rationale for, and formulate, a multi-state duration model for characterizing disconnection trajectories.

2.1. Description of data

The NLSY97 offers an unparalleled data source for assessing not only which youth ever become disconnected, but also for documenting their experiences prior to and after reaching this state. Beginning in 1997, the survey collects data annually to supply elaborate longitudinal information on a sample of 12–17 year olds. The survey also collects retrospective data, which allows us to create event histories starting in 1994, when individuals were aged 9–14. The survey offers a nationally-representative sample as well as supplementary samples for Blacks and Hispanics. The latest survey round used in our analysis is Round 13, conducted in 2009. This ensures we have data on all youth until at least age 24 (and given the forward-looking nature of our spell definitions [see Section 2.2], this allows us to include ages up to 23 in our final estimation samples). These data report detailed, monthly information on each youth’s educational and work activities, experiences with the justice system, and enrollment in government services. In addition, the data contain information on youths’ backgrounds, including socioeconomic characteristics and test scores on the Armed Services Vocational Aptitude Battery (ASVAB).

The NLSY79 provides a comparable data source to NLSY97 for profiling the experiences of youth in the early 1980s. The survey reports data collected annually from a nationally representative cohort of youths who were between 13 and 22 years of age in 1979. The survey also collects retrospective data, but only back to 1978. The latest survey round used in our analysis is conducted in 1990. This again ensures we have data on all youth until at least age 24. The data include monthly employment and enrollment histories, most of the same socioeconomic characteristics as the NLSY97, and test scores from the Armed Forces Qualifying Test (AFQT), a predecessor of the ASVAB.

We briefly describe the construction of key variables and application of sample restrictions using the NLSY97 and NLSY79 data.

2.1.1. Construction of key variables

Key variables used in the analysis are those for school enrollment, employment, marriage, ASVAB/AFQT composite scores, and parental characteristics. We discuss each in turn.

To measure enrollment, we primarily rely on the event history schooling variables in the NLSY, which report enrollment status of each respondent by month. The event history variables correct for apparent inconsistencies in the way respondents answered schooling questions (e.g., the reported highest grade completed moves down over time for some respondents). The NLSY97 event history data for enrollment are available for all years relevant to the analysis – 1994 through 2009. The NLSY79 event history data for enrollment are available from a respondent’s 1980 interview onwards. For months prior to 1980, we utilize enrollment information from the 1979 and 1980 interviews to fill in the event histories ourselves.¹ To ensure that we do not count students who are on summer vacation as not enrolled, we only consider respondents to be not enrolled if the month in which they report being not in school is a member of a set of four or more months of non-enrollment.

To measure employment, we use the event history employment variables in the NLSY. These variables report employment status for each respondent by week. We consider a respondent to be employed in a month if that individual reports “working,” being “associated with an employer,” or being in “active military service” for at least one week in that month. We first aggregate weekly data to be monthly, considering an individual employed if the respondent worked for at least one week in the month. Data for weeks that overlap two months are prorated based on how many days in the week occurred in each month.²

We also obtain information on co-residence with a spouse and ASVAB/AFQT composite scores. In NLSY97, we define living with a spouse using the NLSY cohabitation event history variables. In NLSY79, where these event histories were not available, we instead define living with a spouse to mean married and not separated. ASVAB/AFQT composite scores are based on the ASVAB tests given to most respondents in the two surveys. The test is meant to measure aptitude in a number of basic skills such as math and reading comprehension. The ASVAB composite is created by the Bureau of Labor Statistics and included in NLSY97. The AFQT score is a composite score created by the Department of Defense and available in NLSY79. The two composite scores are calculated in similar ways.

Finally, we obtain information on parent education and government aid receipt. We obtain parents’ educational background from the parent retrospective survey and define an individual’s parental education level as the highest grade completed by either parent.³ The government aid variable identifies whether a respondent’s parent received any form of government aid - including AFDC, SSI, Medicaid, and food aid - from the time they were 18 or their first child was born (whichever was sooner) until 1997. This is only available in NLSY97.

2.1.2. Sample restrictions

We make several sample restrictions. In our estimation results using NLSY97, we drop any individual who attrits by the survey’s 13th wave. We use both the main sample and the Black and Hispanic supplementary samples. In estimation results using NLSY79, we drop respondents who attrit by 1990 (when the youngest cohort will be 25). We exploit the nationally-representative cross-section and the supplementary samples of Blacks and Hispanics, but we exclude the “poor white” and the military sub-samples of the NLSY79. As noted by MaCurdy et al. (1998), the “poor white” supplementary sample was discontinued in 1991, partly because survey designers questioned the representativeness of the sample. Few members of the military sample were interviewed after 1983. For these reasons, we do not consider either group in the analysis. In both the NLSY97 and NLSY79 samples, we remove youth for whom we do not have data available in the first quarter of age 15, as well as youth with missing data on any of the key variables needed for the analysis.

These sample restrictions leave us with 5,056 individuals in the NLSY97 sample and 1,899 individuals in the NLSY79 sample. Details on the sample exclusions are presented in Tables 1 and 2 for the NLSY97 and NLSY79 cohorts, respectively. For each survey, we use panel weights from the latest round in our data (2009 for NLSY97, and 1990 for NLSY79) to correct for attrition/non-response and the use of the supplementary samples.

Table 1.

Data Exclusions to Produce Final Sample (NLSY97 Cohort)

Subgroup	Initial Count	Remaining Count After Removing Youth Who:
		(1) Have Attrition or Non-Response	(2) Have No Age 15Q1 Data	(3) Have Missing Data	(4) Have Disconnected by Age 15	(5) Never Disconnect
All	8,984	5,383	5,194	5,056	5,018	1,367
Black	2,333	1,432	1,373	1,327	1,317	537
White	4,752	2,856	2,792	2,735	2,714	499
Hispanic	1,899	1,095	1,029	994	987	331
Male	4,599	2,547	2,452	2,374	2,356	548
Female	4,385	2,836	2,742	2,682	2,662	819
Received Government Aid	3,923	2,469	2,395	2,318	2,293	859
Non High School Grad Parent	1,835	1,026	837	799	789	381
High School Grad Parent	2,923	1,727	1,727	1,664	1,648	556
Some College Parent	2,110	1,265	1,265	1,241	1,234	263
College Grad Parent	1,167	724	724	718	714	102
Advanced Degree Parent	949	641	641	634	633	65
Non-Urban	2,410	1,498	1,464	1,426	1,415	340

NOTE: Each cell shows the count remaining after the removal indicated in the column. Column (4) shows the final set of individuals used in estimation and Column (5) shows the number who ever disconnect during our sample, using the “N” definition.

Table 2.

Data Exclusions to Produce Final Sample (NLSY79 Cohort)

Subgroup	Initial Count	Remaining Count After Removing Youth Who:
		(1) Have Attrition or Non-Response	(2) Are in White/Military Oversample	(3) Have No Age 15Q1 Data	(4) Have Missing Data	(5) Have Disconnected by Age 15	(6) Never Disconnect
All	12,686	10,436	8,770	2,042	1,899	1,883	710
Black	3,174	2,719	2,646	592	557	554	299
White	7,510	5,988	4,428	1,039	976	969	259
Hispanic	2,002	1,729	1,696	411	366	360	152
Male	6,403	5,112	4,280	1,041	965	958	280
Female	6,283	5,324	4,490	1,001	934	925	430
Non-High School Grad Parent	4,545	3,856	3,203	683	611	599	328
High School Grad Parent	4,844	3,935	3,308	817	764	761	269
Some College Parent	1,432	1,154	979	240	231	230	67
College Grad Parent	1,119	871	753	174	167	167	28
Advanced Degree Parent	746	620	527	128	126	126	18
Non-Urban	2,676	2,197	1,793	415	386	383	132

NOTE: Each cell shows the count remaining after the removal indicated in the column. Column (5) shows the final set of individuals used in estimation and Column (6) shows the number who ever disconnect during our sample, using the “N” definition.

The panel weights are designed to allow for nationally representative estimates when using the populations represented in Column (1) of Table 1 and Column (2) of Table 2. Our estimates of the duration model presented below use as a starting point the sample in Column (3) of Table 1 and Column (4) of Table 2. To the extent that the sample exclusions between these columns disproportionally remove certain groups, the use of the panel weights will no longer yield nationally representative estimates when used in our duration model. To investigate if this is a concern, we use the counts in Table 1 and Table 2 to examine the proportion of each sub-group within each column, comparing Columns (1) and (3) for Table 1 and Columns (2) and (4) for Table 2. Fortunately, we find that the composition of each sub-group shown in the tables is stable when comparing these columns. For NLSY97, for all but one sub-group, the proportions in Table 1 Columns (1) and (3) are within 1.05 percentage points of each other; for the one exception (non-high school graduate parents), the difference is 3.26 percentage points. For NLSY79, for all but two sub-groups, the proportions in Table 2 Columns (2) and (4) are within 2.02 percentage points; for the two exceptions (non-high school graduate parents and high school graduate parents), the differences are 4.35 percentage points and 2.51 percentage points, respectively. While it is possible that differences would be larger among other observable characteristics, these results suggest that the use of panel weights will reflect a population that is approximately nationally representative, though tilted slightly to higher parental education levels than the true overall population in each time period.

2.2. Concept of a disconnection spell and connection spell

In this section, we define a “disconnected month” and “connected month,” and then leverage this definition into a concrete definition of “disconnection spells” and “connection spells,” which we interchangeably refer to also as disconnection and connection episodes. Our focus in most of the remainder of the paper will be on disconnection and connection spells.

We principally consider two notions of disconnected months: (1) months when a youth is not working and not in school (the “N” definition everywhere below); and (2) months when a youth is not working, not in school, and not living with a spouse (the “S” definition). The second notion is important because policy makers may wish to distinguish the circumstances of youth living with a spouse, for instance, from those higher-risk youth who experience disconnection from all means of support, including school, work, and close family. Below, we also describe how profiles would change if one were to narrow definition (2) to exclude from the disconnection pool those youth cohabiting with a partner (the “C” definition). Ideally, we would also consider youths’ residence with parents or close relatives. However, it is not possible to create credible month-by-month histories of residence with parental figures with the NLSY data. Any month that is not a disconnected month is defined as a connected month.

An occasional month of disconnection in a youth’s trajectory is not worrisome, but a more sustained spell of disconnection is more troubling. Therefore, we focus on such sustained spells in our analysis. To do so, we require a clear definition of when a disconnection spell (or disconnection episode) begins and ends. We define that an individual enters a disconnection spell in a given month if that individual is both disconnected in that month (according to the “N,” “S,” or “C” notions of disconnected months above) and will be disconnected in at least 8 of the following 11 months. Similarly, we define that an individual enters a connection spell (or connection episode) in a given month if that individual is both connected in that month and will be connected for at least 3 of the following 11 months. It is possible that in any given month, neither the disconnection spell nor the connection spell criterion is satisfied. For instance, if the current month is connected and 9 or more of the 11 subsequent months are disconnected, neither criterion is satisfied for the current month.

For the duration model, an initial disconnection spell begins when the disconnection spell criterion is satisfied and continues until the connection spell criterion is satisfied. After the initial disconnection spell starts, there may be a long sequence of months where the person is going back and forth between disconnected and connected months and neither criterion is satisfied. Alternatively, there may be many months in which the disconnection spell criterion is satisfied. Either way, the disconnection spell continues throughout this time. It continues until the person reaches a month in which the connection spell criterion is satisfied. At that point, a reconnection spell is initiated. This means that once a disconnection spell is initiated, it does not necessarily last 12 months – it could last just a handful of months or it could last many more than 12 months depending on the sequence of disconnected and connected months following initiation of the spell.

Similarly, for the duration model, a reconnection spell begins when the connection spell criterion is satisfied and continues until the disconnection spell criterion is satisfied. After initiation of the reconnection spell, the person may go through a sequence of months where neither criterion is satisfied, as they oscillate back and forth between connected months and disconnected months. They may also go through months where the reconnection spell criterion is satisfied. Either way, the reconnection spell continues until the person reaches a month in which the disconnection spell criterion is satisfied. At that point, a re-disconnection spell is initiated. Again, once a reconnection spell is initiated, it does not necessarily last 12 months – it could last just a handful of months or it could last many more than 12 months depending on the sequence of disconnected and connected months following initiation of the spell.

Fig. 16 contains several examples to illustrate these points. Each example is structured analogously. The top row indicates months in a person’s longitudinal history (months 1–24). The “Connected” and “Disconnected” rows have zeroes and ones to indicate if the person’s month is connected or disconnected – a one in the connected row indicates the month is connected, and a one in the disconnected row indicates the month is disconnected. The “Meets Connection Spell Criteria” and “Meets Disconnection Spell Criteria” rows have zeroes and ones to indicate if a month meets each of the criteria, with a one indicating that the criteria are met. Gray cells with italicized indicators in these rows indicate the start of a spell. Finally, the bottom row shows the implications for the start point and length of the initial disconnection spell, reconnection spell, and re-disconnection spell. Note that for simplicity, all three examples assume that all months after month 24 are disconnected for the person.

Fig. 16.

Examples of disconnection and connection spells.

To summarize the examples:

• Example 1 shows a simple case. Months 1–7 are disconnected, months 8–10 have a mixture of connected and disconnected months, months 11–14 are connected, and months 15 onwards are disconnected. Only month 1 satisfies the disconnection spell criterion out of months 1–7, but the initial disconnection spell beginning with month 1 continues until the start of the reconnection spell in month 8. The reconnection spell begins in month 8 and continues until the start of the re-disconnection spell in month 15.

• Example 2 shows a case where a reconnection spell begins very soon after the initial disconnection spell starts. While there is a relationship between connected and disconnected months, nothing mechanical forces this to happen. Alternative sequences could cause the reconnection spell to start sooner or later than this example. Rapid reconnection, as in this case, can happen when the disconnection spell is not very intense. This is exactly what we want out of our definition.

• Example 3 shows a case where a re-disconnection spell begins very soon after the reconnection spell starts. Again, nothing mechanical forces this timing. Alternative sequences could cause the re-disconnection spell to start sooner or later than this example. Rapid re-disconnection, as in this case, can happen when the reconnection is tenuous. Again, this is a desirable implication of our definitions.

As our definitions and the examples together suggest, our approach purposely limits the extent to which individuals will be assigned rapid, spurious changes in spell status based on scattered disconnected or connected months that are not reflective of their fundamental experience. Under our approach, a person can initiate a reconnection spell immediately after starting a disconnection spell only if the disconnection spell does not have an intense pattern of disconnected months; a person can initiate a re-disconnection spell immediately after starting a reconnection spell only if the reconnection spell is characterized by a tenuous pattern of reconnected months.

For most of the paper we focus on quarters rather than months. We divide years into four quarters of three months each and collapse the monthly data into quarters. The core reason for this choice is that the use of quarters can speed the empirical computations. This was useful primarily in the long series of initial investigations where we conducted sensitivity analyses for alternative definitions of disconnection episodes and alternative specifications of the duration dependence in the hazard models. Less importantly, this also helps simplify some of the tables and figures and helps reduce the number of observations that must be dropped due to missing values (since a stray month, but not an entire quarter, may have missing information). An individual is considered to begin a disconnection spell in a quarter if a spell begins in at least one month in that quarter. We use an analogous aggregation rule for connection spells. If in the same quarter an individual has one month where they have started a disconnection spell and one month where they have started a connection spell, we consider that individual to be starting a disconnection spell in that quarter; this is a rare situation. Early analyses compared the use of months to the use of quarters and found that patterns of disconnection remained qualitatively similar.⁴

2.3. Rationale for a statistical model specifying spells of disconnection and reconnection

The simplest approach to describing the disconnection trajectories of youth using the NLSY97 and NLSY79 cohorts would be to simply provide summary statistics on disconnection experiences – e.g., rates of entry into disconnection and reconnection spells by a certain age, distributions of the length of spells, etc. Instead, in this paper we apply a multi-state duration model to describe youths’ experiences. The use of the empirical model is necessary for three reasons.

First, the model allows for the assessment of the effects of individual characteristics like race holding other factors constant, even when sub-group sample sizes are small. Tables 1 and 2 illustrate the small sample sizes within key sub-groups in the NLSY97 and NLSY79 cohorts. Column (4) of Table 1 shows the number of NLSY97 youth overall and by sub-group after all core sample exclusions. Column (5) shows the numbers after removing those who never have a disconnection spell (the set of youth for whom questions of reconnection and re-disconnection are relevant). Columns (5) and (6) of Table 2 are fully analogous, except for the NLSY79 cohort. Given the sample sizes here, a non-parametric analysis that simply leverages the richness of the data would not be able to examine effectively the factors that affect variation in disconnection experiences, even for initial disconnection. A duration model allows for this type of examination (with the tradeoff, of course, of the modeling assumptions). The model allows us to understand important factors in the underlying hazard rates for each event of interest, and also provides a statistical framework with which to simulate and compare the complete trajectories of youth from differing backgrounds up to age 23 (see Section 5). Given the starkly limited sample sizes, such comparisons would not be possible without a model.

Second, a model allows for a cleaner understanding of the timing of disconnection, reconnection, and re-disconnection. Simple plots of raw (i.e. Kaplan-Meier) hazards are informative, but very volatile. In Section 4, we show that the Kaplan-Meier hazards for initial disconnection spells for males and females reflect a great deal of noise. This would only be accentuated if one were examining initial disconnection for any particular sub-group or examining reconnection and re-disconnection for any population. A duration model helps smooth these patterns into more easily interpretable results, including graphs displaying cumulative failure/success rates for disconnection/reconnection and simulations summarizing youths’ trajectories comprehensively. Using a simpler approach, such as moving averages of Kaplan-Meier hazards, could be sufficient to understand the time patterns for the overall population, but may not be sufficient to examine key sub-groups given the sample sizes in Tables 1 and 2.

Third, although we can get an unbiased estimate of whether a youth has entered an initial disconnection spell by age 23 without a model, we have right censoring issues when estimating the length of the initial disconnection spell (i.e., time until reconnection) and the length of a reconnection spell (i.e. time until re-disconnection). This is because someone may happen to be in the middle of a disconnection or reconnection spell in the last time period in our final estimation sample (which we have selected to be age 23 for all youth). We could get around this by picking an arbitrary early age (e.g. 19), estimating the length of disconnection/reconnection spells and incidence of reconnection/re-disconnection only for those youth who initially disconnect by that age, and restricting our interpretation of the results to just those youth. However, we are fundamentally interested in understanding the full trajectory of initial disconnection, reconnection, and re-disconnection for all youth in our sample through their early 20s.

2.4. Formulation of statistical model specifying spells of disconnection and reconnection

With these considerations in mind, we apply a multi-state duration model to estimate duration distributions for the time until a disconnection spell, the time until reconnection after disconnecting, and the time until re-disconnecting after reconnecting. We leverage the approach discussed in MaCurdy et al. (2010). As they note, the approach shares commonalities with more standard competing risks models, but has certain practical advantages. Most notably, the approach allows for modeling the various outcomes in a modular manner, rather than standard models that jointly specify and estimate all risk branches. The approach also provides a convenient entry point to modeling non-monotonic duration dependence, which is present in our context.

2.4.1. Model specification

Specifically, the multi-state duration model utilizes a series of discrete-time models with distinct conditioning events (or statuses), where a duration distribution characterizes the likelihood that an individual experiences a given number of quarters of continuous residence in a particular status, given admission into this status and some specification of history prior to the start of the spell. Letting “$i$” designate an arbitrary status, the probability of experiencing exactly $\tau$ quarters in that status takes the form

$$f_i\left(\mathbf{\tau }\right)={\mathit{S}}_{\mathit{i}}(\mathbf{\tau }-1){\mathit{P}}_{\mathit{i}}\left(\mathbf{\tau },{\mathit{Z}}_{\mathit{i}}\right)$$ (2.1)

with

$${\mathit{S}}_{\mathit{i}}(\mathbf{\tau }-1)=\prod _{\mathit{t}=1}^{\mathbf{\tau }-1}\left[1-{\mathit{P}}_{\mathit{i}}\left(\mathit{t},{\mathit{Z}}_{\mathit{i}}\right)\right]$$ (2.2)

The hazard rate $P_i(t,Z)$ designates the likelihood that an individual leaves the current status in the $t-th$ quarter after already having been in the state for $t-1$ quarters. The hazard rate is a function of the time $t$, as well as covariates $Z$ summarizing the demographic characteristics and history at the start of the current episode. In the hazard rate for initial disconnection, $t$ corresponds to age in quarters. When we estimate the duration to an event conditional on another event $K$ – for example, the duration to a connection spell conditional on having entered a disconnection spell – the probabilities are conditioned on the event $K$, and $t$ corresponds to time since that event began.

Estimation can proceed in modular fashion with separate estimation steps for (i) the probability of entering an initial disconnection spell, (ii) the probability of entering a reconnection spell conditional on initial disconnection, and (iii) the probability of entering a re-disconnection spell conditional on reconnection. In each step, we simply subset the data appropriately to those youth who experience the conditioning event. In the final estimation sample, we use data on all respondents up to age 23.

Once the hazard rates associated with duration in every status are obtained, one can construct the probability of any combination of disconnection and connection spell histories and various functions of those histories (e.g. average length of time in a reconnection spell). Therefore, the modular nature of our multi-state duration model allows us to fine-tune model fit separately for each of steps (i)-(iii), while the ultimate outputs can be used to construct the complete trajectories of initial disconnection, re-connection, and re-disconnection.

The empirical specification requires an appropriate functional form for the hazard rates determining exit from each state. Graphs of the Kaplan-Meier hazard rates in our data reveal that empirical specifications of the probabilities $P_i(t,Z)$ must admit non-monotonic duration dependence, which standard empirical specifications typically do not accommodate. We specify the following logistic model for the probabilities $P_i(t,Z)$:

$${\mathit{P}}_{\mathit{i}}(\mathit{t},\mathit{Z})=\frac{1}{1+{\mathit{e}}^{-\mathit{Z}{\mathbf{\beta }}_{\mathit{i}}-{\mathit{g}}_{\mathit{i}}\left(\mathit{t},{\mathbf{\alpha }}_{\mathit{i}}\right)}}$$ (2.3)

where ${\beta }_i$ is a parameter vector, $Z$ is a vector of covariates, and the function $g_i\left(t,{\alpha }_i\right)$ determines the duration properties associated with the spell.

2.4.2. Covariates and duration dependence

In the logit models used to estimate hazard rates for initial disconnection, we include the following covariates (Z): (i) race dummy variables; (ii) a set of dummy variables to indicate the maximum educational attainment of the respondent’s parents; (iii) a set of variables that interact ASVAB scores with a dummy indicating the cohort of the respondent;⁵ and (iv) a government aid variable.⁶ We estimate separate models for males and females. In the logit models for reconnection and re-disconnection, a relatively small number of youth have the conditioning event (see last columns of Tables 1 and 2), and so we pool males and females and include race/sex interactions. We also include dummy variables that indicate whether the individual has had certain experiences, such as child birth, dropping out of school, or a criminal conviction, before the start of the current spell.⁷

We model the function $g_i\left(t,{\alpha }_i\right)$ using a novel construction of spline terms that are distinct for initial disconnection, reconnection, and re-disconnection, and designed to fit the time patterns in the Kaplan-Meier hazard rates. Implicit in conventional spline models, which fit polynomial functions to a series of intervals over duration, is a tradeoff between smoothness and goodness of fit. Fit can be improved by increasing the number of polynomial functions, but non-differentiability at the boundaries requires a sacrifice in smoothness. Limiting the number of intervals yields a smoother curve but diminishes the capabilities of detecting complicated forms of duration dependence.

In view of this challenge, we instead specify $g_i\left(t,{\alpha }_i\right)$ as the general smooth function:

$${\mathit{g}}_{\mathit{i}}\left(\mathit{t},{\mathbf{\alpha }}_{\mathit{i}}\right)=\sum _{\mathit{j}=1}^{\mathit{J}}\left[{\mathbf{\Phi }}_{\mathbf{i}\mathbf{j}-1}(\mathbf{t})-{\mathbf{\Phi }}_{\mathbf{i}\mathbf{j}}(\mathbf{t})\right]\left[{\mathbf{\alpha }}_{\mathbf{i}\mathbf{j}}\right].$$ (2.4)

The quantity ${\Phi }_{ij}(t)$ denotes the cumulative distribution function ($cdf$) of a normal random variable possessing mean ${\mathrm{\mu }}_{ij}$ and variance ${\sigma }_{ij}^2$. The presence of the $cdf’\mathrm{s}$ in (2.4) permits us to incorporate spline features in $g$ so that the parameter ${\alpha }_{ij}$ represents $g$ over only a specified range of $t$.⁸ With the values of the ${\mu }_{ij}$ and the ${\sigma }_{ij}$ set in advance of estimation, $g_i\left(t,{\alpha }_i\right)$ is strictly linear in the parameters $\alpha$. One can control where each spline or polynomial begins and ends by adjusting the values of the $\mu$’s. One can also control how quickly each spline cuts in and out by adjusting the values of the $\sigma$’s, with higher values providing for a more gradual and smoother transition from one $cdf$ to the next.

We choose the parameters of the spline terms such that the fitted hazard rates capture the movements in the Kaplan-Meier hazard rates. In our estimates of the duration until initial disconnection spell, we make the following choices for the spline terms: We set $J=3,{\mathrm{\mu }}_{\mathrm{i}1}=8$ (quarter corresponding to age 15), ${\mathrm{\mu }}_{\mathrm{i}2}=20$ (age 18), and ${\mathrm{\mu }}_{\mathrm{i}3}=28$ (age 20), with the standard deviations ${\mathrm{\sigma }}_{\mathrm{i}1},{\mathrm{\sigma }}_{\mathrm{i}2}$, and ${\mathrm{\sigma }}_{\mathrm{i}3}$ set at 3, 1.5, and 1.5 respectively. We set ${\mathrm{\mu }}_{\mathrm{i}0}$ equal to minus infinity, so that ${\mathrm{\Phi }}_{\mathrm{i}0}(\mathrm{t})=1$ for all $t$, regardless of ${\mathrm{\sigma }}_{\mathrm{i}0}$. For reconnection and re-disconnection, we include a set of three splines, where $J=2,{\mathrm{\mu }}_{\mathrm{i}0}$ equals minus infinity, ${\mathrm{\mu }}_{\mathrm{i}1}=4$ (1 year), ${\mathrm{\mu }}_{\mathrm{i}2}=12$ (3 years), and ${\sigma }_{i1}={\sigma }_{i2}=2$. These choices are common to the NLSY97 and NLSY79 analyses.

As explained thus far, our analysis examines factors affecting the disconnection of those 15 years of age and older. However, some individuals may have disconnected prior to age 15. In order to take account of this, we specify that $P_i(1,Z)=P^{\mathrm{*}}(Z)$ in Eq. (2.2), where $P^{\mathrm{*}}(Z)$ is the probability of having disconnected prior to age 15. To estimate $p^{\mathrm{*}}(Z)$ we first restrict the sample to individuals we observe in the data from at least age 14 onwards. We then calculate the proportion of this sample who disconnects prior to age 15, for males and females separately. For simplicity and because of limited sample size, we assume $P^{\mathrm{*}}(Z)$ depends only on sex and not on other covariates. These estimates of pre-age 15 disconnection are quite small; for instance, for the NLSY97, less than 0.4% of youth are part of this group. When estimating $P_i(t,Z)$ for $t>1$ (i.e., for the first quarter of age 15 and beyond), we need to eliminate those who disconnected prior to age 15. In practice, we implement this by using all individuals who we observe at the fourth quarter of age 14 or before, and who have not disconnected before age 15 in the time we observe them.⁹ It is possible there is some left censoring, in that someone may have disconnected prior to the fourth quarter of age 14 but that earlier quarter is not included in the NLSY data for that person; however, our estimates indicate this is likely to be extremely rare.

3. Assessing our definition of disconnection spells

In this section, we show our notion of “disconnection spells” captures the youth with the most worrisome experiences. We begin by establishing that our definitions of “disconnected month” underlying the construction of disconnection spells yield prevalence rates that are consistent with prior literature on cross-sectional prevalence and correlate with youth characteristics as expected. Table 3 presents the shares of youth experiencing various amounts of disconnected months by ages 20 and 22 (note that this table does not include any information about disconnection spells and does not address right censoring).¹⁰ The first two columns of the table report results using a definition of disconnected month in which youths are not working, not enrolled, and not living with a spouse (“S” definition); the second two columns consider youths to be disconnected if they are not working and not enrolled (“N” definition). Overall, more than 20% of all youths in the NLSY97 cohort accumulate at least a year of not working and not being enrolled by age 22. Groups expected to be at higher risk for worse outcomes have higher numbers of disconnected months. For example, four high-risk groups – teen mothers, high-school dropouts, youths who have been convicted of a crime, and youths who have spent time not living with parents – experience much higher accumulation of disconnected months than other youth.¹¹

Table 3.

Share of Population Experiencing Some Disconnection by Ages 20 and 22 (NLSY97 Cohort)

	“S” Disconnection		“N” Disconnection
	(cumulative number of months, ages 13–23)
	≥12 months	≥24 months	≥12 months	≥24 months
By Age 20
All	11.3	3.5	12.2	3.9
Male	10.0	3.0	10.1	3.0
Female	12.6	4.0	14.5	4.9
Black Male	20.8	7.8	20.8	7.8
Black Female	22.6	8.3	22.8	8.6
Hispanic	14.9	4.0	16.5	5.1
Teen Mother	48.5	20.5	55.6	26.8
Criminal Conviction	27.6	10.4	29.1	11.5
High School Dropout	48.4	23.1	51.9	26.3
Foster Child	22.1	7.8	25.7	9.9
By Age 22
All	18.3	8.2	20.1	9.7
Male	17.0	7.4	17.2	7.4
Female	19.6	9.0	23.0	11.9
Black Male	33.7	19.3	34.0	19.3
Black Female	37.2	20.2	38.8	21.4
Hispanic	23.1	9.4	26.4	12.4
Teen Mothers	59.9	34.3	67.2	44.9
Criminal Conviction	40.1	19.3	41.1	21.4
High School Dropout	59.4	35.3	63.4	40.5
Foster Child	32.4	15.4	37.5	19.9

NOTE: Each cell represents the fraction of youth in the row who have experienced at least the number of disconnected months given in the column heading, according to the “S” or “N” definition of disconnection.

While Table 3 describes accumulated months of disconnection, our analysis below focuses on the incidence of disconnection spells. The chief advantage of defining a disconnected spell is that it clearly defines what it means to be in a state of disconnection (rather than simply accumulating one-off months of unemployment or non-enrollment). One concern with this approach, however, is that we may miss some individuals who accumulate a substantial number of disconnected months without ever experiencing a concentrated spell of disconnection.

Appendix Tables 1 and 2 (referred to as A.1 and A.2) examine this possibility in detail. Table A.1 looks at the proportion of the sample that experiences zero spells of disconnection according to our “9 of 12″ definition, while accumulating at least 12, 18, and 24 total months of disconnection. Separate panels provide results for the “S” and “N” definitions, using all months in the data available from ages 13–23. Of individuals who experience at least 12 months of disconnection by age 22, only around 10% (under the “N” definition) or 11% (under the “S” definition) do not experience any spells of concentrated disconnection. These results suggest that our duration analysis that analyzes spells of disconnection picks up approximately 90% of those individuals who have a substantial number of disconnected months.

Table A.2 looks at a similar question in a different way. For youth with 12 or more disconnected months by age 22, we compare the experiences of youth with zero disconnection spells (left panel) to youth with one or more spells. Very few youth with zero spells have more than 24 disconnected months in all, regardless of the definition. In analysis not shown here, we find that these youth tend to have a burst of disconnected months over a 5–8 month period, and then have several more disconnected months scattered over a very long period. In contrast, youth with at least one disconnection spell have more months of disconnection concentrated within a given time period, with over half having a sequence of 20 consecutive months, regardless of definition. These are exactly the type of youth that we are most interested in, suggesting that our definition of “spell” does not cause us to miss youth with very intense disconnection experiences.

Another potential concern about our definition of disconnection spell is that, while it captures those with intense bouts of unemployment and non-enrollment, it does not necessarily capture greater deprivation in a sense that researchers and policy makers would care about. Table 4 addresses this concern. Panel A summarizes the contacts of youths with various government agencies, with the left panel showing youth who had a disconnection spell and the right panel showing those who never experienced a disconnection spell during their time in the NLSY97 sample. The results reveal substantial use of public assistance among youth with at least one spell; for example, more than 56% participated in some government welfare during the period covered by the survey for the “S” definition of disconnection. Much of this participation came through WIC and Food Stamps. In contrast, 34% of the never-disconnected youth reported any involvement with the welfare system. Meyer and Mittag (2019) summarize a wide body of evidence demonstrating that survey data typically under-count program receipts relative to administrative data, and show evidence from the CPS that “missing” receipts account for a larger share of income among the poorest households. To the extent that program participation under-counting mimics program receipts, and that disconnected youth are predominantly from poorer households, the patterns here may understate the gap between ever-disconnected and never-disconnected youth. Patterns of juvenile justice involvement confirm the differences across the ever-disconnected and never-disconnected population; 40% of “S” disconnected youth were convicted of a crime at some point, compared with only 18% of the never-disconnected youth.¹² Panel B of Table 4 shows that much, though not all, of this contact with public assistance and juvenile justice occurs during and after the initial disconnection spell. These findings suggest that our definition of disconnection spell captures deprivation in multiple senses.

Table 4.

Percent of Youth Having Contact with Government Programs by Disconnection Status (NLSY97 Cohort)

Type of contact	Panel A
	Ever Disconnected			Never Disconnection
	N	S	C	N	S	C
Government Training	19.8%	20.5%	21.7%	15.3%	15.5%	15.5%
AFDC	10.9%	12.3%	12.3%	1.8%	2.0%	3.1%
Food Stamps	32.7%	34.0%	32.7%	9.5%	11.6%	14.5%
WIC	38.4%	36.7%	33.3%	17.1%	20.8%	23.9%
Other Welfare	11.0%	12.3%	12.2%	2.6%	2.9%	3.9%
Any Welfare	56.1%	56.0%	53.9%	31.1%	34.3%	37.4%
Workers Compensation	0.1%	0.1%	0.1%	0.5%	0.5%	0.4%
Unemployment Insurance	20.7%	20.8%	20.7%	14.6%	15.3%	16.0%
Arrest	36.8%	40.0%	40.7%	23.5%	23.2%	24.7%
Found Guilty	34.9%	39.7%	41.7%	19.3%	18.3%	19.7%
Charged	28.6%	31.3%	31.7%	18.9%	18.5%	19.6%

Type of contact	Panel B
	Prior to First Disconnection			During First Disconnection			During/After First Disconnection
	N	S	C	N	S	C	N	S	C
Government Training	10.4%	9.6%	10.1%	5.6%	6.3%	6.9%	11.1%	12.6%	13.8%
AFDC	3.1%	3.4%	4.1%	5.0%	5.9%	5.8%	9.1%	10.7%	10.7%
Food Stamps	10.8%	10.4%	9.6%	16.4%	16.6%	13.9%	29.0%	30.8%	30.0%
WIC	19.4%	17.2%	14.7%	22.8%	20.4%	14.9%	33.2%	32.3%	28.9%
Other Welfare	5.1%	5.5%	5.4%	5.2%	5.7%	5.2%	8.2%	9.4%	9.5%
Any Welfare	29.4%	27.8%	25.3%	34.8%	33.6%	28.4%	48.6%	49.5%	47.7%
Workers Compensation	0.1%	0.1%	0.1%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%
Unemployment Insurance	8.5%	7.8%	7.5%	7.3%	7.3%	6.9%	15.2%	16.1%	16.3%
Arrest	27.3%	29.0%	29.3%	30.9%	33.0%	33.1%	36.8%	40.0%	40.7%
Found Guilty	19.0%	20.8%	21.4%	24.8%	27.5%	28.2%	34.9%	39.7%	41.7%
Charged	19.8%	21.1%	21.1%	23.2%	24.8%	24.6%	28.6%	31.3%	31.7%

NOTE: Each cell represents the lower bound on the fraction of youth who had contact with training, welfare, and criminal justice as specified in the rows. In the top panel, denominators reflect those who experienced any disconnection spell or never experienced a disconnection spell according to the “N,” “S,” and “C” definitions. In the bottom panel, we sub-divide by timing of the program interaction.

We conclude that our definition of disconnection spell captures youth with intense bouts of disconnection and persistent challenges. In terms of understanding how our definition of disconnection spell compares to the existing literature, our results demonstrate that our notion of disconnection spell is capturing a narrower and more concerning population than much of the prior literature on disconnection. Table A.2, for example, shows that a substantial portion of all youth run into at least one period with 5–8 months of consecutive unemployment/non-enrollment, similar to the disconnection definition used by other authors. But examining the distribution of total disconnected months in Table A.2 reveals that this is likely a temporary phenomenon for those youth who never experience a disconnection spell by our definition. These and similar results suggest that our definition of disconnection spell captures an especially disadvantaged population relative to the prior literature.

4. Pervasiveness, persistence, and timing of disconnection and connection spells in the NLSY97

This section investigates how many youth experience a disconnection spell, how long spells last, and when they typically start, end, and resume. The first sub-section discusses initial disconnection, while the second sub-section discusses reconnection and re-disconnection. We focus here on patterns for the overall, male, and female populations, and discuss more specific sub-groups in Section 5. While the results in this section are based on estimates of the hazard models for initial disconnection, reconnection, and re-disconnection, we leave the discussion of the coefficients to Section 5.

4.1. Initial spells of disconnection

We begin by analyzing when youths experience their first episode of disconnection. Figs. 1 and 2 present the raw (“Kaplan-Meier”) hazard rates and fitted hazard rates from our duration model for initial disconnection, for men and women, respectively. We calculate the curves for fitted hazard rates by taking the weighted average of predicted hazard rates across the relevant sample. The fitted hazards correspond closely to the Kaplan-Meier hazard estimates, illustrating that our model captures duration dependence well. The hazard rates for both men and women have a prominent peak between ages 18 and 19, corresponding to the age at which most teenagers conclude high school. The hazard rates for the “N” definition of disconnection and those for the “S” definition track closely for both men and women, with a slightly more prominent gap for women than men. Small differences between men and women here and below should be interpreted with caution, as we do not present formal significance tests of the difference in the patterns between men and women. It is possible to construct a formal test using either the delta method or bootstrapping.

Fig. 1.

Raw and fitted hazard rates for disconnection for young men (NLSY97 Cohort)

NOTE: raw hazard rates use the standard Kaplan-Meier calculation. Fitted hazard rates are based on the estimates in Table 5.

Fig. 2.

Raw and fitted hazard rates for disconnection for young women (NLSY97 Cohort)

NOTE: raw hazard rates use the standard Kaplan-Meier calculation. Fitted hazard rates are based on the estimates in Table 5.

Fig. 3 addresses the question of how pervasive disconnection is in the population during the period covered by the NLSY97 cohort. Specifically, the figure examines the share of men/women predicted to have ever disconnected under the “N” or “S” definition by each age. This share – the “failure rate” – is constructed by first calculating every youth’s probability of reaching a particular age without disconnecting (the “survivor function”). The failure rate is then the weighted average across all relevant youths of one minus the survivor function at each age. By age 22, approximately 18% of males and 23% of females have experienced an episode of disconnection according to the “N” definition. This number falls by roughly 0.5 and 4 percentage points for men and women, respectively, when using the “S” definition. We estimate that by age 23, between 19 and 25% of all youth will have experienced a disconnection spell by either definition.

Fig. 3.

Probability of experiencing a disconnection episode by particular age (NLSY97 Cohort)

NOTE: curves depict the fitted failure rate by each age — i.e., the probability that a youth has experienced a disconnection episode by that age, using coefficient estimates from Table 5.

One may wish to consider cohabitation with a partner to be equivalent to cohabitation with a spouse. To assess the impact of cohabitation, we repeat the analysis but add the “C” definition, in which an individual is disconnected if he/she is not working, not enrolled and not living with a spouse or partner. Appendix Figs. 1 and 2 compare the failure rates under the “N,” “S,” and “C” definitions for men and women, respectively. Results between the three definitions are not substantially different for males, with the “S” and “C” definitions essentially indistinguishable. However, for females, the cumulative probability of having experienced a disconnection spell under the “C” definition by age 23 falls by almost 5 percentage points relative to the “S” definition and 9 percentage points relative to the “N” definition. Keeping in mind the caveat about statistical significance from above, this result demonstrates the importance of cohabitation for females during spells of unemployment and non-enrollment.

Our estimates of the share of youth who have experienced a disconnection spell by their early 20s are as high or higher than estimates of disconnection common to the prior literature, despite the fact that our definition of disconnection is more stringent and that we are examining a representative population rather than a particularly vulnerable sub-population. For example, the point-in-time estimates in Lewis and Gluskin (2018) and Loprest and Nichols (2011) are similar or lower, reflecting the fact that we are able to examine the portion of the population that ever experienced disconnection rather than a point in time. Hair et al. (2009), who use the same longitudinal concept as us but do not use a duration model, have similar or slightly lower estimates likely due to right censoring. Our estimates are also the first, to our knowledge, to illustrate the striking role of cohabitation in reducing these substantial disconnection rates for young women.

4.2. Reconnection and re-disconnection

Two statistical phenomena determine the amount of time a youth spends disconnected: the lengths of spells, and the tendency of a youth who reconnects to re-experience another disconnection episode. To assess the tendency of youths to reconnect and re-disconnect, we estimate the conditional hazard functions described in Section 2.4.1. For reconnection, we condition on entering a first spell of disconnection and examine duration from that point to reconnection. For re-disconnection, we condition on entering a connection spell and examine duration from that point to re-disconnection. We use the covariates summarized in Section 2.4.2.¹³ In order to present separate estimates for males and females, we show the weighted mean of the fitted values and failure/success rates conditional on sex.¹⁴

Figs. 4 and 5 plot the fitted hazard rates for reconnection and the success rates for reconnection (i.e. the probability of having entered a reconnection spell by a particular quarter after starting a disconnection spell) for each quarter from initial disconnection. The figures present results for males and females, and for both the “N” and “S” definitions. The fitted hazard rate in Fig. 4 shows a rapid jump from quarter 0 to quarter 1 – this is mechanical, since no youth can start a reconnection spell in the same quarter that they have started a disconnection spell, by construction. More generally, the forward-looking definition of disconnection and reconnection spells limits the size of hazard rates in the quarters immediately after disconnection (and reduces the variability of the hazards across quarters). Fig. 4 demonstrates that reconnection rates peak a little less than two years after the individual initially begins a spell of disconnection. From Fig. 5, we see that by 6 quarters after disconnection, between 46% and 52% of individuals have already reconnected (depending on the definition). We estimate that more than 29% of youth have not reconnected two years after initial disconnection and 10–14% have not reconnected by four years after initial disconnection.

Fig. 4.

Fitted hazard rate for reconnection of young men and women (NLSY97 Cohort)

NOTE: Curves depict the fitted hazard rate predicted using the coefficient estimates from Table 5.

Fig. 5.

Probability of experiencing a reconnection episode by quarter after initial disconnection (NLSY97 Cohort)

NOTE: Curves depict the fitted success rate for reconnection — i.e., the probability of re-connecting by quarter after initial disconnection, for those who disconnected. Fitted probabilities calculated using estimates in Table 5.

The second component of the amount of time spent in a disconnected state involves the likelihood of re-disconnection. Fig. 6 illustrates the failure functions for a second spell of disconnection (given space constraints, we do not show the underlying hazard rates). A large percentage of youth who experienced a disconnection episode earlier in life are susceptible to a spell of renewed disconnection. Approximately 24–32% of youth experience a second episode of disconnection within the first year and a half after reconnecting, depending on the disconnection definition and sex. Overall, about 56% of women and 48% of men who reconnect experience another spell of disconnection within four years according to the “N” definition. Using the “S” definition, the incidence of second spells of disconnection is lower by 2 to 10 percentage points as of quarter 8 onwards, but still substantial.

Fig. 6.

Probability of experiencing a second disconnection episode by quarter after reconnection (NLSY97 Cohort)

NOTE: Curves depict the fitted failure rate for re-disconnection — i.e., the probability of re-disconnecting by quarter after re-connection, for those who disconnected and then re-connected. Fitted probabilities calculated using estimates in Table 5.

Our results demonstrate substantial persistence in disconnection – a sizeable portion of youth who enter an initial disconnection spell fail to reconnect even multiple years after their initial disconnection, and more than a quarter of those who reconnect do not make it two years before disconnecting again. Hair et al. (2009) find that approximately two-thirds of NLSY97 youth who disconnect are able to reconnect again prior to the end of their study period in 2003. Since the end of the study period is different for every youth in their paper, this result is difficult to interpret. Our model provides an easier-to-interpret result – approximately two-thirds of youth have reconnected within two years of starting their initial disconnection spell under either the “N” or “S” definition. Loprest and Nichols (2011) find that approximately 42% of low income single mothers with a disconnection spell have a spell last 12 or more months. Our results instead suggest that more than 60% of youth with a disconnection spell have the spell last at least 12 months. While the difference in results could be related to our differing definitions and populations, a portion of the difference is also explainable by the fact that Loprest and Nichols (2011) only present numbers for those for whom they observe the start of the disconnection spell. Regarding re-disconnection, to our knowledge there has been no systematic prior study of this topic; our results demonstrate this is a crucial gap in understanding persistence.

5. Variation in disconnection experiences across sub-groups in the NLSY97

Thus far, we have examined the pervasiveness, persistence, and timing of disconnection experiences across the overall population of male and female youth using the NLSY97 cohort. Our empirical framework allows us to take the additional step of examining which sub-groups of youth are most likely to become and remain disconnected. We begin by discussing the coefficient estimates from the hazard models for initial disconnection, reconnection, and re-disconnection. We then translate these estimates into comprehensive profiles of disconnection experiences prior to age 23 for multiple sub-groups of interest using a novel simulation framework.

5.1. Coefficient estimates in hazard models for the NLSY97 cohort

Table 5 presents the coefficient estimates for the hazard models for initial disconnection (left panel), reconnection (middle panel), and re-disconnection (right panel) in the NLSY97 cohort. The table includes models for both the “N” and “S” definitions, indicated by the column headers. The reference group consists of females with parents who did not graduate high school and did not receive government aid, and who are neither Black nor Hispanic; the group of non-Black and non-Hispanic youth in this sample is predominantly White, and so we often refer to the group as White for simplicity. In general, few covariates carry statistically significant effects in the reconnection and re-disconnection hazards; in many case this is due to imprecision from limited sample sizes, but we highlight a handful of important cases where magnitudes of coefficients are actually close to zero. Appendix Tables 3–6 present a series of robustness analyses that explore the effects of using alternative covariates in all these models.

Table 5.

Coefficient Estimates in Hazard Models for Disconnection, Re-Connection, and Re-Disconnection (NLSY97 Cohort)

	Disconnection				Re-Connection		Re-Disconnection
	Males		Females		N (5)	S (6)	N (7)	S (8)
Covariates	N (1)	S (2)	N (3)	S (4)
Male					0.1609 (0.1082)	0.0228 (0.1134)	−0.2794* (0.1413)	−0.0670 (0.1642)
Black	0.5493*** (0.1228)	0.5337*** (0.1232)	−0.0520 (0.1074)	0.1304 (0.1172)	−0.1361 (0.1119)	−0.2975* (0.1186)	−0.1039 (0.1328)	0.2012 (0.1503)
Hispanic	0.1257 (0.1552)	0.0539 (0.1575)	−0.1617 (0.1223)	−0.2267+ (0.1377)	0.1502 (0.1238)	0.0676 (0.1366)	−0.3938* (0.1634)	−0.3876* (0.1908)
Black × Male					−0.0345 (0.1586)	0.0902 (0.1633)	0.2835 (0.1936)	0.0900 (0.2131)
Hispanic × Male					0.2057 (0.1864)	0.2518 (0.1986)	0.0382 (0.2459)	0.1339 (0.2773)
1st Disconnected, Age 17					−0.0293 (0.1602)	−0.0413 (0.1627)	−0.0880 (0.1694)	−0.0716 (0.1811)
1st Disconnected, Age 18					0.1594 (0.1441)	0.0559 (0.1474)	−0.1144 (0.1604)	−0.0502 (0.1710)
1st Disconnected, Age 19+					−0.4339** (0.1429)	−0.4156** (0.1462)	−0.0950 (0.1481)	−0.1726 (0.1569)
HS Graduate, Parents	−0.4202** (0.1475)	−0.4199** (0.1486)	−0.4434*** (0.1208)	−0.4295*** (0.1295)	−0.0634 (0.1010)	−0.0673 (0.1042)	−0.0551 (0.1233)	−0.0191 (0.1304)
Some College, Parents	−0.6747*** (0.1744)	−0.7095*** (0.1767)	−0.6287*** (0.1427)	−0.6814*** (0.1570)	−0.0642 (0.1178)	0.0019 (0.1230)	−0.2170 (0.1420)	−0.2890+ (0.1592)
College Graduate, Parents	−0.5025* (0.2295)	−0.4807* (0.2308)	−0.8333*** (0.2046)	−0.8322*** (0.2270)	−0.0149 (0.1536)	−0.0340 (0.1596)	−0.3466+ (0.1994)	−0.2844 (0.2209)
Advanced Degree, Parents	−0.5190* (0.2437)	−0.4807* (0.2440)	−0.8210*** (0.2207)	−0.9173*** (0.2521)	−0.1368 (0.1803)	−0.1845 (0.1900)	−0.1365 (0.2292)	0.0419 (0.2562)
ASVAB × Cohort 80	−0.0167*** (0.0036)	−0.0174*** (0.0037)	−0.0260*** (0.0032)	−0.0324*** (0.0038)	0.0056* (0.0024)	0.0063* (0.0026)	−0.0046 (0.0035)	−0.0076+ (0.0046)
ASVAB × Cohort 81	−0.0198*** (0.0038)	−0.0211*** (0.0038)	−0.0247*** (0.0032)	−0.0274*** (0.0037)	0.0012 (0.0023)	0.0015 (0.0024)	−0.0071* (0.0033)	−0.0116** (0.0042)
ASVAB × Cohort 82	−0.0148*** (0.0033)	−0.0155*** (0.0033)	−0.0245*** (0.0028)	−0.0263*** (0.0033)	0.0021 (0.0022)	0.0014 (0.0023)	−0.0022 (0.0029)	−0.0024 (0.0033)
ASVAB × Cohort 83	−0.0196*** (0.0036)	−0.0211*** (0.0037)	−0.0295*** (0.0034)	−0.0323*** (0.0039)	0.0006 (0.0024)	0.0037 (0.0025)	−0.0012 (0.0032)	−0.0016 (0.0037)
ASVAB × Cohort 84	−0.0151*** (0.0033)	−0.0161*** (0.0033)	−0.0257*** (0.0033)	−0.0247*** (0.0036)	0.0001 (0.0024)	−0.0013 (0.0024)	−0.0026 (0.0032)	−0.0035 (0.0036)
Gave Birth Prior to 1st Disc.					0.1713+ (0.0969)	0.2019+ (0.1040)	0.0970 (0.1126)	0.0690 (0.1273)
Drop Out Prior to 1st Disc.					−0.0977 (0.0988)	−0.1449 (0.1025)	0.1502 (0.1098)	0.1895 (0.1207)
Convicted Prior to 1st Disc.					0.0078 (0.1163)	−0.1056 (0.1175)	0.1838 (0.1375)	0.1359 (0.1504)
Gave Birth During 1st Disc.							0.2025+ (0.1185)	0.0883 (0.1323)
Received Government Aid, Parents	0.3728** (0.1198)	0.3790** (0.1212)	0.4602*** (0.1002)	0.4290*** (0.1078)	0.1248 (0.0823)	0.0660 (0.0860)	0.1650 (0.1039)	0.1337 (0.1157)
Observations	57,965	57,994	63,038	63,954	11,662	9,837	14,874	14,663

Columns 1, 3, 5, and 7 provide the estimates in hazard models for male disconnection, female disconnection, re-connection, and re-disconnection using the “N” definition. Together, they constitute the multi-state duration model for the “N” definition. Analogously, columns 2, 4, 6, and 8 constitute the multi-state duration model for the “S” definition. Robust standard errors in parentheses.

^***p < 0.001,

^**p < 0.01,

^*p < 0.05,

⁺p < 0.1.

Here, we briefly summarize the implications of the results for differences across sub-groups, focusing on ASVAB/AFQT, parental background, race, and age of initial disconnection in turn. In addition to using the statistical significance of coefficients directly from the tables, the summary below relies on a series of supplementary hypothesis tests. For initial disconnection, this includes tests that all ASVAB coefficients are equal, all ASVAB coefficients are equal to zero, all parental education coefficients are equal, all parental education coefficients are equal to zero, and the Black and Hispanic dummies are equal. For reconnection and re-disconnection, this includes all the tests for initial disconnection, plus tests that the Male and Black*Male coefficients sum to zero, that the Black and Black*Male coefficients sum to zero, that the Male and Hispanic*Male coefficients sum to zero, and that the Hispanic and Hispanic*Male coefficients sum to zero.

ASVAB scores are a consistent predictor of initial disconnection, but a weak predictor of reconnection and re-disconnection. For both the “N” and “S” definitions, higher scores are statistically significantly associated with a lower probability of an initial disconnection, and one can reject the null that they are jointly zero at the 1% level. There is mixed evidence that higher scores are associated with increased likelihood of reconnection and reduced tendency to re-disconnect, as statistical significance varies across cohorts. For reconnection using the “S” definition, one can reject the null that all ASVAB coefficients are jointly zero (10% level). For reconnection using the “N” definition and re-disconnection using either definition, at least one ASVAB coefficient is statistically significant and one cannot reject the hypothesis that all ASVAB coefficients are equal. Importantly, though, this is not a case of large coefficients with large standard errors; in general, the magnitude of ASVAB coefficients in the reconnection and re-disconnection specifications is very small and confidence intervals on the coefficients imply marginal effects are likely not practically significant. For example, in the reconnection “N” specifications, the coefficients that are not significant imply that a 10% increase in ASVAB score from the mean yields approximately a 0.2 percentage point increase in probability of reconnection (when evaluated at the value of the linear latent index consistent with the mean reconnection probability for men at 8 quarters after entering a disconnection spell). Given this and given the size of the standard errors around the coefficient estimates, this suggests that test scores actually predict very little after conditioning on initial disconnection.

Parental background plays an important role in initial disconnection, consistent with a wide array of studies on education, employment, and earnings. Regardless of degree, parental education at the high school graduate level or better is associated with a lower likelihood of initial disconnection. One can reject the null that all parental education covariates are jointly zero at the 1% level. For women, there is a clear ordering of effect sizes by degree level, whereas the story is more ambiguous for men; only for women, one can reject the null that all parental education coefficients are equal (at the 10% or 5% level, depending on disconnection definition). For both men and women, parental receipt of government aid is strongly predictive of greater likelihood of initial disconnection. For reconnection and re-disconnection, one fails to reject that the parental education coefficients and government aid coefficient are different from zero. For reconnection with parental education, this is again a case where the magnitudes of the coefficients are close to zero and this, in combination with the size of the standard errors, implies practically insignificant effects in any case. Examining the college graduate coefficient for the N definition, for instance, the coefficient implies a decline in the probability of reconnection of just 0.25 percentage points (again evaluated using mean reconnection probability 8 quarters after entering a disconnection spell as a reference point).

Race effects are distinctive by sex. Black males are statistically significantly more likely to initially disconnect than White males and Hispanic males (1% level). Except for re-disconnection under the “S” definition, Black males’ circumstances are statistically no worse than White males on reconnection and re-disconnection. The same is not generally true for females. Black females’ initial disconnection patterns are statistically no different than White females, and only different (worse) from Hispanic females for the “S” definition (1% level). For reconnection and re-disconnection, Black females are actually less likely to re-disconnect than White females, statistically significantly so under the “N” definition. However, they are less likely to reconnect (5% and 1% level, depending on definition) and more likely to re-disconnect (10% and 1% level, depending on definition) than Hispanic females.

Finally, experiencing one’s initial disconnection spell at age 19 or later is associated with a statistically significantly lower chance of reconnection in any given quarter (relative to the reference group of those disconnecting prior to age 17). The same is not true for re-disconnection.

5.2. Comprehensive profile of disconnection experiences for the NLSY97 cohort

In this section, we use a simulation framework to present a comprehensive profile of disconnection experiences for the NLSY97 cohort, one that shows the practical implications of the coefficient estimates from the previous section. The simulation approach integrates all the elements of the multi-state duration model to construct quarter-by-quarter pseudo-histories from age 15 through age 23 (36 quarters in all) for all youth in the NLSY97 sample. Using these pseudo-histories, we then provide a detailed set of summary statistics examining every aspect of youths’ disconnection experiences – the rates of initial disconnection, the usual length of disconnection spells, and the incidence of multiple spells.

The simulation framework is straightforward. Consider an individual from the survey with a particular set of characteristics (sex, race, ASVAB score, parents’ education and government aid use).¹⁵ We use the duration model for first disconnection spell to calculate the implied hazard rates and failure function for that set of characteristics. Then, by comparing a random draw from the uniform distribution with the estimated failure function, we determine whether that person ever enters a disconnection spell, and if they do, the age at which they enter.¹⁶ Next, we use the duration model for reconnection to calculate the probability of reconnecting by each quarter, draw another random number for the individual, compare the draw with the estimated survivor function for reconnection for the given profile of characteristics, and then determine whether and when the individual reconnects (if, of course, the individual disconnects in the first step). Finally, we do the same for re-disconnection. This provides a simulated history of disconnection, up through the beginning of a second disconnection spell, for every individual in the sample up to age 23. To overcome the error induced by this mechanism, we repeat this process for the entire sample ten times – with ten different vectors of random draws – and stack all ten sets of pseudo-histories to form a final simulated data set for use in analysis. We present our estimates from the simulation analysis using the “S” definition in Table 6. Note that comparisons across rows in this table are, by construction, comparisons that examine the net effect of all covariate differences across rows; they are not ceteris paribus comparisons toggling just the characteristic in the given row.

Table 6.

Summary of Disconnection Experiences for Youth of Various Characteristics

	Incidence of At Least One Disconnection Spell			Characteristics of First Spell of Disconnection for Youth with At Least One Spell								Incidence of Two Spells of Disconnection		Characteristics of Re-Connected Period for Youth with Two Spells
					Length of First Spell (Quarters)									Length of Re-Connected Period (Quarters)
Group	% All	% At At Least	% All Youth w/Least	Mean Age At of First		Percentiles			Percent with First Spell Longer Than			% All w/ Two	% w/ Two	Percentiles				Percent with Re-Connection Longer Than
Characteristics	Youth	1 Spell	1 Spell	Spell	Mean	20	50	80	1 yr	2 yrs	3 yrs	Spells	Spells	Mean	20	50	80	1 yr	2 yrs
All Youth	100.0%	17.6%	100.0%	18.6	6.5	3	6	9	73.8%	29.8%	10.3%	100.0%	5.0%	6.1	2	5	10	61.6%	31.3%
Male	50.8%	16.7%	48.0%	18.7	6.4	3	6	9	73.3%	28.9%	10.0%	45.5%	4.5%	6.1	2	5	10	63.4%	31.0%
Female	49.2%	18.6%	52.0%	18.5	6.6	3	6	9	74.3%	30.6%	10.6%	54.5%	5.6%	6.2	2	5	10	60.0%	31.5%
Race:
White and Other	72.5%	14.0%	57.5%	18.6	6.3	3	6	9	72.7%	28.4%	9.3%	53.4%	3.7%	6.2	2	5	10	62.4%	31.6%
Male	36.5%	13.0%	26.8%	18.7	6.3	3	5	8	72.8%	28.2%	9.3%	22.5%	3.1%	6.2	2	5	9	65.3%	31.5%
Female	36.0%	15.0%	30.7%	18.5	6.4	3	6	9	72.5%	28.6%	9.2%	31.0%	4.3%	6.1	2	5	10	60.3%	31.6%
Black	15.1%	32.4%	27.7%	18.5	7.2	3	6	10	77.9%	36.0%	14.3%	33.5%	11.1%	5.9	2	4	9	59.5%	29.8%
Male	7.7%	31.4%	13.7%	18.6	7.0	3	6	10	76.4%	34.5%	13.5%	16.2%	10.6%	5.9	2	5	10	61.0%	30.3%
Female	7.4%	33.4%	14.0%	18.4	7.3	3	6	10	79.4%	37.5%	15.0%	17.3%	11.7%	5.9	2	4	9	58.0%	29.3%
Hispanic	12.4%	21.1%	14.8%	18.6	5.9	3	5	8	70.8%	23.5%	7.0%	13.1%	5.3%	6.6	2	5	11	63.4%	34.0%
Male	6.6%	20.3%	7.6%	18.7	5.6	2	5	8	69.4%	21.4%	6.1%	6.9%	5.3%	6.0	2	5	10	62.5%	31.0%
Female	5.8%	22.0%	7.2%	18.5	6.1	3	5	8	72.1%	25.6%	8.0%	6.2%	5.4%	7.2	2	5	11	64.4%	37.3%
Parent’s Education:
HS Dropout	11.2%	35.5%	22.6%	18.5	6.5	3	6	9	74.2%	29.7%	10.2%	26.6%	11.9%	5.8	2	5	9	60.8%	28.3%
Male	5.7%	32.8%	10.6%	18.7	6.4	3	6	9	73.8%	29.1%	10.4%	11.7%	10.3%	5.6	2	5	9	62.7%	27.9%
Female	5.5%	38.4%	12.0%	18.3	6.5	3	6	9	74.6%	30.2%	10.1%	14.8%	13.5%	5.9	2	4	10	59.2%	28.7%
HS Graduate	31.0%	22.3%	39.1%	18.6	6.7	3	6	9	75.1%	31.7%	11.2%	42.2%	6.8%	5.9	2	4	9	58.6%	29.8%
Male	15.8%	20.4%	18.3%	18.7	6.4	3	6	9	74.0%	30.2%	9.9%	18.9%	6.0%	5.8	2	4	9	60.1%	29.2%
Female	15.2%	24.2%	20.8%	18.5	6.9	3	6	9	76.1%	33.0%	12.4%	23.3%	7.7%	6.0	2	4	9	57.4%	30.2%
Some College	26.0%	14.0%	20.7%	18.6	6.4	3	5	9	71.8%	28.8%	9.9%	16.3%	3.2%	6.6	2	5	11	63.2%	35.4%
Male	12.7%	12.6%	9.1%	18.6	6.5	3	5	9	71.9%	28.9%	11.0%	7.1%	2.8%	6.6	2	5	11	62.9%	33.4%
Female	13.3%	15.4%	11.6%	18.6	6.3	3	6	9	71.6%	28.7%	9.0%	9.2%	3.5%	6.5	2	5	10	63.4%	37.0%
College	16.3%	10.7%	9.9%	18.6	6.2	3	6	8	72.7%	26.6%	7.8%	7.7%	2.4%	7.3	3	6	11	71.5%	39.5%
Graduate
Male	8.1%	11.7%	5.4%	18.7	6.1	3	5	9	72.9%	26.4%	8.3%	3.5%	2.2%	7.0	3	6	10	75.2%	41.0%
Female	8.2%	9.7%	4.5%	18.4	6.2	3	6	8	72.5%	26.7%	7.1%	4.1%	2.5%	7.5	2	6	12	68.3%	38.3%
Advanced	15.5%	8.9%	7.8%	18.7	6.4	3	5	9	73.2%	27.7%	10.4%	7.3%	2.4%	6.5	2	5	10	67.5%	33.2%
Male	8.4%	9.9%	4.7%	18.8	6.1	3	5	8	72.6%	26.8%	9.6%	4.2%	2.5%	6.8	2	5	10	70.8%	35.6%
Female	7.0%	7.6%	3.0%	18.5	6.7	3	6	9	74.0%	29.1%	11.5%	3.1%	2.2%	6.1	2	5	10	63.0%	29.9%
Aid Status:
Receive Gov’t	41.1%	25.5%	59.4%	18.5	6.5	3	6	9	73.9%	30.1%	10.3%	65.5%	8.0%	6.0	2	5	10	60.6%	30.5%
Aid
Male	19.9%	23.6%	26.6%	18.7	6.4	3	6	9	72.8%	29.5%	10.4%	28.1%	7.1%	5.9	2	4	9	61.5%	29.5%
Female	21.2%	27.3%	32.8%	18.4	6.5	3	6	9	74.7%	30.5%	10.2%	37.4%	8.9%	6.2	2	5	10	59.9%	31.2%
ASVAB
Percentile:
0–25th	21.6%	42.5%	51.9%	18.5	6.5	3	6	9	75.1%	31.3%	10.9%	59.1%	13.8%	6.0	2	5	9	60.0%	29.7%
Male	11.4%	37.5%	24.2%	18.7	6.4	3	6	9	74.8%	29.8%	10.4%	26.0%	11.5%	5.9	2	4	9	61.3%	29.6%
Female	10.2%	48.1%	27.8%	18.4	6.7	3	6	9	75.2%	32.7%	11.4%	33.1%	16.4%	6.0	2	5	10	58.9%	29.8%
25–50th	21.6%	21.3%	26.1%	18.7	6.1	3	5	8	72.6%	27.2%	8.7%	23.4%	5.5%	6.1	2	5	10	62.7%	32.3%
Male	10.4%	20.1%	11.8%	18.8	6.0	3	5	8	71.4%	26.4%	9.0%	10.1%	4.9%	6.0	2	5	9	66.8%	30.3%
Female	11.2%	22.5%	14.3%	18.6	6.2	3	6	9	73.6%	27.9%	8.4%	13.3%	5.9%	6.2	2	5	10	59.6%	33.9%
50–75th	21.6%	11.0%	13.5%	18.7	6.2	3	6	9	73.6%	28.4%	8.9%	11.3%	2.6%	6.5	2	5	11	61.9%	35.9%
Male	10.5%	11.4%	6.8%	18.7	6.1	3	6	9	72.9%	29.8%	7.8%	5.7%	2.7%	6.4	2	5	11	63.9%	36.0%
Female	11.1%	10.7%	6.7%	18.6	6.4	3	5	9	74.2%	26.9%	9.9%	5.6%	2.5%	6.6	2	5	10	59.9%	35.9%
75–100th	21.7%	6.3%	7.8%	18.6	5.7	2	5	8	68.1%	24.8%	6.7%	6.1%	1.4%	7.0	2	5	11	71.8%	34.0%
Male	11.6%	7.5%	4.9%	18.6	5.7	2	5	8	69.1%	24.8%	7.0%	3.7%	1.6%	6.8	2	6	11	67.4%	35.3%
Female	10.2%	5.0%	2.9%	18.4	5.7	2	5	8	66.5%	24.9%	6.2%	2.5%	1.2%	7.2	3	5	12	78.3%	32.1%

The remainder of this section walks through Table 6, dividing the discussion into three parts: (i) the size and composition of the ever disconnected population; (ii) the length of initial disconnection spells; and (iii) multiple disconnection spells. We do not present formal statistical tests of differences, so small differences should be treated warily. Such tests could in principle be added based on bootstrapping methods. We do, however, focus on patterns of experiences where the underlying coefficient estimates exhibit statistically significant differences across sub-groups (see Section 5.1).

5.2.1. Size and composition of the disconnected population

The first three columns of Table 6 help contextualize the size and composition of the ever disconnected population, by sub-group.¹⁷ The first column shows what percentage of the youth population was made up of by each of the indicated sub-groups, as estimated from the NLSY97 sample and associated weights. The second column provides the percentage of the given sub-group that experienced at least one spell of disconnection by the last quarter of age 23. The third column gives the percentage of all youth who experienced a disconnection spell who fall into each of the given groups. The proportions in this column can be compared with the first column in order to understand which groups are disproportionately represented in the disconnected population.

The results show sizeable differences across sub-groups. The likelihood of experiencing any disconnection spell by age 23 ranges from just 6% in the highest ASVAB quartile to 43% in the lowest; the lowest ASVAB quartile makes up more than half of youth with any disconnection spells. Children of parents with a high school diploma or higher education have a 13–25 percentage point lower likelihood of experiencing any disconnection. Children of parents who receive government aid have an 8 percentage point higher likelihood of any disconnection than the average youth, and make up 59% of the ever disconnected population. Finally, Black males have an 18 and 11 percentage point higher likelihood of ever having a disconnection spell than White males and Hispanic males, respectively.

5.2.2. Length of initial disconnection spells

The next eight columns of Table 6 provide a variety of information about youths’ first spells of disconnection. We focus on differences across White, Black, and Hispanic women, given the pattern of statistically significant differences described above. Black women have a mean length of spell of 7.3 quarters, almost a quarter larger than White women (6.4) and 1.2 quarters larger than Hispanic women (6.1). Moreover, 37.5% of Black women are estimated to have a spell longer than 2 years prior to age 23, compared to 28.6% for White women and 25.6% for Hispanic women. About 15% of Black women have spells longer than 3 years prior to age 23.

5.2.3. Multiple disconnection spells

The remainder of Table 6 presents a wide array of information on experience with multiple spells. Only 5% of all youth are estimated to have a second spell. For this small portion of youth, 62% were reconnected for more than a year and 31% were reconnected more than two years prior to age 23 (this does not count the portion of reconnection spells that may extend past age 23). Again focusing on Black, White, and Hispanic females, Black females make up 7.4% of the population but 17.3% of those with multiple spells; in contrast, these numbers are closer for Hispanic females (5.8% of the population and 6.2% of those with multiple spells) and flipped in order for White females (72.5% of the population and 53.4% of those with multiple spells). Black females’ chances of having a second spell are 6.3 percentage points higher than Hispanic females and 7.4 percentage points higher than White females.

6. Comparisons of disconnection experiences across NLSY79 and NLSY97

This section addresses how disconnection has changed over time by comparing more recent experiences in the NLSY97 cohort with those seen twenty years before in the NLSY79 cohort. To begin with, Table 7 compares the two samples after making the sample exclusions summarized in Tables 1 and 2. The NLSY97 cohort has more Hispanic and fewer White youth than the NLSY79 cohort. The NLSY97 cohort also has a more educated set of parents than the NLSY79 cohort, with more parents that have at least some college education. The macroeconomic environment also changed substantially over time. Fig. 7 plots the unemployment rate and real GDP per capita over the 1970–2008 period. The NLSY97 was aged 12–17 in 1997. They saw a period of unemployment rate declines during the tail end of the economic expansion, before experiencing an economic slowdown in the early 2000s. The NLSY79 cohort, aged 13–15 in 1979, saw substantially higher unemployment rates throughout their adolescence, but with a very different pattern – they experienced an economic slowdown with sharply increasing unemployment at younger ages, followed by a robust economic expansion and sharply declining unemployment rates. In what follows, we do not attempt to decompose the portions of differences across the NLSY97 and NLSY79 cohorts due to youths’ characteristics rather than macroeconomic variables, but these comparisons provide important context for what we find below.

Table 7.

Selected Characteristics of Final NLSY97 and NLSY79 Samples

	Sample Means for Each Cohort
Characteristic	NLSY97	NLSY79
Black	0.15	0.15
White	0.73	0.79
Hispanic	0.12	0.07
Male	0.51	0.51
Received Govt Aid	0.41	N/A
Non-High School Grad Parent	0.11	0.21
High School Parent	0.31	0.44
Some College Parent	0.26	0.14
College Grad Parent	0.16	0.12
Advanced Degree Parent	0.16	0.09
Non Urban	0.32	0.23
ASVAB1980	53.68	N/A
ASVAB1981	53.95	N/A
ASVAB1982	54.87	N/A
ASVAB1983	52.99	N/A
ASVAB1984	53.91	N/A
AFQT63	N/A	46.22
AFQT64	N/A	41.44

NOTE: Each cell represents the mean from the sample in the given cohort.

Fig. 7.

U.S. unemployment rate and real GDP Per Capita, 1970–2008

NOTE: Unemployment rate is sourced from U.S. Bureau of Labor Statistics, retrieved from FRED, Federal Reserve Bank of St. Louis; https://fred.stlouisfed.org/series/UNRATE, March 14, 2022. Real GDP Per Capita is sourced from U.S. Bureau of Economic Analysis, retrieved from FRED, Federal Reserve Bank of St. Louis; https://fred.stlouisfed.org/series/A939RX0Q048SBEA, March 14, 2022.

For the purposes of the comparisons below, we estimate hazard models and construct figures using closely analogous approaches in the NLSY79 cohort as we did for NLSY97. The primary methodological difference is that the NLSY79 hazard model specifications do not include covariates for parental government aid receipt or (for reconnection and re-disconnection) criminal convictions during the initial disconnection spell, due to data availability. In preliminary analysis, we have estimated failure functions in the NLSY97 cohort using a more comparable empirical specification without these covariates, and differences from failure functions for disconnection between the simpler specification and the more complete specification used in Sections 4–5 were less than 0.01 percentage points across all ages. Finally, we do not show any comparisons using the “C” definition because we do not have sufficient information to construct credible cohabitation histories in NLSY79.

Figs. 8–11 present a series of comparisons across the NLSY97 and NLSY79 cohorts for males, and Figs. 12–15 present the analogous comparisons for females. These figures cover the following: (i) fitted hazard rates (Figs. 8, 12); (ii) failure rates for initial disconnection (Figs. 9, 13); (iii) success rates for reconnection (Figs. 10, 14); and (iv) failure rates for re-disconnection (Figs. 11, 15). All figures present results for the “N” and “S” definitions using both the NLSY97 and NLSY79 cohorts. All figures for NLSY79 use the coefficient estimates from the hazard models for initial disconnection, reconnection, and re-disconnection presented in Table 8; those for NLSY97 use the estimates from Table 5. We conduct identical supplementary hypothesis testing for the models in Table 8 to that conducted for the NLSY97 models in Table 5 (see pages 24–25). As with the NLSY97 cohort, we include alternative specifications for the NLSY79 hazard rates in Appendix Tables 7–10.

Fig. 8.

Comparison of fitted hazard rates for disconnection of young men in NLSY79 and NLSY97 Cohorts

NOTE: Fitted hazard rates are based on the estimates in Table 5 (NLSY97) and Table 8 (NLSY79).

Fig. 11.

Probability of young men experiencing a second disconnection episode by quarter after reconnection (NLSY79 and NLSY97 Cohorts)

NOTE: Curves depict the fitted failure rate for re-disconnection — i.e., the probability of re-disconnecting by quarter after re-connection, for those who disconnected and then re-connected. Fitted probabilities calculated using estimates in Table 5 (NLSY97) and Table 8 (NLSY79).

Fig. 12.

Comparison of fitted hazard rates for disconnection of young women in NLSY79 and NLSY97 Cohorts

NOTE: Fitted hazard rates are based on the estimates in Table 5 (NLSY97) and Table 8 (NLSY79).

Fig. 15.

Probability of young women experiencing a second disconnection episode by quarter after reconnection (NLSY79 and NLSY97 Cohorts)

NOTE: Curves depict the fitted failure rate for re-disconnection — i.e., the probability of re-disconnecting by quarter after re-connection, for those who disconnected and then re-connected. Fitted probabilities calculated using estimates in Table 5 (NLSY97) and Table 8 (NLSY79).

Fig. 9.

Probability of experiencing a disconnection episode by particular age for young men in NLSY79 and NLSY97 Cohorts

NOTE: Curves depict the fitted failure rate by each age — i.e., the probability that a youth has experienced a disconnection episode by that age, using coefficient estimates from Table 5 (NLSY97) and Table 8 (NSLY79).

Fig. 13.

Probability of Experiencing a Disconnection Episode by Particular Age for Young Women in NLSY79 and NLSY97 Cohorts

NOTE: Curves depict the fitted failure rate by each age — i.e., the probability that a youth has experienced a disconnection episode by that age, using coefficient estimates from Table 5 (NLSY97) and Table 8 (NSLY79).

Fig. 10.

Probability of young men experiencing a reconnection episode by quarter after initial disconnection (NLSY79 and NLSY97 Cohorts)

NOTE: Curves depict the fitted success rate for reconnection — i.e., the probability of re-connecting by quarter after initial disconnection, for those who disconnected. Fitted probabilities calculated using estimates in Table 5 (NLSY97) and Table 8 (NLSY79).

Fig. 14.

Probability of young women experiencing a reconnection episode by quarter after initial disconnection (NLSY79 and NLSY97 Cohorts)

NOTE: Curves depict the fitted success rate for reconnection — i.e., the probability of re-connecting by quarter after initial disconnection, for those who disconnected. Fitted probabilities calculated using estimates in Table 5 (NLSY97) and Table 8 (NLSY79).

Table 8.

Coefficient Estimates in Hazard Models for Disconnection, Re-Connection, and Re-Disconnection (NLSY79 Cohort)

	Disconnection				Re-Connection		Re-Disconnection
	Males		Females		N (5)	S (6)	N (7)	S (8)
Covariates	N (1)	S (2)	N (3)	S (4)
Male					0.4944*** (0.1465)	−0.1346 (0.1715)	−0.7750*** (0.2123)	−0.0874 (0.2534)
Black	0.1604 (0.1662)	0.2320 (0.1685)	0.2120 (0.1457)	0.7959*** (0.1752)	−0.3053* (0.1436)	−0.7505*** (0.1659)	−0.1314 (0.1869)	0.2454 (0.2711)
Hispanic	−0.0771 (0.1970)	−0.0793 (0.2046)	−0.1678 (0.1686)	0.0511 (0.2057)	−0.0637 (0.1607)	−0.4661* (0.1913)	−0.1950 (0.2061)	0.2111 (0.2810)
Black × Male					0.2394 (0.1970)	0.7707*** (0.2174)	0.6502* (0.2685)	0.3602 (0.3372)
Hispanic × Male					0.0275 (0.2397)	0.4775+ (0.2728)	0.3124 (0.3289)	−0.1181 (0.3942)
1st Disconnected, Age 17					0.1119 (0.2101)	−0.0197 (0.2333)	0.2005 (0.2327)	−0.2673 (0.2644)
1st Disconnected, Age 18					0.3736+ (0.2021)	0.0302 (0.2232)	0.1764 (0.2209)	−0.0233 (0.2509)
1st Disconnected, Age 19+					0.3325 (0.2030)	0.1276 (0.2166)	−0.1342 (0.2135)	−0.3024 (0.2498)
HS Graduate, Parents	−0.3300 (0.1822)	−0.3266 (0.1873)	−0.1603 (0.1420)	−0.2241 (0.1704)	0.1071 (0.1189)	0.0422 (0.1281)	−0.2451 (0.1537)	−0.0211 (0.1755)
Some College, Parents	−0.0707 (0.2546)	−0.1068 (0.2637)	−0.3671 (0.2313)	−0.1541 (0.2738)	0.0594 (0.1815)	0.0613 (0.1962)	−0.5926* (0.2757)	−0.6214+ (0.3350)
College Graduate, Parents	−0.3497 (0.3334)	−0.3020 (0.3369)	−0.8768* (0.3449)	−0.5002 (0.4068)	0.4951+ (0.2671)	0.7318* (0.3080)	−0.2061 (0.3483)	−0.3715 (0.4810)
Advanced Degree, Parents	−0.5600 (0.4933)	−0.5241 (0.4982)	−0.7454+ (0.3815)	−0.4960 (0.4542)	0.1324 (0.3204)	0.0405 (0.3527)	0.1749 (0.3903)	0.0711 (0.4670)
Revised AFQT × Cohort 62					−0.0007 (0.0053)	0.0030 (0.0069)	−0.0006 (0.0075)	−0.0116 (0.0095)
Revised AFQT × Cohort 63	−0.0336*** (0.0056)	−0.0322*** (0.0058)	−0.0269*** (0.0039)	−0.0288*** (0.0052)	0.0051 (0.0034)	0.0064 (0.0039)	−0.0050 (0.0042)	−0.0035 (0.0050)
Revised AFQT × Cohort 64	−0.0247*** (0.0053)	−0.0231*** (0.0054)	−0.0309*** (0.0046)	−0.0293*** (0.0060)	0.0050 (0.0034)	0.0057 (0.0038)	−0.0131** (0.0050)	−0.0073 (0.0058)
Gave Birth Prior to 1st Disc.					−0.1807 (0.1382)	−0.1833 (0.1611)	−0.2542 (0.1936)	0.0368 (0.2224)
Drop Out Prior to 1st Disc.					0.0943 (0.1161)	0.0617 (0.1239)	0.4501** (0.1579)	0.3793* (0.1866)
Gave Birth During 1st Disc.							0.1884 (0.1573)	0.3070 (0.2308)
Observations	26,114	26,219	22,490	23,880	7,315	5,354	9,557	9,858

Columns 1, 3, 5, and 7 provide the estimates in hazard models for male disconnection, female disconnection, re-connection, and re-disconnection using the “N” definition. Together, they constitute the multi-state duration model for the “N” definition. Analogously, columns 2, 4, 6, and 8 constitute the multi-state duration model for the “S” definition. Robust standard errors in parentheses.

^***p < 0.001,

^**p < 0.01,

^*p < 0.05,

⁺p < 0.1.

The remaining sub-sections use the figures, tables, and supplementary hypothesis tests to compare the NLSY97 and NLSY79 cohorts on overall disconnection experiences (Section 6.1) and variability across sub-groups (Section 6.2).

6.1. Differences in disconnection experiences across time for young women and men

Males in the NLSY97 cohort have a similar or better set of circumstances than males in the NLSY79 cohort in terms of initial disconnection, but potentially worse circumstances for reconnection and re-disconnection. Fig. 9 shows that in the NLSY97, approximately 13% of males experience a disconnection spell by age 20, while the NLSY79 cohort has a higher rate of 17–18%. Hypothesis tests comparing the predicted hazard rates for a baseline profile of male youth (White, non-high school graduate parents, mean ASVAB score, and – for NLSY97 – mean share of government aid receipt) show that one cannot reject equality across the NLSY97 and NLSY79 cohorts. Figs. 10 and 11 show that among those who initially disconnect, the experiences of those in the NLSY97 cohort are, if anything, somewhat worse than those in the NLSY79 cohort – there are lower reconnection rates and higher re-disconnection rates across spell duration.

The patterns for females point to clearer improvements in circumstances over time. The probability of ever experiencing a disconnection spell is substantially higher in the NLSY79 cohort than in the NLSY97 cohort for the “N” definition, with the gap rising to 12 percentage points by age 23. Hypothesis testing comparing underlying hazard rates across the two cohorts using the same baseline profile above shows the NLSY79 rates are statistically significantly higher (at the 10% level) than the NLSY97 rates in the age 18 time period. We have not yet tested the resulting impact on the difference in failure functions across ages (though looking at the difference in failure functions at a single age of 20.5 yields a z-statistic of 1.48, implying a P-value of approximately 0.14). Using the “S” definition, one still sees improvements over time, but at a smaller magnitude; moreover, the underlying hazard rates are not statistically significantly different from one another. (Given that cohabitation is likely more common in the NLSY97 cohort than the NLSY79 cohort, though, some of the cohort gap reduction through marriage is likely artificial; it is possible, though speculative, that the cohort gap under the “C” definition would therefore more closely resemble that under the “N” definition.) In any case, these differences between the “N” and “S” definitions suggest that the enrollment and employment circumstances of young women have improved appreciably across cohorts. Moreover, the smaller gap between failure rates for the “N” and “S” definitions in the NLSY97 cohort suggest that fewer young women are relying on living with a spouse to avoid disconnection spells. Figs. 14 and 15 suggest that the NLSY97 cohort of women enjoys similar reconnection and re-disconnection rates to the NLSY79 cohort (though the curves for the “S” definition in Fig. 15 suggest some deterioration over time). In fact, the relative position of women to men has improved over time when it comes to reconnection and re-disconnection. White females exhibit statistically significantly lower reconnection rates than White men in the NLSY79 cohort under the “N” definition; they are statistically indistinguishable, however, in the NLSY97 cohort.

Given the gap between NSLY97 and NLSY79 cohorts in initial disconnection hazard rates during usual school leaving ages for men and women (Figs. 8 and 12), we believe that changes in patterns of educational attainment play a large role in any improvements we are observing over time. Appendix Figs. 3 and 4 investigate this hypothesis further. Appendix Fig. 3 reports the failure functions for first drop-out, using hazard models analogous to those for disconnection, reconnection, and re-disconnection.¹⁸ Overall, the figures suggest that male and female dropout rates in later ages were about 7 and 3 percentage points lower, respectively, for the NLSY97 cohort than the NLSY79 cohort. Dropping out does not necessarily correspond to disconnection because some youths immediately find employment or re-enroll in school. Appendix Fig. 4 reports the probabilities that a youth experienced entry into the labor market or a return to school by a given quarter since dropping out.¹⁹ Two quarters after dropping out, the NLSY97 cohorts were about 6–7 percentage points more likely to have returned to school relative to the NLSY79 cohorts. While these differences should be interpreted cautiously since we are not conducting formal significance tests, the results show that the NLSY97 cohort were less likely to leave school than the NLSY79 cohort and, conditional on leaving school, were slightly more likely to return to school quickly. This, and the fact that labor market participation conditional on dropping out was nearly identical across surveys, suggests that changes in enrollment patterns over the last few decades are partially responsible for the reduction in disconnection between the 1979 and 1997 surveys.

6.2. Differences in sub-groups more inclined to disconnect and reconnect across cohorts

Finally, we examine differences between the NLSY97 and NLSY79 cohorts in the sub-groups more inclined to disconnect and reconnect. We briefly summarize the results for ASVAB/AFQT, parental background, and race in turn, focusing most of the discussion on striking differences over time by race.

The qualitative effects of ASVAB/AFQT scores on initial disconnection, reconnection, and re-disconnection are very similar across the NLSY97 and NLSY79 cohorts. Higher scores strongly predict reduced rates of initial disconnection in a way that is similar across cohorts, but there is little evidence of a correlation with reconnection and re-disconnection. Supplementary hypothesis testing shows that the only qualitative differences are that in NLSY79, the ASVAB coefficients are jointly different from zero for the re-disconnection “N” definition (10% level) and no longer jointly different from zero for the reconnection “S” definition. The coefficient size, and implied magnitude of effect on hazard rates, continues to be close to zero for reconnection. Given these results, the role of ASVAB/AFQT in disconnection experiences appears quite stable over time.

Parental education plays a less clear role in NLSY79 than in NLSY97. The exception is for female youth in the “N” definition. In this case, two of the parental education coefficients are statistically significantly different from zero (at 10% level or better) and one can reject the null that all parent education coefficients are jointly zero, just as in the NLSY97 cohort. For all other cases of the initial disconnection hazards, one fails to reject that the parental education coefficients are jointly zero. For reconnection and re-disconnection, one also fails to reject that the parental education coefficients are jointly zero, despite the fact that some covariates are individually statistically significant at the 10% level or better.

The most important differences between the NLSY79 and NLSY97 results concern race, with the role of race changing differentially by sex. The circumstances of Black males have deteriorated over time, both in absolute terms and relative to White and Hispanic males. To compare absolute circumstances, we compare the predicted hazard rates for initial disconnection of Black males in NLSY79 and NLSY97 using a baseline profile of covariates (Black, non-high school graduate parents, mean ASVAB score, and – for NLSY97 – mean share of government aid receipt). We observe a statistically significant increase in the NLSY97 cohort (at the 1% level), relative to the NLSY79 cohort, for both the “N” and “S” definitions around age 20 (and no differences at earlier ages). We have not yet tested the resulting impact on the difference in failure functions across ages (though looking at the failure function at a single age of 20.5 does not yield a statistically significant difference). The situation of Black males has also deteriorated in relative terms. For initial disconnection, in NLSY97 Black males are more likely to disconnect than White males and Hispanic males (1% level) for both the “N” and “S” definitions. In NLSY79, Black males were statistically indistinguishable for all but the case of Hispanic males and the “S” definition. The relative position of Black males to White males is qualitatively more similar over time for reconnection and re-disconnection (with perhaps some improvement for re-disconnection under the “N” definition).

In sharp contrast, the circumstances of Black females have markedly improved over time, both in absolute terms and relative to White and Hispanic females. To examine absolute changes, we compared predicted hazards for initial disconnection using the analogous exercise to that of Black males. We find that for both the “N” and “S” definitions, the NLSY97 predicted hazard rate is statistically significantly lower than the NLSY79 predicted hazard rate around ages 18–19. We have not tested the difference in failure functions across multiple ages, but at age 20.5 the difference is statistically significant at 1%. The situation of Black females has also improved in relative terms. The magnitude of hazard coefficients and implied impact on the gap between Black females and White females has fallen sharply. For the baseline profile above, for instance, in NLSY79 Black females have a probability of ever disconnecting under the “S” definition by age 20.5 of 32%, as compared to 16% for analogous White females. In NLSY97, this gap closes to 19% versus 17%. There is also evidence of improvement relative to White females for reconnection, where the NLSY79 cohort exhibits statistically significantly lower reconnection rates for Black females than Whites for both the “N” and “S” definition, while NLSY97 exhibits this only for the “S” definition; this result is not sensitive to the inclusion of the government aid covariate in NLSY97.

7. Discussion and conclusion

This paper provides a complete characterization of youths’ transitions to and from disconnection statuses, one that permits an understanding of the timing and duration of disconnection and reconnection events for the overall population and key sub-groups, as well as how these have changed from the 1980s to 2000s. The key findings of the paper include:

• A large fraction of youth experience at least one disconnection spell. In the early 2000s, by age 22 approximately 18% of males and 23% of females have experienced a disconnection spell according to the “N” definition. This number falls by roughly 0.5 and 4 percentage points for men and women, respectively, when using the “S” definition. Through the use of complete longitudinal histories on youth through age 23, we avoid right censoring issues that have led to downward biases in previous studies’ estimates. Our definition of disconnection spell also improves upon previous studies by focusing on intense bouts of disconnection, while not counting cases where youth are accumulating many scattered disconnected months over a long period.

• Entry into an initial disconnection spell is most common at ages 18–19. For both men and women, and regardless of “N” or “S” definition, hazard rates for initial disconnection start off very low, peak between ages 18 and 19, and then decline before stabilizing at a medium level above that of the early ages. This strongly suggests that large segments of youth are facing their initial disconnection spell as they finish high school and attempt a transition to the workforce or college. This pattern remains true across time, appearing in both the NLSY97 and NLSY79 cohorts.

• Parental education and government aid receipt, youth test scores, and race are important predictors of the likelihood of initial disconnection. In the early 2000s, having parents without a high school degree, parents who receive government aid, and lower ASVAB scores is associated with dramatically higher incidence of disconnection spells by age 23, with effect sizes ranging from 8 to 25 percentage points more than the overall population after accounting for all differences in characteristics. Black males have strikingly higher incidence than the overall population, with rates of experiencing initial disconnection that are 9 percentage points higher by age 21.

• Once they occur, a surprising portion of initial disconnection spells last more than one year. In the early 2000s, more than 70% of initial disconnection spells last more than one year, 30% last more than two years, and 10% last more than four years. These estimates fully account for any right censoring arising from youth being in the middle of a disconnection spell in their last survey round. Existing studies of disconnection spell length are generally contaminated by right and left censoring issues that tend to understate the length of disconnection spells.

• Entry into a reconnection spell is most common 8 quarters after initial disconnection. For both men and women, and regardless of “N” or “S” definition, hazard rates for reconnection exhibit a roughly symmetric distribution with a peak two years after the start of an initial disconnection spell. This pattern remains true across time, appearing in both the NLSY97 and NLSY79 cohorts.

• The majority of reconnecting youth are able to sustain reconnection for a long period, but an important minority experiences a second spell quickly. In the early 2000s, regardless of sex and definition, our duration model implies that more than 60% of youth have their reconnection spell last more than two years, and more than 50% of youth have it last at least three years. Still, approximately 24–32% of youth experience a second episode of disconnection within the first year and a half after reconnecting, depending on the disconnection definition and sex. These estimates, the first such estimates we are aware of, fully account for any right censoring arising from youth being in the middle of a reconnection spell in their last included survey sample year.

• Those experiencing their initial disconnection at age 19 or later are least likely to reconnect quickly, but few other characteristics predict reconnection or re-disconnection. In the early 2000s, those who initially disconnect later in life are statistically significantly less likely to have a reconnection spell in any given quarter after disconnection. Few other characteristics are statistically significant predictors of reconnection and re-disconnection. This is not simply due to imprecision of estimates, as many key coefficients (e.g. ASVAB) are close to zero with very small implied marginal effects and small standard errors.

• In comparison to the 1980s, the likelihood of ever experiencing an initial disconnection spell has fallen appreciably for young women, and has stayed stable for young men. For young women, the probability of ever experiencing a disconnection spell is substantially higher in the NLSY79 cohort than in the NLSY97 cohort for the “N” definition, with the gap rising to 12 percentage points by age 23. Young men also observe a small decline, but the underlying hazards in NLSY79 and NLSY97 appear statistically indistinguishable.

• In comparison to the 1980s, circumstances have improved substantially for Black women but have deteriorated for Black men. The probability of ever experiencing a disconnection spell has declined for Black women over time, both in absolute terms and relative to other women. The probability of ever experiencing a disconnection spell has, in contrast, increased for Black men over time, both in absolute terms and relative to other men.

For policy makers aiming to attack the problem of disconnected youth, this paper therefore shows the size of the problem, for whom it is most pressing, and at what time points key spell transitions happen. It also places more recent patterns into historical context, with changes over time offering important clues for what could produce sustainable improvements in the future. In doing so, it uses a novel disconnection spell definition, more comprehensive data, and a duration modeling framework to remedy measurement biases inherent to past studies’ examination of the problem, biases that have tended to lead to an under-estimation or mis-appreciation of the incidence and persistence of disconnection spells.

Going forward, our empirical framework provides a vehicle that one can adapt straightforwardly to evaluate the role of more detailed family circumstances, the economic environment, and policy interventions in preventing youths from falling into disconnected statuses and in assisting youths to escape from disconnected states when prevention fails. For instance, the NLSY offers an array of opportunities for integrating important additional covariates into the multi-state duration model. For researchers interested in family circumstances, this could include NLSY variables describing family routines and household risks. For researchers interested in the role of labor market opportunities, it would be possible to integrate NLSY geocode variables on county economic variables, such as employment by industry. For researchers interested in policy interventions, it would be possible to exploit differential changes in Medicaid or other programs across states. In all these cases, useful descriptive results could be obtained simply by integrating new covariates into the existing multi-state duration model. For those interested in accounting for endogeneity of certain variables – e.g. type of criminal conviction – more adaptation would of course be needed.

More generally, our modeling framework could be easily adapted to other longitudinal data settings where right censoring is a substantial problem and the sequences of outcomes of interest are complex and numerous. The modularity of the framework offers empirical researchers considerable flexibility to break complex trajectories with multiple states into manageable pieces. Researchers can then implement descriptive analysis quickly and easily with standard statistical software. This initial descriptive analysis can provide useful empirical guidance to inform the design of more sophisticated economic models and eventual estimation of structural parameters.