## Abstract

Disability status—experiencing a functional limitation caused by a health condition—is dynamic throughout the life cycle, even during adolescence and young adulthood. We use data from the 1997 cohort of the National Longitudinal Survey of Youth to better understand these dynamics, examining how health condition and limitation statuses evolve during adolescence and young adulthood as well as how changes in these characteristics are related to survey nonresponse and attrition. Health condition and limitation dynamics are evident in our data: the proportion of sample members who reported having a limitation in their activities for any interview increased from approximately 12 % during the initial interview (when sample members were 12 to 17 years old) to almost 25 % 13 years later. Multivariate analyses revealed that women are more likely than men to report changes in health condition or limitation status. Those with mild limitations were relatively less likely than those without limitations or with severe limitations to experience changes in limitation status. Somewhat surprisingly, a survival analysis of survey participation outcomes found limited correlation among health conditions, limitations, and either missing a survey interview for the first time or permanently leaving the survey sample.

## Introduction

Although health and disability factors are typically considered important aspects of older adulthood, they can also potentially affect youth and young adults’ ability to transition into adulthood. However, measures of health and disability in commonly used national sources, such as the Current Population Survey (CPS) and American Community Survey (ACS), capture individuals at a single point in time and typically rely on adult-based measures that may not completely capture the dynamic and evolving experiences of youth (Halfon and Newacheck 2010).

Because of the particularly dynamic nature of health during adolescence and early adulthood, it may be important to take a broader view than can be measured during a point-in-time survey contact and to understand how health status can affect survey participation. This information can guide policymakers, researchers, and other stakeholders in developing policies and interventions that improve outcomes for all youth, not just those identified with a health condition at a specific point in time. In this study, we consider disability dynamics using definitions that focus on health conditions and their resulting limitations. We capitalize on a longitudinal survey of youth ages 12 to 16 on December 31, 1996, to track health condition prevalence over a 15-year period to understand how condition and limitation prevalence changes as youth age into young adulthood. We seek to answer two questions. First, how does the prevalence of health condition and limitation severity status change as adolescents become young adults? Second, to what extent are changes in condition and limitation status predictive of survey nonresponse and attrition?

Our findings suggest that at any given observation, slightly less than 10 % of the sample members had a condition that limited them; however, over the entire observation period, almost one-half of youth had a potentially limiting condition, and almost one in four had a limitation in their activities because of a health condition at any point. Further, the presence of a health condition or limitation is a strong predictor of future changes in a condition or limitation, respectively, reflecting the dynamic patterns of health for this population. Finally, neither condition nor limitation status was typically strongly associated with survey nonresponse and attrition.

In this article, we provide background information on measuring health and disability in youth and young adults along with a consideration of how having a health condition might affect survey participation. We follow with an explanation of our data and methods before presenting our results. We conclude with a discussion of the implications of our results for both policy and survey methods.

## Background

### Disability Prevalence of Youth

The current standard of defining disability, based on the International Classification of Functioning, Disability and Health (ICF) model, allows for a systematic and comprehensive view (Weathers 2009; World Health Organization (WHO) 2001). Under the ICF model, disability is a function of one’s health, environment, and personal factors: a health condition may result in a disability through an impairment that affects one’s body structure or function, an activity limitation that affects one’s ability to execute a task or action, or a participation limitation that affects the individual’s ability to be involved in a societal or life situation (such as with work or friends). Further, and perhaps more importantly, various environmental and personal factors can exacerbate one or another disability definition or eliminate it altogether—an important consideration for youth, given their transition from school to adult environments. For example, an individual with a learning condition may have a disability as measured by impairment in a school context, but the condition may not affect her ability in other activities (such as work) because of appropriate environmental supports. This definition is similar to, but distinct from, previous medical or social disability schemes, such as Nagi (1965) (in which a condition leads to an impairment, which can lead to a limitation, which can lead to a disability in one’s performance in social roles) or Verbugge and Jette’s (1994) disablement process (which builds on Nagi (1965), but includes risk factors that influence impairments and intra-/extraindividual factors that influence limitations). We use the ICF framework because it provides a broader conceptualization of disability that is not linear—that is, an impairment may not necessarily lead to an activity limitation, but could lead to a participation limitation—and separates our roles into those for the self and those involving a societal interaction. However, although each of the ICF disability definitions is specific and precise, the term “disability” itself becomes quite broad and imprecise.

Disability prevalence—estimated in terms of impairments, activity limitations, and/or participation limitations—for youth and young adults varies based on the specific survey and its questions, often because of differences with the concepts measured as well as the use of measures designed for adults (Honeycutt and Wittenburg 2012). Point-in-time surveys such as the CPS and the ACS, which use a limited set of questions to measure disability, suggest that approximately 5 % to 6 % of individuals ages 15 to 24 have a disability. Other surveys that include a broader range of measures related to disability, such as the Survey of Income and Program Participation (SIPP), show higher disability prevalence rates (around 11 % to 13 %). Still other surveys, such as the National Survey of Children’s Health (NSCH), a cross-sectional survey of households that focuses on health care access and use among children ages 17 and younger, identify even higher numbers of youth with conditions because of the surveys’ broader set of health-related questions. Based on information in the NSCH, between 6 % and 19 % of youth ages 16 and 17 years old have a sensory, physical, or mental condition, and 31 % have an impairment, activity limitation, or participation limitation. However, these statistics are for individuals with varying degrees of impairments; the percentages decrease as the impairment definition is restricted to those having moderate or severe impairments.

Reported disability prevalence changes for youth as they age into adulthood, although detailed information on how prevalence changes for individuals over this period is limited. Broadly, we know that disability prevalence increases over the life course and is highest among older adults. For example, disability prevalence in the SIPP rises from 10 % for those 15 to 24 years old to 29 % for those 55 to 64 years old (Brault 2012). Most information in this area comes from cross-sectional studies, but a major challenge is that surveys often include relatively limited information about onset, and this information is usually retrospective. For example, using a cross-sectional survey of adults, Jamoom et al. (2008) examined data that asked individuals with a disability about the onset of their condition. Of the 12,000 individuals with a disability, 11 % had onset before age 21, and their condition severity tended to be lower than that of respondents whose onset occurred after that age. Other evidence has suggested that individuals with a disability onset before and after age 22 have different adult outcomes. For example, those with early onset had lower educational attainment but higher employment rates (depending on other factors, such as disability benefit receipt; Loprest and Maag 2007). However, these studies are based on retrospective data and thus may not capture some of the subtleties in how disability status changes over time. They may also be subject to recall bias.

### Using Longitudinal Data to Measure Disability

Longitudinal disability statistics for youth and young adults are not common, with the National Longitudinal Survey of Youth 1997 (NLSY97) being one of the best data sources for this information. Despite its wide use in observing the experiences of youth with disabilities (e.g., Shandra 2011; Shandra and Hogan 2008), its use in understanding the dynamics of disability for youth has been limited. Mann and Honeycutt (2014) pursued this line of research to provide evidence of the variability in disability status over time for youth and young adults. They found that one-half of youth ages 12 to 17 had ever had health conditions by the time they were ages 18 to 23, although far fewer (about 15 %) had health conditions that resulted in a limitation in their activities.

### Issues With Missing Data in Longitudinal Surveys

Although longitudinal data have multiple benefits, those benefits can be diminished because of challenges with missing data across survey rounds. In addition to item nonresponse (in which a survey respondent fails to complete a specific item, a problem in cross-sectional surveys as well), survey nonresponse can occur when individuals do not participate in one or more follow-up survey waves: some respondents do participate in some follow-up survey waves, and others do not participate in any more follow-up survey waves (attrition). Survey nonresponse can be problematic to statistical inference when there are systematic reasons or patterns among those who do and those who do not complete surveys. That is, if survey nonresponse is related to the variables of interest, results based only on survey completers may be biased (Allison 2001; Little and Rubin 1987; Rubin 1976). Statistically appropriate responses to survey nonresponse can include approaches such as multiple imputation and weighting to account for missing data.

Variables associated with survey nonresponse for youth and young adults are similar to patterns observed for adults more broadly. Demographic variables (being older, male, nonwhite) are typically identified as being correlated with survey nonresponse, as are indicators of lower employment and socioeconomic status, such as experiencing chronic poverty and having lower educational attainment (Ahern and Le Brocque 2005; Macurdy et al. 2001). Health factors may also play a role in attrition, although the findings are not as consistent as with other variables. Among older adults, those in poor health or with more reported health conditions are more likely to attrite or die (Zhivan et al. 2012); in the general population, factors associated with survey nonresponse include poor mental health status (Torvik et al. 2012). In contrast, comparisons of health-related variables in a longitudinal survey with those in cross-sectional surveys find overall agreement in the representativeness of the surveys, suggesting no critical issues with nonresponse related to health (Schoeni et al. 2013). For youth, the findings are similar. Some analyses have found that health variables, such as having anxiety or depression, are associated with being more likely to miss interviews (Davey et al. 2001), while other research has suggested no differences in response by health status (such as mental illness) after controlling for background factors, such as SES (de Graaf et al. 2013; Jeličić et al. 2010). The consensus seems to be that the types of variables related to survey nonresponse can vary from survey to survey and even across statistical models and within different groups for the same survey (Ahern and Le Brocque 2005; Macurdy et al. 2001); that is, health variables are not always found to be predictors of survey nonresponse and may depend on the population, the context, and the situation.

### Research Questions

Given this background, we set out to answer two research questions using data from the NLSY97 to understand disability dynamics for youth, using questions on health condition and limitations as proxies for disability status. First, how does the prevalence of health condition and limitation severity status change as adolescents become young adults? Our main hypothesis is that prevalence and severity increase over time. Related to this question, we also explore how those changes are associated with individual demographic and health characteristics. Second, to what extent are changes in condition and limitation status predictive of survey nonresponse and attrition? We hypothesize that neither condition nor limitation status affects survey participation, but changes in these statuses are associated with survey nonresponse. To further explore this question, we examine survey nonresponse related to specific conditions and limitations as well as to specific demographic and background characteristics. We also compare nonresponse for attrition and for those who missed an interview.

## Method

### Data

Our data source for this study is the NLSY97, a survey that is well positioned to provide insights into condition and limitation status transitions. The NLSY97 is a nationally representative, longitudinal survey that has tracked and is still tracking a cohort of Americans who were 12 to 16 years old on December 31, 1996. The sample tracks 8,984 people—6,748 from a nationally representative cross-section, and 2,236 from an oversample of blacks, Hispanics, and Latinos. We refer to the people who are being tracked by the NLSY97 as respondents. The NLSY97 contains a wealth of information relevant to our study, such as missed interview status, marital status, educational attainment, health, and household poverty status. Data for 15 NLSY97 annual survey rounds were available for our analysis. Except for our analysis of the change in limitation status (described later), we use the entire NLSY97 sample. Annual interviews are mostly completed in person, although some are completed by phone; very few (no more than 14 in a given survey round) are completed by proxy. In all analyses, respondents are weighted to be nationally representative as of December 31, 1996.1

The NLSY97 collected health condition information from respondents during several survey rounds. During the initial round, respondents’ parents were asked about their children’s health condition status. However, for the 6th and 11th, 12th, and 13th survey rounds, the health condition data were collected directly from survey respondents. This change in data collection may influence differences in condition status between the 1st and 6th rounds. Each respondent’s health condition information was recorded only once during the 11th, 12th, and 13th survey rounds—whichever survey round the respondent participated in first (and hereafter referred to as the 11th round, for brevity). For example, if a respondent missed his or her 11th survey round interview but participated in the next two interview rounds, the NLSY97 would record that person’s health condition information in the 12th survey round.

The NLSY97 asks about four types of health conditions along with limitation status. The questions for youth are as follows:

• Have you ever had trouble seeing, hearing or speaking?

• Have you ever had a part of your body that was deformed or missing?

• Have you ever been diagnosed with any other chronic health condition or life threatening disease such as [asthma, cardiovascular or heart condition, anemia, diabetes, cancer, epilepsy, HIV/AIDS, sexually transmitted disease other than HIV/AIDS, other]?

• Have you ever had an eating disorder, a learning or emotional problem or a mental condition that has limited your ability to attend school regularly, do regular school work, or work at a job for pay?

Similar questions were asked of parents about their child in Round 1 with slight wording changes, the most notable of which is not including “eating disorder” in the fourth question. We classify positive answers to the preceding four questions as indicating a condition, although the questions themselves are imprecise as to whether respondents report a condition, an impairment, or a limitation because respondents provide information on specific conditions with a positive response. Those answering in the affirmative are asked about the specific type of condition (if a condition is present), followed by a question on whether the condition currently limits one’s activities (Does the [condition] currently limit your activities?) with response options of “Yes, limits a little;” “Yes, limits a lot;” or “No, not currently limited by this condition.” Because the question is not specific, any affirmative answer to the limitation question roughly corresponds to the ICF’s activity and participation limitation categories.

Three unique aspects of the NLSY97 health questions are worth emphasizing. First, these questions ask about ever having a condition, rather than asking about current or recent conditions. However, the limitation status questions ask whether the respondent’s condition has caused a limitation at the time of the interview—not ever. Second, the question on mental conditions is fundamentally different from the other three condition questions because it asks about having a mental condition that is associated with a participation limitation (in work or school). Third, comparisons of disability prevalence between the NLSY97 and other surveys are tenuous because the NLSY97 questions do not correspond well to those in other surveys.

As in Mann and Honeycutt (2014), we group conditions into three broad categories: sensory conditions (the first question), physical conditions (the second and third questions), and mental conditions (the fourth question). We also construct two binary measures that capture a respondent’s limitation status (mild or severe) at the last time a respondent’s condition data were collected; that is, we construct limitation status for when condition data were last collected and the respondent did not miss that interview. Lacking more specific information, the limitation status corresponds to both the activity and participation limitations of the ICF model.

Because the NLSY97 health condition questions ask whether a condition was ever present, the NLSY97 health condition measures should in theory be cumulative over time. However, NLSY97 health condition reports frequently violate this principle: as we show later, many respondents who had a health condition reported in one round did not report ever having that condition in a subsequent round. For some of our analysis, we create and use condition status measures that are cumulative over time, regardless of changes in responses by round. However, we also present the responses to the questions at the time of the interviews in the descriptive tables. In the discussion, we consider the implications of these temporal differences in reporting health.

In addition to condition and limitation status measures and other control variables, we create four binary outcome measures: changes in condition status, changes in limitation status, first missed survey-round participation status, and attrition status. The change in condition or limitation status variables capture whether a status changed between the last two survey rounds the status was measured. We define first missed survey-round participation as the first survey round in which a respondent did not participate. All respondents participated in the initial survey round, so missed interviews can first occur in the second survey round. An attrited respondent, however, has missed more than just one survey round: he or she has stopped participating in the survey altogether. Attrition begins at the first survey round at and after which a respondent has not participated in any further survey rounds. However, because the 15th survey round is the last survey round available, we do not consider those who participated in the 14th survey round but not the 15th survey round to have left the survey. If we did, it would artificially inflate the attrition rate. Thus, we exclude all data from the 15th survey round from the attrition analyses.

### Analytical Approach

Our analysis employs three quantitative tools: (1) basic summary statistics presented in descriptive tables, (2) logistic regressions, and (3) survival analyses. In this section, we describe the logistic regression and survival analysis models. We also outline in this section how we use multiple imputation (MI) to improve our survival analyses estimates.

We specify logistic regression models to examine changes in condition or limitation status between certain survey rounds. The estimated models reveal the relationship between condition or limitation status and subsequent changes in that status. For each dependent variable, we specify and estimate separate models for examining changes between the 1st and 6th survey rounds and the 6th and 11th survey rounds. We examine the two gaps between reporting periods separately because of the differences in who reports the information at the beginning and end of each period. Within periods, we further subdivide the sample into two groups: those who did and those who did not have a condition or limitation at the beginning of the period. This additional subdivision allows us to examine separately the results for respondents acquiring or losing a condition or limitation. These models all take the same basic functional form:
$PrEVENTi=1=11+e−fifi=αXi+βZi+εi,$
1
where i is the individual youth; EVENT is a Bernoulli random variable indicating a change or decline in limitation status between the previous two condition data collection points; X is a vector containing non-health-related control variables; and Z is a vector of condition or limitation status indicators, which are the explanatory variables of interest. We measure all explanatory variables included in the models at the beginning of the period being examined and estimate these models using only those who had their condition or limitation status reported at the beginning and end of the period being analyzed. The estimated models are weighted so that the results are nationally representative.
We use survival analysis to estimate the probability that a respondent would fail to participate for the first time during a given survey round or attrite from the survey during a given round. In both cases, the survey round–specific estimate is predicated on past survey participation: in other words, the likelihood that a respondent misses a survey wave given that he or she has not missed any so far. The advantage of these models is that they take into account the respondent’s past survey participation, which is likely a strong predictor of future survey participation. The survival models we use, which are known as proportional hazard models, take the following functional form:
$limdt→0Prt≤CHANGEt|Xt,Ztdt=λt|Xt,Zt=λ0texpγXi+δZt+υt,$
2
where t is time, CHANGE is a change in survey participation status, and λ0 is the baseline hazard. The hazard function λ captures the probability of an imminent change in survey participation status relative to the probability that a change in survey participation status has not yet occurred.2 Similar to the logistic regression models, the condition and limitation status variables in Z are the explanatory variables of interest. The survival models are estimated using all NLSY97 respondents, who are weighted during estimation to make the results nationally representative. To better understand their meaning, we report the survival analysis estimation results as hazard ratios instead of coefficients.

We use MI to create condition and limitation status variable values for 1,095 NLSY97 respondents whose parents did not provide health condition status information during the first survey round. MI is a method for filling in missing data that, under the correct conditions, produces unbiased parameter estimates and properly adjusted standard errors. This technique creates multiple data sets, each with its own imputed values for the missing data. The imputed values are created using predictions from an estimated model but contain a random component. The analysis equation is estimated separately for each data set, with standard errors for the final MI parameter estimates being a function of the variability within each data set (uncertainty in the data) as well as the variability between data sets (uncertainty introduced by the imputation process). A key assumption underlying MI is that the data are missing at random (the missing data pattern is only a function of observed characteristics) or missing completely at random (the missing data pattern is not a function of any characteristics) (Rubin 1987). We do not impute information for the 6th or 11th survey rounds because most of the condition or limitation data, as well as other data, are missing as a result of survey round nonresponse, which limits our ability to use other round-specific covariates to predict condition or limitation status well. Both survival analyses use data that contain multiple imputations for missing Round 1 condition and limitation status information. The limitation status change analyses, which are confined to those who reported condition information at the beginning and end of the period being examined, and the descriptive tables do not use imputed data. We use 50 imputations in our MI process to impute missing Round 1 health condition and limitation data for 1,095 respondents. Various characteristics at survey Round 6 (such as missed interview status, attrition status, education, marital status, limitation status, and condition type) as well as characteristics at survey Round 1 (including gender, race/ethnicity, age, geographic region, urbanity, poverty level, and health status) are used to impute condition and limitation status at Round 1. The results for analyses that use imputed data are robust to other methods for addressing missing data, such as listwise deletion.

## Results

### Changes in Reported Condition and Limitation Status

In the NLSY97, youths’ conditions and limitation statuses—as reported by the respondents and their parents—evolve substantively between health condition reporting rounds. Table 1 shows the distribution of conditions and limitations for the sample used to conduct the condition and limitation change analyses. At the initial survey round, 29 % of NLSY97 respondents’ parents reported that their children ever had a health condition, and 11 % did not provide a condition status. However, at the 6th survey round, just 22 % of respondents reported having ever had a health condition. Approximately 11 % of those with, 12 % of those without, and 19 % with those missing a Round 1 condition had a missing Round 6 condition status. Among those with a Round 6 condition, a small fraction (that is, less than 0.7 %) had a missing limitation status. With about 50 % having missing information, respondents who had missing condition status information in Round 6 were relatively most likely to have missing Round 11 condition status information. However, almost all respondents who reported a condition in Round 11 also reported their limitation status. No one missed the 1st survey round, 13 % of respondents missed the 6th survey round, and 12 % missed the 11th survey round.

Condition and limitation rates for NLSY97 respondents whose statuses were reported at a given round present a somewhat different picture than do the cumulative statuses (Table 2). Changes in reports of condition status differed by condition type. The proportion of respondents who ever had mental or sensory conditions was largest at Round 1, declined by 6 to 7 percentage points at Round 6, and then increased by Round 11 (although not to Round 1 levels). The proportion of those who ever had physical conditions grew steadily over time from 13 % to 20 %. Despite the increase in the proportion reporting ever having a condition, limitation status remained fairly constant over time. The proportion of respondents with a mild limitation decreased from 10 % in Round 1 to 8 % by Round 11. Consistently across survey rounds, approximately 2.5 % to 3 % of respondents had a severe limitation.

### Predicting Changes in Condition and Limitation Status

Among those without a condition in a given round, few characteristics were associated with the acquisition of a condition in the next round. As shown in Table 3, being female was consistently positively correlated with condition acquisition for both periods. For condition status changes between Rounds 1 and 6, living in the South (relative to the Northeast) or being in fair or poor health were positively associated with reporting a condition. Respondents younger than age 18 and without a diploma (who were either still in school or had dropped out of school) were less likely to have reported a condition. Age at first interview, race/ethnicity, and household income were not associated with condition acquisition.

When we examine those with a condition in a given round, those with limitations were more likely to still have a condition in the following round. Between Rounds 1 and 6, all condition and limitation status variables were associated with being less likely to report a condition (Table 4). However, between Rounds 6 and 11, the condition variable estimates lose their significance at the p < .05 significance threshold, although they maintain negative signs. The estimates for the limitation variables were significant and negative in both models. In the Round 1 to Round 6 period, being female or in fair or poor health at Round 1 was associated with not reporting a condition.

For both observation periods, nearly all condition and limitation status variables were linked in some way with increases in limitation severity. The first two sets of columns in Table 5 include the regression results showing the association of condition and demographic characteristics with increases in limitation severity. Limitation severity increased when someone without a limitation in the first observation reported a mild or severe limitation in the second observation and also increased when a person with a mild limitation in the first observation reported a severe limitation in the second. Respondents with mental or physical conditions at Round 1 or any condition type by Round 6 were relatively more likely than those without such conditions to eventually report an increase in their limitations by the end of the period being examined. However, those who already had mild limitations were less likely to eventually report severe limitations compared with those without limitations acquiring mild or severe limitations. In addition, being female or in fair or poor health was linked with being more likely to report more severe limitations.

In contrast to increases in limitation severity, condition status was less strongly correlated with losing a severe or mild limitation (Table 5). We marked decreases in limitation severity when someone with a mild or severe limitation in the first observation reported either no limitation or (for those with a severe limitation in the first observation) a mild limitation in the second observation. Having sensory and physical conditions in Round 1 or having mental conditions by Round 6 were correlated with being less likely to have decreases in one’s limitation status in the following round. However, having a mild limitation (relative to a severe limitation) was consistently associated with being less likely to see a decrease in limitation severity. In addition, those in fair or poor health were less likely to have decreases in their limitation status from Rounds 1 to 6, whereas Hispanic respondents with limitations in Round 6 were more likely to have decreases in their limitation status by Round 11.

### Survey Response Rates and Missed Interviews

Survey response rates, missed interviews, and attrition rates changed over the course of the survey’s administration. For the entire (unweighted) sample, the survey response rate declined steadily between the 1st and 11th survey rounds, starting at 100 % and then ending at about 81 %, where it stayed through Round 15 (Table 6). After survey Round 1, during which there were no missed interviews, the first missed interview rate followed a pattern somewhat similar to that of the response rate, with a steady decline through the 11th survey round from 7 % to 2 %, followed by rates in the 1 % to 2 % range. The first missed interview rate likely declined in part because each respondent can miss a first interview only once. If a respondent continually missed interviews and never participated again, we consider that respondent to have left the sample (attrition). Sample attrition climbed steadily over time, increasing by 1 to 2 percentage points year to year (data not shown). By the 14th survey round, 15 % of respondents had left the survey.

Relative to the entire sample, those who had a condition at any point by a given survey round had higher survey response rates and appeared less likely to miss an interview or leave the sample. We include the unweighted sample sizes by survey round because the group of those who ever (or never) had a condition varied over time as more health condition information was reported. The response rate for those who had a condition at any point decreased rather steadily over the 15 survey rounds, declining from 100 % to 86 %. The first missed interview rate began at 6 % in Round 2 and decreased to under 2 % by Round 11 and thereafter. Attrition climbed steadily but at a slower rate than for the overall unweighted sample. By the 14th survey round, 11 % of those who had a condition at any point had left the survey.

Conversely, respondents who never had a condition by a given survey round typically had lower response rates and higher attrition than the overall unweighted sample. In the 11th and 15th survey rounds, the response rate for respondents who never had a condition dipped below 80 %. Starting in Round 6, those who never had a condition experienced attrition at a higher rate than did the overall sample; 16 % left the survey by the 14th survey round.

Table 7 shows the distribution of conditions and limitations for the sample used to conduct the survival analyses. The key difference between the analysis samples described in Tables 3 and 8 is that the analysis sample described in Table 8 has imputed values for all missing Round 1 condition and limitation information. Overall, differences between the two samples are minor. For instance, similar to the nonimputed sample, approximately 11 % of those with and 14 % of those without a Round 1 condition in the imputed sample had a missing Round 6 condition status.

Approximately 11 % of those with and 14 % of those without a Round 1 condition had a missing Round 6 condition status. However, among those with a Round 6 condition, a small fraction (less than 0.3 %) had a missing limitation status. With about 50 % having missing information, respondents who had missing condition status information in Round 6 were relatively most likely to have missing Round 11 condition status information. However, all respondents who reported a condition in Round 11 also reported their limitation status.

### Predicting Survey Nonresponse and Attrition

Condition and limitation statuses were typically not strongly linked with missing one’s first survey round or leaving the survey sample. Table 8 reports results from the first missed interview and attrition survival analyses, which were both estimated using imputed data. Hazard ratios with values less than 1.0 indicate that a variable is less likely to be associated with an outcome, whereas the opposite is true for hazard ratios larger than 1.0.

Only two condition variables were correlated with missing an interview, conditional on having not missed an interview before that point. Specifically, respondents who ever had a sensory (p < .05) or physical (p < .10) condition at last condition report were less likely than those who did not have such conditions to miss an interview for the first time (Table 8). Limitation status at last interview was not associated with this outcome. Those more likely to miss an interview did not have a high school diploma and were missing either urban status or health status information; those less likely to miss an interview included females, being 17 years old or younger and not having a diploma, or living in the North Central part of the continental United States.

Condition status was also not a strong predictor of survey attrition. Someone who missed reporting his or her condition information in survey Round 6 or survey Round 11 for any reason was about twice as likely to leave the survey and not return, relative to respondents who had consistently reported their condition status (Table 8). (In the other survival analysis, the missing condition information variable captured only missing condition information not related to a missing interview because the analysis was examining predictors of the first missed interview.) Among condition and limitation variables, individuals with mild limitations at the last condition report (p < .05) or who had ever had mental conditions (p < .10) were less likely to have left the sample. Several demographic characteristics were also associated with attrition. Being non-Hispanic black or Hispanic (relative to being white or affiliated with another racial or ethnic minority) was associated with a lower probability of attrition, as was being 17 years of age or younger and not having a diploma. Alternatively, not having any postsecondary degree or a high school diploma or being in the highest income group was associated with a higher probability of attrition. Having missed at least one interview in the past and having missing data for one of the demographic variables were very strong predictors of survey attrition.3

## Conclusions and Discussion

Our multivariate regression analyses highlight the dynamics of health status in adolescence, as measured by condition and limitation measures. Across 15 years of observations, results from the NLSY97 indicate that almost one-half of respondents had a health condition during adolescence or young adulthood, and almost one-quarter had a limitation in their activities associated with that condition. A sizable proportion—about 1 in 10—reported having a condition consistently across survey waves. These findings stand in contrast with the perception that health issues are a problem largely for older adults.

Changes in condition or limitation status were typically correlated with initial condition or limitation status, although there were exceptions. That health conditions are common among young adults highlights the need for continued access to health care as youth leave their homes, through both health insurance coverage (which was recently bolstered through the Affordable Care Act for those under age 26, who can remain on their parents’ health insurance plans) as well as through appropriate connections to health providers. Survey respondents’ current condition statuses and limitation statuses are often strong predictors of future changes in those statuses, suggesting that health conditions and the limitations they cause are dynamic early in the life cycle. Specifically, among those who already have a condition, having a specific condition type—mental, physical, or sensory—or a limitation is typically associated with being less likely to eventually no longer have a condition. The result that those with mild or severe limitations are less likely to have a change in condition status by Round 6 suggests that those with limitations have conditions that tend to last for longer periods and on which both the parent and youth concur; that is, having a limitation may be correlated with having a more chronic condition.

Few demographic characteristics were consistently associated with condition and limitation changes. We found that individuals who were female and in poorer health were more likely to have had condition and limitation acquisitions, less likely to have cessation, and more likely to have had increases in limitation severity, with other characteristics (such as not having a diploma or being non-Hispanic) inconsistently associated with these outcomes. The literature is not conclusive on the demographic factors associated with condition and limitation changes for youth. Although many factors are frequently associated with youth prevalence in cross-sectional studies, we might not expect similar patterns for acquisition. For example, youth who are older, male, non-Hispanic white, and have lower SES have higher rates of special health needs (Bethell et al. 2008), whereas limitations are more common for older, male, and non-Hispanic youth, with mixed patterns for socioeconomic status (Halfon and Newacheck 2010). In contrast, information from retrospective studies on disability onset for young adults point to an association with being black and having lower education levels or lower levels of household poverty, but not gender or age (Loprest and Maag 2007). The lack of consistent findings between our study and other studies may be due to the prospective nature of our study, tracking youth from a relatively earlier point in time, or a lack of statistical power to detect differences for some characteristics. Another consideration, though, could be that demographic associations with condition and limitation transitions are significant for specific conditions, but such patterns are dampened with our aggregate measures.

We also found that those who had a condition at any point by a given survey round had higher survey response rates and appeared less likely to miss an interview or leave the sample. The converse was also true. These results suggest that NLSY97 interviewers have a relatively easier time locating those with conditions or that those with conditions are more likely to cooperate with interviewers (that is, not refuse to participate in an interview). We speculated about how having a condition could affect someone’s ability to move to a new residence (inability to locate a respondent is a common reason given for why a respondent missed an interview). Although physical conditions may be associated with less mobility, mental conditions may be associated with less stability (and hence, more mobility). Thus, the combination of being young and having a condition—not just having a condition alone—may make respondents less likely to move (or move and not provide updated contact information) and therefore easier to locate.

Although new reports of conditions across rounds are suggestive of condition acquisition in the five years between rounds, reports of not having a condition after having one in an earlier round suggest the presence of measurement error. Unfortunately, it is hard to predict how our results would be affected by potential sources of measurement error. We see a large difference between Round 1 (when parents report on their child’s health condition) and Round 6 (when respondents first report on themselves). More than one-half of nonmissing cases with a condition in Round 1 no longer had a condition in Round 6. Multiple factors may account for this pattern, but the most likely reason is that the reports come from different sources. Youth and parents might have different perceptions of the youth’s health; parents might have an incentive to identify health conditions to obtain needed school or other services for the youth; and youth (older at the time of their reports) might have left school and entered the adult world, where the incentives differ for reporting health conditions. The condition-reporting differences between Rounds 6 and 11 are not as pronounced as for Rounds 1 and 6, but they are still present. Of those with data at Rounds 6 and 11, about one in five youth with a Round 6 condition did not have a condition in Round 11. Across the rounds, perhaps conditions that are less severe, less limiting, or more transient are prone to not being recalled or noted in later waves. In addition, just as disability depends on the specific context, responses about health conditions also depend on the individual’s situation at the time of the survey.

Our results on health and limitation dynamics for youth and young adults point to two important implications, one survey-related and one policy-related. First, for youth and young adults, a broad set of questions may be needed to identify those with disabilities or who are at risk of acquiring a disability. The cumulative prevalence of condition and limitation status observed into young adulthood far exceeds the survey’s point-in-time estimates, which themselves are almost twice as large as the disability rates observed in the CPS and ACS with questions on activity and participation limitations (Honeycutt and Wittenburg 2012). The questions in those national surveys may not sufficiently capture the health and disability status of a younger population—a situation that has implications for using such surveys to estimate the size of the population who might qualify for specific programs or be in need of services. Policymakers and researchers should consider that estimates from cross-sectional surveys may underestimate the proportion of youth and young adults with conditions and limitations. Second, the fact that so many youth have conditions that are limiting—or potentially limiting—during a period in which they are transitioning into the world of work raises the question of whether youth have the right supports for success and building their human capital. Although many youth with health conditions have access to services while in secondary school, after they leave high school—or if they acquire conditions afterward—the system for young adults with disabilities to promote employment and independence is highly fragmented and difficult to navigate (Currie and Kahn 2012; Osgood et al. 2010; U.S. Government Accountability Office (GAO) 2012).

Only part of our main hypothesis on survey nonresponse and attrition was supported by the analysis results. Condition or limitation status was typically not associated with first missed interview or survey attrition, although ever having had a sensory condition had some predictive power for first missed interview. This finding is consistent with other research that does not observe an association between attrition and health conditions for youth (de Graaf et al. 2013; Jeličić et al. 2010). Although sample attrition was weakly associated with certain health conditions, missing survey rounds previously was overwhelmingly the strongest predictor of leaving the sample. Thus, it seems that health condition status was not a significant factor for either missing an interview wave or overall attrition, at least for youth and young adults. This nonresult, however, might reflect the nature of the NLSY97 in following up with youth or the administration of the instrument; we encourage those administering other surveys to consider health status—and changes in health status—as a potential impediment to survey response.

The preceding conclusions should be considered in light of the study’s limitations. Gaps in health condition reporting across NLSY97 survey rounds provide us with incomplete information about condition status. Consequently, we cannot monitor how limitation status evolves from year to year. In addition, the relationships we document are not causal; that is, we did not measure the extent to which condition or limitation status causes future status changes, missed interviews, or attrition. Rather, we measured correlations between condition status and various outcomes that could address our research questions and provide helpful information to survey researchers, policymakers, and other stakeholders. Finally, the measures of condition status in the NLSY97 are broad; the inclusion of measures more typically used for identifying disability status (such as having a condition that limits or prevents work) might have resulted in different observations.

The NLSY97 provides potential opportunities to examine the dynamics of condition and limitation status for youth and young adults and how health status affects long-term transition and adult outcomes. As more years of survey data become available, it will be important to build on this and other research to document the effects of having conditions and limitations as a youth—even if just temporarily—on adult outcomes such as employment, family, and social involvement. Further, given the wealth of data available in the NLSY97, additional work in this area could identify how different environmental and educational characteristics affect these relationships, along with how they affect long-term work ability.

## Acknowledgments

The authors appreciate the assistance of Nora Paxton for programming support, Jody Schimmel Hyde for helpful comments on the analysis, and Jane Nelson for production support. Funding for this study was provided by the Research and Training Center on Disability Statistics and Demographics (StatsRRTC) at the University of New Hampshire, which is funded by the U.S. Department of Education, National Institute for Disability and Rehabilitation Research (NIDRR) (Grant No. H133B100015). The contents do not necessarily represent the policy of the U.S. Department of Education and you should not assume endorsement by the federal government (Edgar, 75.620 (b)).

## Notes

1

The weights used in the study are either included in the NLSY97 data or constructed using a program obtained from the Bureau of Labor Statistics (BLS). The NLSY97 includes the weights that make the entire NLSY97 sample nationally representative. However, for our analyses that involve subsamples of NLSY97 respondents, we must construct custom weights. The program from the BLS can create weights that make any subsample of NLSY97 respondents nationally representative.

2

A key assumption of a proportional hazard model is that when an explanatory variable’s value changes, the hazard function moves relative to the baseline hazard—an assumption that is testable (Grambsch and Therneau 1994). Results (not shown) revealed that the proportional hazard assumption was not rejected for the attrition analysis but was rejected for the first missed interview analysis. We therefore also estimated the first missed interview survival model assuming an underlying distribution for the hazard function. Using the Akaike information criterion to compare model results across distributional assumptions, we found that the Gompertz survival distribution fit best. However, the results from the Gompertz survival model did not differ qualitatively from the proportional hazard model results. Hence, to minimize the number of models we need to describe in this article, we present only the proportional hazard model results for the first missed interview analysis.

3

For a diagnostic test, in all regression analyses, we assessed the colinearity of the variables of interest using the variance inflation factor (VIF). We did not find a VIF large enough to warrant a concern of multicolinearity.

## References

Ahern, K., & Le Brocque, R. (
2005
).
Methodological issues in the effects of attrition: Simple solutions for social scientists
.
Field Methods
,
17
,
43
69
. 10.1177/1525822X04271006
Allison, P. D. (
2001
).
Missing data
.
Thousand Oaks, CA
:
Sage
.
Bethell, C. D., Read, D., Blumberg, S. J., & Newacheck, P. W. (
2008
).
What is the prevalence of children with special health care needs? Toward an understanding of variations in findings and methods across three national surveys
.
Maternal and Child Health Journal
,
12
,
1
14
. 10.1007/s10995-007-0220-5
Brault, M. W. (
2012
).
Americans with disabilities: 2010
(Current Population Report P70-131).
Washington, DC
:
U.S. Census Bureau
.
Currie, J., & Kahn, R. (
2012
).
Children with disabilities: Introducing the issue
.
Future of Children
,
22
(
1
),
3
11
. 10.1353/foc.2012.0001
Davey, A., Shanahan, M. J., & Schafer, J. L. (
2001
).
Correcting for selective nonresponse in the National Longitudinal Survey of Youth using multiple imputation
.
Journal of Human Resources
,
36
,
500
519
. 10.2307/3069628
de Graaf, R., van Dorsselaer, S., Tuithof, M., & ten Have, M. (
2013
).
Sociodemographic and psychiatric predictors of attrition in a prospective psychiatric epidemiological study among the general population. Result of the Netherlands Mental Health Survey and Incidence Study-2
.
Comprehensive Psychiatry
,
54
,
1131
1139
. 10.1016/j.comppsych.2013.05.012
Grambsch, P. M., & Therneau, T. M. (
1994
).
Proportional hazards tests in diagnostics based on weighted residuals
.
Biometrika
,
81
,
515
526
. 10.1093/biomet/81.3.515
Halfon, N., & Newacheck, P. W. (
2010
).
Evolving notions of childhood chronic illness
.
Journal of the American Medical Association
,
303
,
665
666
. 10.1001/jama.2010.130
Hemmeter, J., & Gilby, E. (
2009
).
The age-18 redetermination and postredetermination participation in SSI
.
Social Security Bulletin
,
69
(
4
),
1
25
.
Honeycutt, T., & Wittenburg, D. (
2012
).
Identifying transition-age youth with disabilities using existing surveys
.
Princeton, NJ
:
Mathematica Policy Research
.
Jamoom, E., Horner-Johnson, W., Suzuki, R., Andresen, E. M., & Campbell, V. A. (
2008
).
Age at disability onset and self-reported health status
.
BMC Public Health
,
8
,
10
10.1186/1471-2458-8-10
Jeličić, H., Phelps, E., & Lerner, R. M. (
2010
).
Why missing data matter in the Longitudinal Study of Adolescent Development: Using the 4-H Study to understand the uses of different missing data methods
.
,
39
,
816
835
. 10.1007/s10964-010-9542-5
Little, R. J., & Rubin, D. B. (
1987
).
Statistical analysis with missing data
.
New York, NY
:
John Wiley & Sons
.
Loprest, P. J., & Maag, E. (
2007
).
The relationship between early disability onset and education and employment
.
Journal of Vocational Rehabilitation
,
26
,
49
62
.
Macurdy, T., Mroz, T., & Gritz, R. M. (
2001
).
An evaluation of the National Longitudinal Survey of Youth
.
Journal of Human Resources
,
33
,
345
436
.
Mann, D. R., & Honeycutt, T. (
2014
).
Is timing everything? Disability onset of youth and their outcomes as young adults
.
Journal of Disability Policy Studies
,
25
,
117
129
.
Nagi, S. (
1965
).
Some conceptual issues is disability and rehabilitation
. In Sussman, M. (Ed.),
Sociology and rehabilitation
(pp.
100
113
).
Washington, DC
:
American Sociological Association
.
Osgood, W., Foster, M., & Courtney, M. E. (
2010
).
Vulnerable populations and the transition to adulthood
.
Future of Children
,
20
(
1
),
209
229
. 10.1353/foc.0.0047
Rubin, D. B. (
1976
).
Inference and missing data
.
Biometrika
,
63
,
581
592
. 10.1093/biomet/63.3.581
Rubin, D. B. (
1987
).
Multiple imputation for nonresponse in surveys
.
New York, NY
:
John Wiley & Sons
.
Schoeni, R. F., Stafford, F., McGonagle, K. A., & Andreski, P. (
2013
).
Response rates in national surveys
.
ANNALS of the American Academy of Political and Social Science
,
645
,
60
87
. 10.1177/0002716212456363
Shandra, C. L. (
2011
).
Life course transitions among adolescents with and without disabilities: A longitudinal examination of expectations and outcomes
.
International Journal of Sociology
,
41
,
67
86
. 10.2753/IJS0020-7659410104
Shandra, C. L., & Hogan, D. P. (
2008
).
School-to-work program participation and the post-high school employment of young adults with disabilities
.
Journal of Vocational Rehabilitation
,
29
,
117
130
.
Torvik, F. A., Rognmo, K., & Tambs, K. (
2012
).
Alcohol use and mental distress as predictors of non-response in a general population health survey: The HUNT Study
.
Social Psychiatry and Psychiatric Epidemiology
,
47
,
805
816
. 10.1007/s00127-011-0387-3
U.S. Government Accountability Office (U.S. GAO)
. (
2012
).
Students with disabilities: Better federal coordination could lessen challenges in the transition from high school
(Publication No. GAO-12-594).
Washington, DC
:
U.S. GAO
.
Verbugge, L. M., & Jette, A. M. (
1994
).
The disablement process
.
Social Science & Medicine
,
38
,
1
14
. 10.1016/0277-9536(94)90294-1
Weathers, R. R.II (
2009
).
The disability data landscape
. In Houtenville, A., Stapleton, D. C., Weathers, R. R.II, & Burkhauser, R. V. (Eds.),
Counting working-age people with disabilities
(pp.
27
68
).
Kalamazoo, MI
:
Upjohn Institute
.
International classification of functioning, disability and health
. (
2001
).
Geneva, Switzerland
:
WHO
.
Zhivan, N. A., Ang, A., Amaro, H., Vega, W., & Markides, K. S. (
2012
).
Ethnic/race differences in the attrition of older American survey respondents: Implications for health-related research
.
Health Services Research
,
47
,
241
254
. 10.1111/j.1475-6773.2011.01322.x