This study adhered to the reporting standards outlined in the Preferred Reporting Items for Complex Sample Survey Analysis (PRICSSA; Seidenberg et al., 2023), a 17-item checklist designed to promote rigorous reporting of complex survey designs. Institutional Review Board (IRB) approval was obtained from the authors' affiliated institution.
Data were drawn from the National Longitudinal Study of Adolescent to Adult Health (Add Health; Harris, 2018), a nationally representative panel study. Wave I (1994–1995) surveyed 20,745 adolescents (grades 7–12) and 17,700 parents from 134 schools using a stratified sampling design based on region, urbanicity, school type, demographic composition, and size. Follow-up assessments were conducted in 1996 (Wave II), 2001–2002 (Wave III), 2007–2008 (Wave IV), and 2015–2016 (Wave V). Response rates across waves ranged from 71.8% to 88.6%.
The analytic sample (N = 7,683) included respondents who participated in Waves I, IV, and V. Only those in grades 9–12 at Wave I were retained, as the construct of decent education was defined in relation to high school. Wave III was excluded due to limited workforce participation among respondents (mean age = 23). Longitudinal weights were applied following recommendations by Chantala and Chen (2014) to adjust for attrition.
Sample demographics were as follows: 56.9% male, 43.1% female; 57.5% non-Hispanic White, 18.8% non-Hispanic Black, 16.1% Hispanic, and 6.3% Asian. Reported family income at Wave I ranged from $0 to $999,000 (M = $49,429, SD = $54,534; median = $40,000).
Descriptions below are abbreviated; full details are available in Supplemental Table 1.
Economic resources (T1) were derived from parent-reported Wave I data (Serido et al., 2014), comprising:
- Household income (continuous)
- Financial stability (binary)
- Receipt of public assistance (six binary items, reverse-coded)
A composite score was constructed using standardized values of income, financial stability, and the summed public assistance indicators.
Decent education (T1) was measured using six items from Wave I, five from the adolescent survey and one from the parent survey. Items captured key dimensions (Duffy et al., 2022):
- Physical safety
- Psychological safety
- Equitable treatment
- School belongingness
- Instructional quality
Responses were recorded on a 5-point Likert scale and reverse-coded so that higher scores reflected greater perceived educational quality.
Decent work (T2, T3) was assessed using five items from Wave IV and four items from Wave V, aligned with the five-factor model of decent work (Duffy et al., 2017). Domains included:
- Health care access (binary)
- Compensation (continuous in Wave IV; ordinal in Wave V)
- Value alignment (5-point satisfaction scale)
- Time and rest (binary and 4-point Likert)
Racial and ethnic categories were derived from Wave I responses on racial background and Hispanic origin. Participants were categorized as non-Hispanic White, non-Hispanic Black, Hispanic, or Asian. Native American and “Other” categories were excluded due to low frequency (<1%). Biological sex, reported at Wave I, was used to classify participants as male or female. Data on gender identity or sexual orientation were not available and are addressed in the study limitations.
A multigroup structural equation modeling (SEM) framework was employed to test the hypothesized model and assess variation in path estimates by race and sex. The analytic strategy followed a two-step approach (Kline, 2005):
1. Estimation of the measurement model
2. Estimation of the structural model
For longitudinal consistency, correlated residuals were specified for identical indicators of decent work across Waves IV and V (Little, 2024). All variables were coded such that higher values reflected higher levels of the latent construct.
Measurement invariance across racial and sex groups was tested in three stages:
- Configural invariance: equivalent factor structure
- Metric (weak) invariance: equivalent factor loadings
- Scalar (strong) invariance: equivalent loadings and intercepts
Following Putnick and Bornstein (2016), weak invariance was considered sufficient to permit group comparisons. Fit comparisons between models relied on both chi-square difference testing and practical fit indices, including ΔCFI < .01 and ΔRMSEA < .015 (Chen, 2007).
Analyses were conducted using Mplus 8.10 (Muthén & Muthén, 1998–2017). Indirect effects were computed via the delta method. Longitudinal sampling weights were applied as recommended by Chantala and Chen (2014).
Model fit was evaluated using the following criteria (Hu & Bentler, 1999):
- Comparative Fit Index (CFI): ≥ .90 (acceptable), ≥ .95 (good)
- Root Mean Square Error of Approximation (RMSEA): ≤ .08 (acceptable), ≤ .05 (good)
- Standardized Root Mean Square Residual (SRMR): ≤ .10 (acceptable), ≤ .05 (good)
The chi-square test statistic was not used to assess model fit due to its sensitivity to large sample sizes (Kline, 2005).