What are we modeling? An evaluation of depressive symptom trajectory models from adolescence to early midlife in the Add Health cohort

Abstract

It is critical to understand the development of depressive symptoms across life stages. Existing research has primarily explored this from a life course perspective, yielding inconsistent depressive trajectories, and raising questions as to whether life course processes best characterize the evolution of depressive symptoms across life stages. This study compares ten longitudinal models from four theoretical perspectives (life course, enduring, autoregressive, and hybrid) to identify the best-fitting, theoretically-informed model of depressive symptom development from adolescence to early midlife. Results indicate a hybrid model that combines enduring and autoregressive perspectives outperforms traditional life course models and best fits the data. This hybrid model suggests depressive symptom levels at baseline remain relatively stable across life stages, with past symptom levels predicting future levels. Additionally, it reveals racial/ethnic and gender differences in symptom levels in early adolescence, as well as racial/ethnic differences in longitudinal patterns. These findings advance theoretical understanding of depressive symptom development among US young adults across early portions of the life course.

1. Introduction

Understanding how depressive symptoms unfold across early portions of the life course, and whether there are differences in this process across sociodemographic groups, is of longstanding interest to social scientists who study mental health. Health status in adolescence and young adulthood, for example, has important implications for later life health and wellbeing (Harris, 2010), and elevated depressive symptoms in early phases of life are associated with lower educational attainment and reduced earnings in adulthood, as well as higher odds of subsequent chronic conditions, mental disorders, and premature mortality (Fletcher, 2010, 2013; Johnson et al., 2018; Keenan-Miller et al., 2007; National Academies of Sciences, Engineering, and Medicine, 2021; Needham, 2009). Higher depressive levels across life stages are documented among women and racial/ethnic minorities (e.g., Hargrove et al., 2020). Moreover, recent studies reveal differences in depressive levels across cohorts of United States (U.S.) midlife adults (Infurna et al., 2021; Luo et al., 2023), challenging conventional understanding of a “U” shaped depressive trajectory across the life course (Carr, 2023). Altogether, the implications of early life depressive symptoms for later health and wellbeing, along with possible varying patterns across cohorts and sociodemographic groups, highlight the continued importance of studying depressive symptom development throughout the life course. Such knowledge can help to identify life phases when mental health interventions may reduce development of more severe physical and mental illnesses among middle-aged and older U.S. adults, and disparities therein.

Several social science studies have investigated the evolution of depressive symptoms across life stages. Life course or developmental (hereafter: life course) theoretical perspectives have largely framed prior research, suggesting that longitudinal changes in depressive symptoms reflect the cumulative impact of exposure to various biological, social, cultural, and contextual risks across the lifespan (Harris, 2010). As such, researchers have frequently implemented longitudinal latent growth models to test life course hypotheses, yielding depressive trajectories that describe initial symptom levels and the socio-contextual factors associated with change as people age. Such studies have produced important evidence that sociodemographic and health-related factors – including race/ethnicity, gender, social support, marital status, employment status, stress exposure, parental support, excessive drinking, and smoking – are all associated with the unfolding of depressive symptoms across early portions of life (e.g., Adkins et al., 2008; Galambos et al., 2006; Hargrove et al., 2020; Meadows et al., 2006; Walsemann et al., 2009; T. Wickrama and Wickrama, 2010). Prior studies, however, also reveal a variety of patterns of change (e.g., linear, cubic) in depressive symptoms over time, providing inconsistent evidence of how depressive symptoms unfold from adolescence to early midlife.

It is important to accurately characterize the processes which govern changes in depressive symptoms across life stages. While life course studies have yielded nuanced descriptions of how depressive symptoms may evolve as people age, evidence from community and clinical psychology studies suggests that change processes beyond, or in combination with, life course processes may describe symptom development as well (e.g., Cerniglia et al., 2020; Ialongo et al., 2001; Kim et al., 2011). Explicit tests of alternative theories of depressive symptom change, however, are largely absent from the literature, raising questions as to whether life course processes best represent depressive symptom development. This theoretical gap has important methodological implications, as substantive theory should ideally guide researchers’ statistical modeling strategies. In the absence of clear theoretical guidance, researchers may choose longitudinal modeling strategies that are common for the field of life course studies, but fail to fully capture the unique attributes of depressive symptom change (Bianconcini and Bollen, 2018; Bollen and Gutin, 2021; Orth et al., 2021). Identifying the theorie(s) that most precisely characterize the evolution of depressive symptoms can therefore help researchers select the best types of longitudinal models to fit to their data, yielding analyses that ultimately refine our understanding of depressive trajectories and their correlates of change.

This study therefore aims to identify the best-fitting, theoretically guided model of depressive symptom development from adolescence to early mid-life, and to determine whether and how this model varies across racial/ethnic and gender groups. I use data from the National Longitudinal Study of Adolescent to Adult Health (Add Health) – one of the most commonly used datasets to answer questions about depressive symptom development – to test and compare multiple types of longitudinal models which correspond with four different theories of depressive symptom change. I then apply the best-fitting model to explore racial/ethnic and gender differences in depressive symptom development. Findings suggest a hybrid longitudinal model that combines enduring and autoregressive processes best fit the underlying data. This hybrid model suggests individuals have their own baseline level of depressive symptoms that remains relatively unchanged across life stages, while at the same time, past depressive symptom levels predict future depressive symptom levels. Findings also reveal racial/ethnic and gender differences in depressive levels in early adolescence, as well as racial/ethnic differences in the longitudinal patterns of depressive symptom development into adulthood. Altogether, this study’s findings sharpen our understanding of the longitudinal processes associated with depressive symptom development among a diverse cohort of US young adults. Moreover, the analytic approach presented here illustrates how social scientists can conduct rigorous longitudinal model comparisons to identify the most appropriate modeling strategy for their data, yielding more nuanced analyses of depressive symptom development in the US population as a whole.

2. Background

2.1. Life course perspectives on depressive symptom trajectories

Much of what is known about the patterning of depressive symptoms across life stages comes from life course studies. The life course perspective suggests the interplay of biological and socio-contextual factors, social roles, and the timing of life events and transitions, will dynamically shape the development and temporal dynamics of depressive symptoms as people age (Elder et al., 2015; Harris, 2010). Scholars across disciplines have frequently tested life course hypotheses of depressive development using longitudinal latent growth models (LLGM), which include Finite Mixture Models (FMM)¹ and growth curve models. For example, a review of the methodological approaches of life course studies of depressive trajectories, as measured by the Center for Epidemiologic Studies-Depression Scale (CES-D), in U.S. or Canadian populations under age 50 over the last 20 years reveals that they all used a LLGM approach (see Appendix Table 1).

Almost 30 % of the reviewed studies employed Finite Mixture Models (FMM) (Appendix Table 1). A key underlying assumption of FMM models is that the population of interest comprises multiple, distinct, latent subgroups (Berlin et al., 2014). These studies therefore identified the number and unique depressive symptom trajectory characteristics of two or more population subgroups over time without a priori knowledge of each group’s defining characteristics. Life course studies that used FMM models have shown substantial heterogeneity in the number of identified depressive trajectory classes (i.e., between 3 and 6 groups) and in the severity (i.e., low, medium, high) and stability (i.e., stable, increasing, decreasing) of symptom levels within classes over time (Musliner et al., 2016). Some combination of “stable high,” “stable low,” “increasing,” and “decreasing” trajectories were most often documented, with rare reports of trajectories characterized by a “U” or “reverse-U” shape (Musliner et al., 2016). Notably, scholars have critiqued FMM approaches as not always appropriate for studying a repeated measure with a nonnormal distribution (as is often the case with depressive symptom measures). In such cases, FMM approaches can erroneously produce estimates that suggest multiple latent trajectory classes exist even if the population under study is homogenous and only one latent subgroup actually exists (Bauer and Curran, 2003).

More than 70 % of the reviewed studies used methods from the second LLGM group – growth curve models – to investigate life course hypotheses related to depressive symptom change (Appendix Table 1). Most of these studies documented quadratic (“U”- or “reverse U″- shaped) or cubic (“S”-shaped) trajectories, and largely found that depressive symptoms rise from early adolescence to late adolescence, decline through the transition to adulthood and then increase across the late 20s and early 30s with the approach of midlife (Appendix Table 1). However, there are some exceptions to these patterns. A few studies document negative linear patterns, which suggest that depressive levels are highest in adolescence and steadily decline in the years leading to midlife (e.g., Galambos et al., 2006; Schieman et al., 2002). Increased depressive symptoms during adolescence have been attributed to the stressors that accompany rapid physical, emotional, and social development (Harris, 2010), while decreases in adulthood are associated with young people leaving home, completing their education, entering the labor market, and forming long-term relationships (Mirowsky and Ross, 1992).

Depressive symptom levels from adolescence to early midlife also vary across racial/ethnic groups. Most recently, Hargrove and colleagues (2020) found a cubic pattern whereby depressive symptoms rise gradually across ages 12–19, decline from ages 20–30, and then rise again from ages 31–42, with Black and Hispanic young adults exhibiting higher symptom levels than Whites across all ages under study. Earlier studies, however, documented negative linear (J. S. Brown et al., 2007; Walsemann et al., 2009) or negative quadratic (Adkins et al., 2009) associations between age and depressive symptoms, while also yielding differences in symptom levels across racial/ethnic groups with age (J. S. Brown et al., 2007; Walsemann et al., 2009). Altogether, prior research shows that racial/ethnic disparities in depressive symptoms generally emerge in early adolescence, with minoritized racial/ethnic groups experiencing higher symptom levels than Whites from adolescence to early midlife. Chronic exposure to racism-related stressors as well as systemic exclusion from social and economic resources are hypothesized explanations for these patterns (T. N. Brown et al., 2013).

Life course studies have documented gender differences in depressive symptom development from adolescence to early midlife as well. These studies show negative linear (J. S. Brown et al., 2007; Galambos et al., 2006; Galambos and Krahn, 2008; Meadows et al., 2006; Walsemann et al., 2009; Zeiders et al., 2013), quadratic (Kamis and Copeland, 2020), negative quadratic (Adkins et al., 2008, 2009; Natsuaki et al., 2009), or cubic (Hargrove et al., 2020; Rawana and Morgan, 2014; Zeiders et al., 2013) associations between age and depressive symptoms. The majority of studies reported that adolescent girls exhibited higher depressive symptom levels than adolescent boys at ages 12 to 18 and maintained higher symptom levels across their twenties and into their mid-thirties or early forties (Adkins et al., 2009; J. S. Brown et al., 2007; Galambos et al., 2006; Galambos and Krahn, 2008; Hargrove et al., 2020; Kamis and Copeland, 2020; Meadows et al., 2006; Rawana and Morgan, 2014; Walsemann et al., 2009). One study, however, documented a convergence in the symptom levels of young women and men by age 26 (Adkins et al., 2008). Research suggests that exposure to adverse childhood events, the timing of pubertal onset and menarche, and stressors related to gendered socialization into particular social roles as well as conflicts between those roles across young adulthood may help to explain why women consistently experience higher depressive symptom levels than men across portions of early life (Ge et al., 2001; Piccinelli and Wilkinson, 2000).

Collectively, the findings from life course studies reveal inconsistent evidence regarding how depressive symptoms develop and unfold from adolescence to early midlife, raising questions as to whether life course processes best characterize depressive symptom development. It is possible that these inconsistencies reflect differences in age ranges across study samples. It is also possible that other change processes beyond, or in combination with, life course processes best describe how depressive symptoms unfold, but this possibility remains underexplored. For example, none of the studies reviewed (Appendix Table 1) reported whether various types of longitudinal models were rigorously compared prior to adopting an LLGM approach (see Walsemann et al., 2009 for exception). This may reflect a tendency among researchers to test their hypotheses using modeling conventions of the field, rather than systematically testing for the presence of different forms of change in their data. Put differently, social scientists may select a longitudinal model for their analyses based on the analytical approaches of past studies, rather than systematically comparing across different types of longitudinal model specifications that emphasize different types of change to confirm their model best fits the underlying structure of their data (Bauldry and Bollen, 2018; Bianconcini and Bollen, 2018; Bollen and Gutin, 2021; Orth et al., 2021). A systematic comparison of theories of depressive symptom change could, however, refine understanding of the processes that characterize depressive development.

2.2. Alternative longitudinal perspectives of depressive symptom development

While few extant studies have directly tested alternative theories of depressive symptom development, leaving the most appropriate modeling approach unclear, this gap is not unique to mental health scholarship. For example, Bollen and Gutin (2021) documented little consensus regarding the most theoretically appropriate trajectory of self-rated health (SRH). They attributed this finding to the near absence of prior studies which systematically compared across multiple longitudinal model specifications for SRH to ensure that the chosen modeling strategy best fit the underlying data, in addition to comparing within-model specifications (Bollen and Gutin, 2021). For example, when fitting a growth-curve model, it is standard practice to compare various functional-forms of “growth” (e.g., no-growth, linear, quadratic) before choosing a final model. This practice of making within-model comparisons guides selection towards a model that parsimoniously, yet accurately, reflects the growth dynamics observed in the data. Following Bollen and Gutin’s (2021) approach, I outline three plausible alternative perspectives of depressive symptom development below.

First, the enduring perspective hypothesizes that depressive symptoms are trait-like. Individuals may be predisposed (due to underlying genetic or personality characteristics, for example) to a certain baseline depressive symptom level that remains relatively stable or unchanged across life stages, despite the presence of external factors that could exacerbate or mitigate symptoms. There is some empirical support for this perspective (Kim et al., 2011; Lamers et al., 2012; van Eeden et al., 2019). Cerniglia and colleagues (2020), for example, identified a stable component of depression that endured over a nine-year period in a community-based sample of mothers. Moreover, two studies of adolescents found that depressive symptoms remained stable across multiple follow-up periods for up to one year (McLaughlin and King, 2015; Tram and Cole, 2006). These studies suggest that enduring processes may contribute to depressive symptom development. While almost every latent growth analysis of depressive symptom development indirectly tests the enduring perspective (e.g., a mean latent slope of zero), as described above, few if any studies have reported model fit statistics which compare unconditional intercept-only models with latent growth models. As such, we know less about the degree to which enduring processes contribute to depressive symptom development across life stages.

Second, the autoregressive perspective suggests depressive symptoms are not necessarily predictable across life stages, but are instead temporally dependent on recent states. As such, higher symptom levels in the past are predictive of higher symptom levels in the future. Within this perspective, the nature of depressive symptoms may be trait-like, state-like (reflecting the presence of an acute stressor, for example), or a combination of both. One study, for example, found that reports of depressed mood among first graders predicted their diagnosis of Major Depressive Disorder by age 14 (Ialongo et al., 2001). Another study found those with elevated depressive symptom levels during mid- and late-adolescence also reported elevated depressive symptoms in early adolescence (Ge et al., 2001). These empirical findings suggest the need to explicitly test the autoregressive perspective in studies of depressive symptom development across life stages.

Finally, a hybrid perspective suggests that two or more longitudinal processes may operate simultaneously to shape how depressive symptoms develop across portions of early life (Bollen and Gutin, 2021). For example, both autoregressive and life course processes, or enduring and autoregressive processes, could jointly characterize depressive symptom trajectories. Yet, to my knowledge, this hybrid perspective remains largely underexplored in extant studies of depressive trajectories.

Altogether, the above-reviewed scholarship suggests that in addition to the life course perspective, three alternative perspectives could plausibly describe depressive symptom development. Yet, most of these alternative perspectives have not been explicitly considered in the extant literature on depressive trajectories. This represents an important gap regarding the best way to characterize depressive symptom development across life stages. When accounting for the need to compare both across (e.g., enduring vs. hybrid perspectives) and within (e.g., linear vs. quadratic vs. cubic) longitudinal model specifications, these four theoretical perspectives yield ten plausible best-fitting models of depressive symptom development. This study rigorously tests and compares these ten longitudinal models, with the goal of refining understanding of how depressive symptoms develop across portions of the early life course and the emergence of racial/ethnic and gender disparities therein.

3. Material and methods

3.1. Data

I use data from the National Longitudinal Study of Adolescent to Adult Health (Add Health), one of the most commonly-used data sources for answering questions related to the longitudinal patterning of depressive symptoms. Add Health is a nationally representative study of U.S. adolescents who were enrolled in grades 7–12 in 1994 and 1995. Researchers followed respondents into adulthood across five waves of data collection to date. Each wave contains rich details about respondents’ sociodemographic characteristics, and their physical and mental health status (Harris et al., 2019). This analysis uses data from all five waves. Waves I (ages 11–21) and II (ages 12–22) correspond with adolescence, Wave III corresponds with the transition to adulthood (ages 18–26), Wave IV with young adulthood (ages 24–32), and Wave V with early midlife (ages 33–43). The analytic sample includes non-Hispanic White (hereafter White), non-Hispanic Black (hereafter Black), and Mexican Origin² respondents who were non-missing on the variables for race/-ethnicity, gender, nativity, and age and had valid longitudinal sampling weights (N = 9,805).

3.2. Measures

Depressive Symptoms.

Participants’ responses to the following negative affect subscale items of the Center for Epidemiologic Studies-Depression scale (CES-D) were used to construct latent variables for depressive symptoms at each wave: “I could not shake off the blues,” “I felt depressed,” “I felt happy (reverse coded),” “I felt sad,” “I felt life is not worth living” [0 = ”never/rarely” to 3 = ”most of the time”] (Radloff, 1977). Perreira and colleagues (2005) validated these items as most appropriate for detecting depressive symptom differences across U.S. racial/ethnic groups.

Controls.

All models adjust for age at Wave I (mean centered and assessed continuously), as age differences may shape initial depressive symptom levels as well as subsequent depressive symptom trajectories. Each pooled model adjusts for nativity (0 = U.S. born, 1 = immigrant), gender (0 = men, 1 = women), and race/ethnicity. Race/ethnicity was self-reported at Wave I and entered into the models as two indicator variables (0 = White, 1 = Black; 0 = White, 1 = Mexican Origin). The multigroup model, which examines racial/ethnic differences, is adjusted for age, nativity, and gender. Similarly, the multigroup model that examines gender differences is adjusted for age, nativity, and race/ethnicity.

3.3. Analytic approach

To identify the best-fitting, theoretically guided model of depressive symptom development from adolescence to early midlife, I first used confirmatory factor analysis to develop a measurement model of latent depressive symptoms at each Add Health wave. These models linked the wave’s observed depressive symptom measures (i.e., the CES-D items) to a theoretically-derived, unobserved latent variable for depressive symptoms, with the errors of the CES-D indicators only correlated through their relationship with their respective latent variable and their respective CES-D indicator at all other waves (Bollen, 1989). There is substantial heterogeneity between depressive symptoms that is not well accounted for when depressive symptom scale items are combined into a summary score (van Eeden et al., 2019). Thus, the measurement models correct for measurement error that may otherwise bias longitudinal model estimates.³

Next, I tested ten theoretically plausible longitudinal models of depressive symptoms, which collectively enabled across- (e.g., enduring vs. hybrid perspectives) and within- (e.g., quadratic vs. cubic form) model specification comparisons of the four theoretical perspectives of interest. Path diagrams of the enduring, autoregressive, and life course perspectives of depressive development, are shown in Figs. 1–3, respectively. Path diagrams of the hybrid perspective of depressive development are illustrated in Fig. 4 (life course and autoregressive perspectives) and Fig. 5 (enduring and autoregressive perspectives).

Fig. 1.

Structural Equation Model corresponding to the Enduring Perspective on Depressive Symptom Development

Note: Latent time-invariant model. Control variables influence the latent depressive symptoms variables and the intercept, and are not depicted to simplify the model diagram. Small arrows refer to the errors of the continuous underlying latent variables. L_j = latent depressive symptoms at each wave. α = the latent intercept.

Fig. 3.

Structural Equation Models corresponding with the Lifecourse Perspective on Depressive Symptom Development

Note. Growth curve models. Control variables influence the latent depressive symptom variables, intercept, and slope(s), and are not depicted to simplify the model diagrams. Small arrows refer to the errors of the continuous underlying latent variables. L_j = latent depressive symptoms at each wave. α = the latent intercept. β_i = the slope.

Fig. 4.

Hybrid Structural Equation Models which combine the Lifecourse and Autoregressive Perspectives on Depressive Symptom Development

Note. Latent Variable Autoregressive Trajectory (LV-ALT) models. Control variables influence the intercept, slope(s), and latent depressive symptom variables and are not depicted to simplify the model diagrams. Small arrows refer to the errors of the continuous underlying latent variables. L_j = latent depressive symptoms at each wave. α = the latent intercept. β_i = the slope.

Fig. 5.

Hybrid Structural Equation Models which combine the Autoregressive and Enduring Perspectives on Depressive Symptom Development

Note. LV-ALT Intercept Only Model and Latent Time Invariant Model. The intercepts of the latent depressive variables are constrained to 0 across L₂ - L₅ in the LV-ALT Intercept Only model, and are freely estimated in the Latent Time Invariant with Autoregressive model. Control variables influence the intercept and latent depressive symptom variables and are not depicted to simplify the model diagrams. Small arrows refer to the errors of the continuous underlying latent variables. L_j = latent depressive symptoms at each wave. α = the latent intercept.

I first test the enduring perspective using a latent time invariant model (Fig. 1), which corresponds with equation (1):

$$L_{it}={\alpha }_t+{\xi }_i+{\epsilon }_{it}$$ (1)

In this equation, $L_{it}$ represents the variable for latent depressive symptoms for an individual, $i$, at time, $t$; ${\alpha }_t$ represents the intercept for time; ${\xi }_i$ represents a time-invariant component that remains stable for each individual, and ${\epsilon }_{it}$ corresponds with random error. Here, the error is expected to vary across people over time, to have a mean of 0, and to not correlate with the time-invariant component of the model.

Next, I test the autoregressive perspective (Fig. 2), which is shown in equation (2):

$$L_{it}={\alpha }_t+{\rho }_{t,t-1}{L}_{i,t-1}+{\epsilon }_{it}$$ (2)

Fig. 2.

A Structural Equation Model corresponding with the Autoregressive Perspective on Depressive Symptom Development

Note. Autoregressive Trajectory Model. Control variables influence the latent depressive symptom variables and are not depicted to simplify the model diagram. Small arrows refer to the errors of the continuous underlying latent variables. L_j = latent depressive symptoms at each wave.

Once again, $L_{it}$ represents the latent variable for depressive symptoms for an individual, $i$, at time, $t$, and ${\alpha }_t$ represents the intercept for time. The time-invariant component in equation (1) is replaced with the autoregressive coefficient, ${\rho }_{t,t-1}$, which estimates the magnitude of lagged effects. ${\rho }_{t,t-1}$ is multiplied by $L_{i,t-1}$, which is the lagged value of depressive symptoms. Finally, ${\epsilon }_{it}$ is a term for random error, which is again, expected to vary across people over time and have a mean of 0. ${\epsilon }_{it}$ is also expected to remain uncorrelated with the lagged value for depressive symptoms.

I then test the life course/developmental perspective with a series of latent growth curve models (e.g., linear, quadratic). These models are depicted in panels a, b, and c of Fig. 3. As an example, I provide an equation for a linear growth curve here (equation (3)), which corresponds with panel a of Fig. 3.

$$L_{it}={\alpha }_i+{\lambda }_t{\beta }_i+{\epsilon }_{it}$$ (3)

In this equation, $L_{it}$ represents the latent variable for depressive symptoms for an individual, $i$, at time, $t$. In contrast to equation (2), the intercept in equation (3) (${\alpha }_i$) is random and person-specific, which enables each individual’s depressive symptom trajectories to start at varying levels. ${\beta }_i$ is the coefficient for the random slope. ${\beta }_i$ varies across individuals to enable depressive symptoms to change at varying rates and is multiplied by a time trend for the number of years since the first time point, ${\lambda }_t$. The final component of the model is random error $\left({\epsilon }_{it}\right)$, which varies across people over time, has a mean of 0, and does not correlate with the intercept or the slope.

I investigate nonlinear changes in depressive symptoms across time in two ways. I, first, add a quadratic term to equation (3) above to test for quadratic growth (Fig. 3, panel b). I then use a freed-loading growth model (Fig. 3, panel c) to test for non-parametric form. This entailed fixing the first two loadings on the slope factor to 0 and 1, respectively, and freely estimating the loadings at the remaining time points. A benefit of this model is that it allows the fit of the polynomial function to reflect the characteristics of the underlying data (Bollen and Curran, 2006).

Finally, I test the hybrid perspective using a series of latent variable autoregressive latent trajectory (LV-ALT) models. This flexible model allows me to simultaneously test and compare various combinations of life course and autoregressive processes (Fig. 4; Bianconcini and Bollen, 2018), as well as enduring and autoregressive perspectives (Fig. 5; Bollen and Brand, 2010) in the absence of strong substantive theory as to how these processes might jointly shape depressive symptoms over time. An example of an LV-ALT model that combines a linear growth curve with an autoregressive process is shown in Panel a of Fig. 4, with corresponding equation (4):

$$L_{it}={\alpha }_i+{\lambda }_t{\beta }_i+{\rho }_{t,t-1}{L}_{i,t-1}+{\epsilon }_{it}$$ (4)

Here, L_it represents the latent variable for depressive symptoms for an individual, i, at time, t. The random person-specific intercept (α_i), random slope (β_i), and the time trend (λ_t) components of the model capture the life course perspective. The autoregressive perspective is captured by the autoregressive coefficient and lagged value of depressive symptoms (ρ_{t,t−1Li,t−1}). Random error (ε_it), varies across people over time, has a mean of 0, and does not correlate with the intercept, slope, or autoregressive terms.

Finally, an example of a hybrid model which combines the autoregressive and enduring perspectives is shown in panel a of Fig. 5, with corresponding equation (5):

$$L_{it}={\alpha }_t+{\xi }_i+{\rho }_{t,t-1}{L}_{i,t-1}+{\epsilon }_{it}$$ (5)

Here, $L_{it}$ represents the latent variable for depressive symptoms for an individual, $i$, at time, $t$, and ${\alpha }_t$ represents the intercept for time. ${\xi }_i$ represents the time-invariant component that remains stable for each individual. ${\rho }_{t,t-1L_{i,t-1}}$ represents the autoregressive component, and ${\epsilon }_{it}$ corresponds with random error. ${\epsilon }_{it}$ is expected to vary across people over time, to have a mean of 0, and to not correlate with the time-invariant or lagged components of the model.

To identify the best fitting model of depressive symptom development, I examined the chi-square test in combination with several other measures of ‘overall fit’: Bayesian information criterion (BIC), comparative fit index (CFI), Tucker-Lewis index (TLI), and one minus the root mean square error of approximation (1-RMSEA) (Bentler, 1990; Bollen, 1989, 1999; Steiger, 2016; Tucker and Lewis, 1973). A statistically insignificant Chi-square test provides evidence in support of the hypothesized model.⁴ Negative BIC values that are larger in magnitude also provide evidence in support of the hypothesized model (Raftery, 1995). For the CFI, TLI, and 1-RMSEA, values between 0.95 and 1 represent a better fit while values less than 0.9 represent an inadequate fit (Hu and Bentler, 1999). When the values for the CFI, TFI, and 1-RMSEA all indicated a good overall fit, I used the BIC to discriminate between models (Bollen and Gutin, 2021).

To identify whether and how the best fitting, theoretically derived longitudinal model of depressive symptoms varies across racial/ethnic and gender groups, I fit multigroup structural equation models (SEM) for race/ethnicity and gender, respectively. Notably, racial/ethnic and/or gender group differences in model coefficients could be attributable to true sociodemographic group differences in mental health status or to sociodemographic group differences in the psychometric responses to scale items. Therefore, I also tested for measurement invariance to identify whether there were meaningful differences (i.e., invariance) in the means, intercepts, and/or relationships between latent depressive symptoms as they develop across life course stages (Svetina et al., 2020). This entailed adding sequentially restrictive equality constraints to the multigroup models. Sufficient invariance allows for a direct comparison of coefficient estimates across racial and ethnic or gender groups (Cheung and Rensvold, 2002). Significant Chi-square difference tests and substantial changes in other measures of overall fit across the nested models (i.e., 1-RMSEA, CFI) were evidence of non-invariance (Satorra, 2000).

All of the above-described models were estimated using the weighted least squares mean and variance adjusted (WLSMV) estimator, which is appropriate for models with endogenous ordinal variables (i.e., the CES-D items). This estimator uses diagonally weighted least squares (DWLS) to estimate model parameters, and computes robust (Huber-White) standard errors and a mean- and variance-adjusted test statistic using the full weight matrix (Rosseel, 2012). Missingness was addressed using “pairwise” deletion (Rosseel and Jorgensen, 2019). I conducted all analyses using the SEM package “lavaan” in R (Rosseel, 2012).

4. Results

4.1. Descriptive statistics

Table 1 displays the means⁵ of the CES-D items across each life stage for the sample of pooled respondents, as well as by race/ethnicity and gender. Among the pooled sample, respondents reported an average of 0.48 depressive symptoms in adolescence. Symptoms declined across late adolescence to reach a low of 0.38 by the transition to adulthood. In young adulthood, mean depressive symptoms were higher than at any other time point ($\overline{x}=0.52$), but declined modestly to 0.49 by early midlife.

Table 1.

CES-D means of add health respondents.

	N	Adolescence (1994–1995)	Late Adolescence (1996)	Transition to Adulthood (2001–2002)	Young Adulthood (2008)	Early Midlife (2016–2018)
Overall	9,805	0.48 (0.01)	0.47 (0.01)	0.38 (0.01)	0.52 (0.01)	0.49 (0.01)
Race/Ethnicity
Black	2,248	0.53 (0.01)	0.54 (0.02)	0.44 (0.02)	0.65 (0.03)	0.54 (0.02)
Mexican Origin	770	0.55 (0.02)	0.60 (0.03)	0.43 (0.03)	0.52 (0.03)	0.48 (0.03)
White	6,787	0.46 (0.01)	0.45 (0.01)	0.37 (0.01)	0.49 (0.01)	0.48 (0.01)
Gender
Women	5,662	0.54 (0.01)	0.53 (0.01)	0.44 (0.01)	0.55 (0.01)	0.51 (0.01)
Men	4,143	0.42 (0.01)	0.40 (0.01)	0.32 (0.01)	0.49 (0.01)	0.47 (0.01)

Notes: Data are from the National Longitudinal Study of Adolescent to Adult Health (Add Health). CES-D means are derived from the available negative affect subscale items at each Wave of Add Health (range: 0–3). For the purpose of calculating descriptive statistics, I treated the CES-D as a continuous variable at each Wave, and weighted each mean using the respective Wave’s cross-sectional survey weights.

There are also subgroup differences in average depressive symptoms across life stages. Specifically, among White respondents, depressive symptoms followed the same pattern as the pooled sample. Among Black respondents, however, average depressive symptoms increased slightly from early adolescence ($\overline{x}=0.53$) to late adolescence ($\overline{x}=0.54$) before declining in the transition to adulthood ($\overline{x}=0.44$), increasing again in young adulthood ($\overline{x}=0.65$), and declining again by early midlife ($\overline{x}=0.54$). The pattern among Mexican Origin respondents was similar to that of Black respondents. In addition, average depressive symptoms among Black and Mexican Origin respondents were higher than those of White respondents at each life stage, with the exception of early midlife where average symptoms between Mexican Origin and White respondents were the same. Among women and men, the pattern of mean depressive symptoms across life stages exhibited the same pattern as that of the pooled sample, and women reported higher mean depressive symptoms than men at each life stage.

4.2. Identifying a best-fitting, theoretically derived longitudinal model of depressive symptoms

Table 2 shows the fit statistics for ten longitudinal models that correspond with the enduring, autoregressive, life course, and hybrid perspectives. The findings reveal that every model except for two (i.e., the Latent Time-Invariant model with Freely Estimated Intercepts and the Autoregressive model) fit the underlying data very well, as evidenced by their negative BICs, and CFIs, TLIs, and 1-RMSEAs above 0.95. Thus, the models that solely reflect the life course perspective, as well as the hybrid models, fit the underlying data better than the models that solely reflect the enduring or autoregressive perspectives. Importantly, the findings also suggest that the LV-ALT Intercept Only model is the best fitting, theoretically derived model. The BIC for the LV-ALT Intercept Only model was negative and exhibited the highest absolute magnitude (−738.316) of any of the tested models (Raftery, 1995). Moreover, the CFI and TLI for this model were very close to 1, and the 1-RMSEA was among the highest of any of the models tested. Collectively, these metrics provide very strong evidence that the LV-ALT Intercept Only model fit the data exceptionally well.

Table 2.

Comparison of fit statistics for pooled longitudinal models of depression (CES-D).

Model	Perspective	X 2	p	df	BIC	CFI	TLI	1-RMSEA	Notes
Latent Time-Invariant with Freely Estimated Intercepts	Enduring	4202.897	0.000	295	1419.656	0.989	0.992	0.963	Negative error variance for “I felt depressed” indicator at Wave 5
Autoregressive	Autoregressive	4799.312	0.000	293	2106.452	0.988	0.990	0.960	Negative error variance for “I felt depressed” indicator at Wave 5
Linear Growth Curve	Life Course	2204.134	0.000	285	−415.201	0.995	0.996	0.974	Negative error variance for “I felt depressed” indicator at Wave 5
Quadratic Growth Curve	Life Course	1974.347	0.000	275	−553.081	0.995	0.996	0.975	Negative error variance for “I felt depressed” indicator at Wave 5
Freed Loading Growth Curve	Life Course	1917.135	0.000	282	−674.628	0.996	0.996	0.976	Negative error variance for “I felt depressed” indicator at Wave 5
LV-ALT Linear Growth	Life Course + Autoregressive	1980.116	0.000	280	−593.265	0.995	0.996	0.974	Negative error variance for “I felt depressed” indicator at Wave 5
LV-ALT Quadratic Growth	Life Course + Autoregressive	1894.849	0.000	270	−586.626	0.995	0.996	0.975	Negative error variance for “I felt depressed” indicator at Wave 5
LV-ALT Freed Loading Growth	Life Course + Autoregressive	1917.781	0.000	289	−738.316	0.996	0.996	0.976	Negative error variance for “I felt depressed” indicator at Wave 5; Model that included slope would not converge therefore the estimates presented here reflect model with no slope.
a.LV-ALT Intercept Only	Autoregressive + Enduring	1917.781	0.000	289	−738.316	0.996	0.996	0.976	Negative error variance for “I felt depressed” indicator at Wave 5
Latent Time Invariant with Autoregressive	Autoregressive + Enduring	1985.457	0.000	286	−643.068	0.995	0.996	0.975	Negative error variance for “I felt depressed” indicator at Wave 5

Notes. National Longitudinal Study of Adolescent to Adult Health, Waves I-V, N = 9805, All models adjust for age at the first wave, race/ethnicity, gender, and nativity.

^a.The LV-ALT Freed Loading Growth model did not reach convergence when the slope was included in the model, but did converge when the slope was removed from the model. Removing the slope from the model reduced the functional form of the LV-ALT Freed Loading Growth model to the LV-ALT Intercept Only Model.

The LV-ALT Intercept Only model suggests depressive symptoms unfold from adolescence to early midlife in a hybrid pattern that incorporates both enduring and autoregressive processes (Fig. 5, Panel a). Specifically, this model assumes that latent depressive symptom levels in adolescence are predetermined and correlated with a latent model intercept. Additionally, the starting points of depressive trajectories across late adolescence to early mid-life vary across individual respondents, with enduring and autoregressive processes collectively predicting depressive symptom levels at each subsequent life stage. Thus, among the pooled sample (Appendix Table 2), the negative model intercept suggests that, on average, most Add Health respondents exhibit low depressive symptom levels, or no symptoms at all, in adolescence, and their baseline symptom levels remain relatively unchanged across life course stages. Additionally, the positive regression coefficients at each life stage, net of the influence of the latent intercept, suggest that higher depressive symptom levels at one life stage are associated with higher depressive symptom levels at subsequent life stages. While these effect sizes are somewhat small (Gomer et al., 2019), they retain substantive meaning, as small effect sizes are commonly observed in longitudinal autoregressive models. This is in part because controlling for “stability” at prior time points reduces the magnitude of the effect at later time points (Adachi and Willoughby, 2015). Moreover, differences in the magnitudes of the coefficients between adolescence and late adolescence, and young adulthood and early midlife may reflect variation in the number of years between waves of data collection within Add Health.⁶

4.3. Racial/ethnic differences

Recall that the multigroup models and corresponding invariance tests reveal whether the relationships between the observed and latent variables in the LV-ALT Intercept Only model described above are equivalent across Black, Mexican Origin, and White respondents. Moreover, sufficient model invariance facilitates unambiguous, direct comparisons of the estimates derived for each racial/ethnic group. The measures of global fit across increasingly constrained models show sufficient invariance to interpret any racial/ethnic differences in the model coefficients (Appendix Table 2).

As such, I present the coefficients from the unconstrained racial/ethnic multigroup model in Table 3. First, the findings show differences in the intercepts, suggesting initial symptom levels are highest among Mexican Origin adolescents and lowest among White adolescents. Second, the regression coefficients show that net of the latent intercept, there are racial/ethnic differences in the longitudinal patterns by which depressive symptoms develop across adolescence to early midlife, particularly among Mexican Origin as compared with Black and White respondents. Among Mexican Origin respondents there are positive associations between depressive symptoms in early and late adolescence, late adolescence and the transition to adulthood, the transition to adulthood and young adulthood, and young adulthood and early midlife. In contrast, among Black and White respondents, there were positive associations between depressive symptom levels at each life stage except for the period between late adolescence and the transition to adulthood.

Table 3.

Racial/ethnic differences in unconstrained from LV-ALT intercept only model.

	Black	Mexican Origin	White
	β (SE)	β (SE)	β (SE)
Regressions
LDep5 ~ LDep4	0.325 (0.06) ***	0.221 (0.10) *	0.249 (0.03) ***
LDep4 ~ LDep3	0.117 (0.05) *	0.211 (0.10) **	0.115 (0.03) ***
LDep3 ~ LDep2	0.052 (0.04)	0.228 (0.10) **	0.042 (0.03) †
LDep2 ~ LDep1	0.527 (0.08) ***	0.543 (0.18) ***	0.460 (0.04) ***
LDep1 ~ Age	0.083 (0.01) ***	0.074 (0.02) ***	0.080 (0.01) ***
LDep1 ~ Nativity	0.008 (0.16)	−0.004 (0.08)	0.130 (0.07) †
LDep1 ~ Women	0.428 (0.05) ***	0.341 (0.07) ***	0.278 (0.03) ***
Intercepts
LV-ALT Intercept Only	−0.508 (0.03) ***	−0.475 (0.06) ***	−0.558 (0.02) ***
Means
LDep1	−0.503 (0.05) ***	−0.331 (0.07) ***	−0.457 (0.03) ***
Variances
Depress1	0.224 (0.03) ***	0.122 (0.03) ***	0.169 (0.01) ***
Depress2	0.321 (0.06) ***	0.300 (0.07) ***	0.252 (0.02) ***
Depress3	0.101 (0.02) ***	0.204 (0.06) ***	0.043 (0.01) ***
Depress4	0.148 (0.02) ***	0.182 (0.04) ***	0.124 (0.01) ***
Depress5	0.156 (0.03) ***	- 0.115 (0.03) ***	0.083 (0.01) ***
Blues1	0.698 (0.06) ***	0.578 (0.08) ***	0.453 (0.03) ***
Blues2	0.698 (0.06) ***	0.518 (0.08) ***	0.375 (0.03) ***
Blues3	0.484 (0.05) ***	0.497 (0.09) ***	0.434 (0.03) ***
Blues4	0.410 (0.03) ***	0.406 (0.07) ***	0.335 (0.02) ***
Blues5	0.404 (0.04) ***	0.413 (0.06) ***	0.292 (0.02) ***
Happy1	0.815 (0.05) ***	0.644 (0.07) ***	0.449 (0.02) ***
Happy2	0.818 (0.06) ***	0.522 (0.07) ***	0.423 (0.02) ***
Happy4	0.793 (0.06) ***	0.566 (0.06) ***	0.382 (0.01) ***
Happy5	0.598 (0.04) ***	0.454 (0.04) ***	0.356 (0.01) ***
Sad1	0.190 (0.01) ***	0.176 (0.02) ***	0.164 (0.01) ***
Sad2	0.178 (0.01) ***	0.081 (0.02) ***	0.155 (0.01) ***
Sad3	0.175 (0.02) ***	0.181 (0.03) ***	0.170 (0.01) ***
Sad4	0.117 (0.01) ***	0.123 (0.02) ***	0.143 (0.01) ***
Sad5	0.166 (0.01) ***	0.046 (0.01) ***	0.124 (0.01) ***
Living1	1.376 (0.18) ***	1.062 (0.21) ***	0.999 (0.10) ***
Living2	0.722 (0.11) ***	1.277 (0.31) ***	0.914 (0.10) ***
Living5	0.795 (0.12) ***	1.282 (0.40) ***	0.679 (0.07) ***
LDep1	0.798 (0.07) ***	0.611 (0.08) ***	0.713 (0.03) ***
LDep2	0.817 (0.14) ***	0.703 (0.16) ***	0.675 (0.06) ***
LDep3	0.507 (0.06) ***	0.532 (0.11) ***	0.466 (0.03) ***
LDep4	0.561 (0.06) ***	0.556 (0.12) ***	0.501 (0.04) ***
LDep5	0.181 (0.03) ***	0.878 (0.17) ***	0.767 (0.05) ***
N	2,248	770	6,787

Notes. National Longitudinal Study of Adolescent to Adult Health, N = 9,805.

Adjusted for age, nativity, and women.

^†p < 0.10

^*p < 0.05

^**p < 0.01

^***p < 0.001.

Fit Statistics: Degrees of Freedom = 747; CFI = 0.94; TLI = 0.994; 1-RMSEA = 0.978; Robust χ² = 1957.860; BIC = −4907.546.

4.4. Gender differences

Next, I turn to the findings from the gender-multigroup model and corresponding invariance tests (Appendix Table 4). The findings from the X² difference tests in combination with the other measures of global fit derived from invariance testing suggest sufficient model invariance to interpret any gender differences in the model. A direct comparison of the coefficients in the unconstrained gender-multigroup model (Table 4) show gender differences in the intercept, collectively suggesting that initial depressive symptom levels are higher among women. Moreover, the regression coefficients reveal that the longitudinal patterns of depressive symptoms are similar for women and men across adolescence to early midlife. Specifically, both gender groups exhibited positive associations between depressive symptom levels at each subsequent life stage.

Table 4.

Gender differences in unconstrained parameters from LV-ALT intercept only model.

	Women	Men
	β (SE)	β (SE)
Regressions
LDep5 ~ LDep4	0.189 (0.03) ***	0.403 (0.05) ***
LDep4 ~ LDep3	0.075 (0.03) **	0.189 (0.05) ***
LDep3 ~ LDep2	0.043 (0.02) *	0.096 (0.03) **
LDep2 ~ LDep1	0.502 (0.04) ***	0.539 (0.05) ***
LDep1 ~ Age	0.065 (0.01) ***	0.102 (0.01) ***
LDep1 ~ Nativity	0.075 (0.07)	0.073 (0.07)
LDep1 ~ Black	0.163 (0.03) ***	0.071 (0.04) †
LDep1 ~ Mexican Origin Intercepts	0.294 (0.05) ***	0.226 (0.06) ***
LV-ALT Intercept Only Means	−0.447 (0.02) ***	−0.500 (0.03) ***
LDep1 Variances	−0.142 (0.02) ***	−0.447 (0.03) ***
Depress1	0.159 (0.01) ***	0.202 (0.02) ***
Depress2	0.306 (0.03) ***	0.284 (0.03) ***
Depress3	0.060 (0.01) ***	0.070 (0.02) ***
Depress4	0.112 (0.01) ***	0.156 (0.02) ***
Depress5	0.083 (0.01) ***	−0.048 (0.02) **
Blues1	0.497 (0.03) ***	0.605 (0.05) ***
Blues2	0.455 (0.03) ***	0.435 (0.04) ***
Blues3	0.426 (0.03) ***	0.513 (0.05) ***
Blues4	0.343 (0.02) ***	0.399 (0.03) ***
Blues5	0.260 (0.02) ***	0.202 (0.02) ***
Happy1	0.570 (0.02) ***	0.491 (0.02) ***
Happy2	0.519 (0.02) ***	0.486 (0.02) ***
Happy4	0.503 (0.02) ***	0.434 (0.02) ***
Happy5	0.408 (0.02) ***	0.426 (0.02) ***
Sad1	0.161 (0.01) ***	0.188 (0.01) ***
Sad2	0.132 (0.01) ***	0.198 (0.01) ***
Sad3	0.174 (0.01) ***	0.182 (0.01) ***
Sad4	0.127 (0.01) ***	0.147 (0.01) ***
Sad5	0.128 (0.01) ***	0.164 (0.01) ***
Living1	1.146 (0.10) ***	1.032 (0.12) ***
Living2	0.855 (0.09) ***	0.957 (0.14) ***
Living5	0.756 (0.08) ***	0.820 (0.10) ***
LDep1	0.721 (0.03) ***	0.661 (0.05) ***
LDep2	0.834 (0.07) ***	0.653 (0.07) ***
LDep3	0.536 (0.04) ***	0.456 (0.04) ***
LDep4	0.513 (0.04) ***	0.458 (0.04) ***
LDep5	0.751 (0.05) ***	0.957 (0.08) ***
N	5,662	4,143

Notes. National Longitudinal Study of Adolescent to Adult Health, N = 9,805.

Adjusted for age, nativity, and women.

^†p < 0.10

^*p < 0.05

^**p < 0.01

^***p < 0.001.

Fit Statistics: Degrees of Freedom = 538; CFI = 0.994; TLI = 0.995; 1-RMSEA = 0.978; Robust χ² = 1825.138; BIC = −3118.655.

5. Discussion

Social scientists and mental health scholars have longstanding interests in understanding the evolution of depressive symptoms across early portions of the life course and variations in these patterns across sociodemographic groups. Extant research across social science disciplines has largely investigated depressive symptom development through a life course lens. These studies have yielded nuanced insights into how depressive symptoms may change as people age, and advanced understanding of the sociodemographic, social, and health-related factors that are correlated with change over time (Adkins et al., 2008, 2009; J. S. Brown et al., 2007; Chen et al., 2011; Eagleton et al., 2016; Estrada-Martínez et al., 2019; Ferro et al., 2015a; Finan et al., 2018; Galambos et al., 2006; Galambos and Krahn, 2008; Hargrove et al., 2020; Kamis and Copeland, 2020; Marshal et al., 2013; Meadows et al., 2006; Natsuaki et al., 2009; Rawana and Morgan, 2014; Walsemann et al., 2009; K. a. S. Wickrama et al., 2009; Zeiders et al., 2013). Prior research, however, has also yielded inconsistent trajectories describing how depressive symptoms unfold across life stages. This raises questions as to whether life course processes best describe changes in depressive symptoms over time and indicates a need to test alternative theories of the change processes that may characterize depressive development.

This study addressed this open question by systematically comparing ten different types of longitudinal models derived from the life course, enduring, autoregressive, and hybrid perspectives of depressive symptom development within the Add Health cohort. This study’s primary contribution is the identification of a best-fitting, theoretically guided model of how depressive symptoms unfold from adolescence to early midlife: the LV-ALT Intercept Only model. This hybrid longitudinal model combines enduring and autoregressive properties, suggesting that individuals have their own baseline level of depressive symptoms that remains relatively unchanged across life course stages, while at the same time, past depressive symptom levels predict future depressive symptom levels. This finding adds an important twist to extant scholarship and aligns with empirical findings from community psychology studies (Cerniglia et al., 2020; Ge et al., 2001; Ialongo et al., 2001; McLaughlin and King, 2015; Tram and Cole, 2006), as well as findings from research conducted in populations diagnosed with clinical depression (Kim et al., 2011; Lamers et al., 2012; Minor et al., 2005; van Eeden et al., 2019).

This study’s results make an important theoretical contribution to the social science literature on mental health by suggesting that a hybrid longitudinal process may best characterize depressive symptom development across early portions of the life course. More specifically, enduring individual-level properties (e.g., personality characteristics, genetic predisposition) may contribute to a stable underlying component of depressive symptoms, making some young adults more prone towards higher depressive symptom levels than others. Net of these enduring properties, the effects of elevated depressive symptom levels at one life stage may linger across subsequent life stages. For example, experiencing a “shock” (e.g., parental incarceration, death of a friend, sudden unemployment) that elevates depressive levels in one life stage may place one at higher risk of elevated depressive symptom levels later in life. While major life events are not necessarily stable or cyclical, they can predict one’s mood for several days or several years. Thus, if individuals experience a proliferation of stressful events, their depressive symptom levels may be more reflective of their stress exposure, rather than personality characteristics or a predisposed underlying genetic trajectory. Altogether, the findings suggest that while some young people may be inclined towards a certain level of depressive symptoms that remain stable across life stages, it is also important to consider the influence of recent depressive status in depressive development across the life course.

This study also investigated whether there are racial/ethnic and gender differences in how depressive symptoms unfold across early portions of the life course. Regarding the findings for racial/ethnic differences, Mexican Origin and Black respondents had higher intercepts in early adolescence than White respondents. Moreover, among Mexican Origin, Black, and White respondents, early adolescence, the transition to adulthood, and young adulthood were all critical periods in which depressive symptoms in these life course stages were positively associated with depressive symptoms at subsequent life course stages. But, among Mexican Origin respondents, late adolescence was a critical period as well. Black and Mexican Origin adolescents and young adults may face myriad structural barriers in comparison to their White counterparts as they transition from adolescence to adulthood (Christensen et al., 2021; Pager et al., 2009; Slopen, 2016), which could increase their risk of psychological distress (Adkins et al., 2009; Walsemann et al., 2009). This finding aligns with prior scholarship that used Add Health data and documented elevated depressive symptom levels among minoritized racial/ethnic groups as compared with Whites across early portions of the life course (Adkins et al., 2009; J. S. Brown et al., 2007; Chen et al., 2011; Hargrove et al., 2020; Walsemann et al., 2009). The findings that Mexican Origin and Black respondents exhibited higher intercepts in early adolescence than White respondents suggests the need to examine the processes that contribute to depressive symptom development during childhood and adolescence. Future scholarship should examine the processes that contribute to depressive symptom development in the earliest life course stages, which may help to explain the racial/ethnic differences detected in the intercepts here.

Turning to the findings for gender, I documented differences in the intercepts of depressive symptoms such that they were higher among girls than boys in early adolescence. This finding aligns with prior research (e.g., Hargrove et al., 2020; Kamis and Copeland, 2020) and underscores the importance of this life course stage for supportive mental health intervention. At the same time, depressive symptom levels at each phase of the early life course positively predicted subsequent depressive symptom levels across life course stages, suggesting that there are minimal gender differences in the longitudinal patterns that contribute to how depressive symptoms develop among women and men. The reasons for this contrasting finding warrant future research.

This study also has important methodological implications for future research. Gaps in our understanding of whether life course processes best describe depressive trajectories may have stemmed from a tendency among researchers to adopt the modeling conventions of prior studies, rather than systematically testing for different forms of optimal change in their data (Bianconcini and Bollen, 2018; Bollen and Gutin, 2021; Orth et al., 2021). While researchers are often most interested in analyzing the correlates of depressive symptom development, this analysis underscores the importance of testing multiple types of longitudinal models, and selecting the model which best fits the underlying data, as a first analytical step. For example, by rigorously testing across and within multiple longitudinal model specifications, the findings confirmed that prior life course studies that used latent growth curve models provide valuable information about how depressive symptoms unfold. However, the findings also revealed that a hybrid model outperformed latent growth curve models, refining theoretical understanding of the processes that may govern changes in depressive symptoms among U.S. young adults over time. As illustrated here, general panel and autoregressive latent trajectory (ALT) models provide a flexible framework that researchers can use to facilitate careful model selection (Bauldry and Bollen, 2018; Bianconcini, 2012; Bianconcini and Bollen, 2018; Bollen and Brand, 2010; Bollen and Curran, 2004, 2006). Carefully comparing various types of unconditional longitudinal models for depressive symptoms can help researchers identify the modeling assumptions that are most appropriate for their data (Bollen and Gutin, 2021). Better models can ultimately help researchers more precisely characterize and interpret the processes that contribute to depressive symptom development across life stages.

This study is not without limitations. First, the conclusions presented here are based on analysis of five waves of unequally spaced data. Data structure limitations prevented the use of alternative time metrics, such as chronological age. Use of wave as the metric for time, however, may have prevented detection of granular temporal changes in depressive symptoms across life stages. Additional research with more waves of data is needed to more precisely assess the evolution of depressive symptoms across life stages, and to confirm that such changes are primarily driven by enduring and autoregressive processes.

Second, to test the life course perspective, this analysis focused on comparing latent growth curves given their prominence in the depressive trajectory literature (Appendix Table 1). While examination of FMM models was beyond the scope of the present study, this represents a notable limitation which should be addressed in future research. Third, the computational complexity of the models presented here precluded examination of group differences based on intersections of race/ethnicity and gender. Prior work by Hargrove and colleagues (2020), for example, found racial/ethnic-gender heterogeneity in depressive symptoms across portions of the early life course. Future scholarship on alternative longitudinal models of depressive symptom development should also test for subgroup differences across intersections of race, gender, and other axes of oppression/privilege. Fourth, this study only examined a single mental health outcome, and the longitudinal patterns identified here may not be generalizable to other types of mental health symptoms and conditions. Fifth, it is possible that the findings presented here are not generalizable to other birth cohorts or in other geographic contexts; therefore, future scholarship should replicate this study using other datasets.

Collectively, this study identified a hybrid model that combines autoregressive and enduring properties as the best-fitting, theoretically guided model of depressive symptom development across early portions of the life course in the Add Health cohort. This model suggests depressive development reflects a combination of enduring traits which predispose individuals towards certain depressive symptom levels across life course stages, net of the lingering effects of recent depressive status. The findings also suggest there are racial/ethnic and gender differences in depressive symptom levels in early adolescence and racial/ethnic differences, but no gender differences, in the patterning of depressive symptoms across adolescence to early midlife. Altogether, this study’s findings sharpen theoretical understanding of the longitudinal processes that characterize depressive trajectories across early portions of the life course.

Appendix 1

Longitudinal Models of Depressive Symptoms (CES-D) across Adolescence to Early Mid-Life in Prior Research

Study	Models Considered	Final Model	Comparison across Different Longitudinal Model Specifications	Correction for Measurement Error
Longitudinal Finite Mixture (FMM) Models
Wickrama and Wickrama (2010)	Random effects (GMM: 1, 2, 3, 4, 5 classes)	Random effects (GMM 4 class solution)	No	No
Briere et al. (2015)	Random effects (GMM: 1, 2, 3, 4, 5, 6, 7, 8 classes)	Random effects (GMM, 5 classes)	No	No
Ferro et al. (2015b)	Fixed effects (LCGA: 3, 4 classes)	Fixed effects (LCGA: 3 groups)	No	No
a Musliner et al., (2016)	Reviewed 25 FMM studies of depressive symptoms		No	No
Ellis et al. (2017)	Fixed effects (LCGA:1, 2, 3, 4, 5, 6 classes), Random effects (GMM: 4, 5 classes)	Random effects (GMM: 4 classes)	Nob	No
Essau et al. (2020)	Random effects (GMM: # of classes tested not reported)	Random effects (GMM: 3 classes)	No	No
Kent and Bradshaw (2020)	Fixed effects (GBTM: # of classes tested not reported)	Fixed effects (GBTM: 4 classes)	No	No
Minh et al. (2021)	Random effects (GMM: # of classes tested not reported)	Random effects (GMM: 4 classes)	No	No
Growth Curve Models
Galambos et al. (2006)	Random effects (linear)	Random effects (linear)	No	No
Meadows et al. (2006)	Random effects (linear)	Random effects (linear)	No	No
Brown et al. (2007)	Random effects (linear)	Random effects (linear)	No	No
Adkins et al. (2008)	Random effects (linear, quadratic, cubic)	Random effects (quadratic)	No	No
Galambos and Krahn (2008)	Random effects (linear)	Random effects (linear)	No	No
Adkins et al. (2009)	Random effects (linear, quadratic, piecewise linear)	Random effects (quadratic)	No	No
Natsuaki et al. (2009)	Random effects (linear, quadratic)	Random effects (quadratic)	No	No
Walsemann et al. (2009)	Fixed effects, Random effects (linear)	Random effects (linear)	Yes (not shown)	No
Wickrama et al. (2009)	Random effects (linear, quadratic, piecewise linear)	Random effects (piecewise linear)	No	No
Chen et al. (2011)	Random effects (linear)	Random effects (linear)	No	No
Marshal et al. (2013)	Random effects (linear, freed loading)	Random effects (linear)	No	No
Zeiders et al. (2013)	Random effects (linear, quadratic, cubic)	Random effects (cubic)	No	No
Rawana and Morgan (2014)	Random effects (linear, quadratic, cubic)	Random effects (cubic)	No	No
Ferro et al. (2015a)	Random effects (linear, quadratic, cubic)	Random effects (cubic)	No	No
Eagleton et al. (2016)	Random effects (quadratic)	Random effects (quadratic)	No	No
Finan et al. (2018)	Random effects (linear, quadratic)	Random effects (quadratic)	No	No
Estrada-Martínez et al. (2019)	Random effects (linear, quadratic, cubic)	Random effects (cubic)	No	No
Hargrove et al. (2020)	Random effects (linear, quadratic, cubic)	Random effects (cubic)	No	No
Kamis and Copeland (2020)	Random effects (linear, quadratic)	Random effects (quadratic)	No	No

^aFor parsimony, this table excludes descriptions of the LFMM studies included in the thorough review conducted by Musliner et al. (2016).

^bModel selection for this study proceeded in a stepwise fashion whereby the researchers analyzed the best fitting models from the LCGA using GMM.

Appendix Table 2.

Select Parameters from Pooled LV-ALT Intercept Only Model

	Estimate	Standard Error	P-value
Regressions
LDep5 ~ LDep4	0.284	0.03	0.000
LDep4 ~ LDep3	0.110	0.03	0.000
LDep3 ~ LDep2	0.049	0.02	0.011
LDep2 ~ LDep1	0.481	0.03	0.000
Intercepts
LV-ALT Intercept Only	−0.565	0.01	0.000
Mean
LDep1	−0.484	0.02	0.000
R-Squared
Depress1	0.811
Depress2	0.827
Depress3	0.914
Depress4	0.858
Depress5	–
Blues1	0.633
Blues2	0.656
Blues3	0.666
Blues4	0.712
Blues5	0.893
Happy1	0.312
Happy2	0.332
Happy4	0.520
Happy5	0.455
Sad1	0.652
Sad2	0.677
Sad3	0.694
Sad4	0.699
Sad5	0.665
Living1	0.527
Living2	0.530
Living5	0.642
LDep1	0.067
LDep2	0.461
LDep3	0.343
LDep4	0.348
LDep5	0.307

Notes. National Longitudinal Study of Adolescent to Adult Health, Waves I-V, N = 9,805.

All models adjust for age at the first wave, race/ethnicity, gender, and nativity.

Appendix Table 3.

LV-ALT Intercept Only Model Measurement Invariance Test, Race/Ethnicity

Model	DF	CFI	TLI	1-RMSEA	Robust χ2	BIC	Model Comparison	χ2 Difference Test	BIC Difference
A: No constraints except model structure	747	0.994	0.994	0.978	1957.860	−4907.546
B: Model A + All factor loadings constrained to equality	781	0.994	0.994	0.978	1999.882	−5178.014	B vs. A	0.164	−270.468
C: Model B + All intercepts of observed variables constrained to equality	859	0.994	0.995	0.979	2076.129	−5818.637	C vs. B	0.023*	−640.623
D: Model C + all regressions constrained to equality	879	0.994	0.995	0.980	1977.133	−6101.446	D vs. C	0.232	−282.809

Notes. National Longitudinal Study of Adolescent to Adult Health, N = 9,805, χ² difference test calculated using the Satorra (2000) method.

^†p < 0.10,

^*p < 0.05

^**p < 0.01

^***p < 0.001.

Appendix Table 4.

LV-ALT Intercept Only Model Measurement Invariance Test, Gender

Model	DF	CFI	TLI	1-RMSEA	Robust χ2	BIC	Model Comparison	χ2 Difference Test	BIC Difference
A: No constraints except model structure	538	0.994	0.995	0.978	1825.9138	−3118.655
B: Model A + All factor loadings constrained to equality	555	0.994	0.995	0.978	1914.342	−3186.467	B vs. A	0.000***	67.812
C: Model B + All intercepts of observed variables constrained to equality	594	0.993	0.994	0.976	2285.151	−3174.094	C vs. B	0.000***	−12.374
D: Model C + all regressions constrained to equality	606	0.993	0.995	0.977	2219.320	−3350.213	D vs. C	0.002 **	176.119

Notes. National Longitudinal Study of Adolescent to Adult Health, N = 9,805, χ² difference test calculated using the Satorra (2000) method.