Flexibly modeling age trends in the prevalence of co-occurring patterns of substance use and mental health disorders using time-varying effects and latent class analysis

Abstract

Background:

Substance use disorders (SUDs) and mental health disorders may change and co-occur in complex patterns across adult ages, but these processes can be difficult to capture with traditional statistical approaches.

Objective:

To elucidate disorder prevalence and comorbidities across adult ages by using time-varying effect models (TVEM), latent class analysis (LCA), and modeling latent class prevalences as complex functions of age.

Methods:

Data were drawn from participants ages 18–65 years old in the National Epidemiologic Survey on Alcohol and Related Conditions III (n = 30,999; 51% women) and a subsample who reported a past-year PTSD, mood, anxiety, or SUD based on DSM-5 diagnoses (n = 11,279). TVEM and LCA were used to examine age trends and comorbidity patterns across age.

Results:

SUD prevalence peaked at age 23 (31%) and decreased thereafter, while mental health disorder prevalence was stable (20%–26% across all ages). The prevalence of five classes of individuals based on specific combinations of mental health and SUDs varied by age: the Alcohol Use Disorder class had the highest prevalence at age 26, whereas the Mood and Anxiety Disorder classes peaked around age 63. Interestingly, the Poly-Disorder class prevalence was greatest at age 18, but decreased sharply across young adulthood; however, the prevalence of the other high comorbidity class, PTSD with Mood or Anxiety Disorder, remained fairly constant across age, peaking at age 44.

Conclusions:

Multimorbid mental health disorders (excluding SUDs) persist in prevalence across adult ages. LCA, TVEM, and their integration together hold substantial potential to advance addiction research.

Worldwide, mental health and substance use disorders (SUDs) contribute significantly to morbidity and mortality and are projected to account for a $16 trillion loss over the next 20 years (1–3). A key barrier to successfully treating these disorders is their complexity, including variation in how they change across adult ages, and the reality that they often co-occur. Often, epidemiological studies report prevalence rates across wide age bands which may obscure distinct ages that present unique risks. For example, an age band of 18–29 aggregates key age-related neuro- and social- changes (e.g., legal drinking age, maturing of the prefrontal cortex) which influence SUD risk (4–6). Similarly, age bands of 30–45 or 45–65 obscure ages when life events and new social roles are common and might increase risk of depression or anxiety disorders (e.g., becoming a parent, retirement) (e.g., 7, 8, 9).

Among certain populations, multimorbidity (i.e., two or more chronic conditions) of mental health disorders and SUDs may be the rule more than the exception. Approximately 20% of persons with a past-year SUD report a mood disorder and 18% report an anxiety disorder (10, 11). The prevalence of co-occurring mood or anxiety disorder is even greater in individuals seeking treatment for SUD or alcohol use disorders (AUD). Co-occurring mood and anxiety disorders are common among people with a past-year AUD (40% and 33%, respectively) or an SUD (60% and 42%, respectively) (10). Prior work has demonstrated the role of externalizing symptoms and disorders such as ADHD or conduct disorder as important predictors of substance use initiation, often described as “deviance pathways” to substance use (12, 13). However, internalizing disorders like mood and anxiety disorders and PTSD, involving negative changes in mood and cognitions (e.g., persistent negative beliefs and thought patterns), may be particularly important in the persistence and maintenance of SUDs among adults. These internalizing disorders can exacerbate aversive feelings that reinforce goal-oriented substance use behaviors to relieve negative affect (i.e., anxiety, depressive symptoms) (14). Among adults 18 and older presenting to SUD treatment, rates of internalizing disorders have been shown to persist or increase at older adult ages, whereas, prevalence of externalizing symptoms have been shown to reduce considerably for those in their mid-twenties and continue to decrease in prevalence at older adult ages (15). Thus, a specific focus on internalizing disorders along with approaches to adequately characterize disorder heterogeneity, may better delineate how, among adults, internalizing disorders covary with SUDs.

Specifically, it is not well understood if or how disorders might co-occur to define subgroups of people with particular patterns of multimorbidity. Identification of these subgroups may be critical for tailoring interventions that fulfill both SUD and mental health treatment needs. One way to clarify this heterogeneity is to distinguish underlying subgroups of people based on specific disorders or sets of co-occurring disorders. The Diagnostic and Statistical Manual of Mental Disorders (DSM-5) is an important measurement tool for the classification of mental health disorders and SUDs, but perhaps, as limited prior work suggests, there may be complex interactions of co-occurring disorders among distinct subgroups of people (18). Further, elucidating presentations of co-occurring disorders across age may help explain shared psychopathology across disorders. These insights will inform interventions that consider both SUD and mental health disorders holistically.

Innovative statistical methods are now available to examine nuanced age trends in the prevalence of individual SUDs and comorbid disorders. Specifically, time-varying effect modeling (TVEM; 19, 20) offers flexibility in capturing change across time without requiring parametric assumptions about those changes (e.g., linear or quadratic), allowing for more precise prevalence estimates at each age to identify age-specific widows of risk. Latent class analysis (LCA; 21, 22) is ideal for identifying unique, underlying latent classes of individuals characterized by profiles of singular or co-occurring disorders. Integrating principles of TVEM with LCA can provide age-varying prevalence estimates of distinct multi-diagnosis classes; this methodological approach has been largely underutilized (23, 24). It is possible to unpack heterogeneity both in comorbidity patterns and across age by estimating latent class prevalences as a continuous function of age. In the present example, insights may help clinicians anticipate age-specific presentations of adult multimorbidity and thus inform better delivery of appropriate supports. Incidentally, our methodological demonstration may also serve to underscore the different types of questions TVEM, LCA, and concepts integrated from both frameworks may serve to answer in the study of SUDs. Thus, the current study aims to build on prior work integrating concepts from TVEM and LCA and demonstrates these approaches through three aims:

1. Use TVEM to model the age-varying prevalence of individual past-year SUDs, mood disorders, anxiety disorders, and PTSD among a nationally representative US sample of adults ages 18–65 years;

2. Use LCA to examine distinct subgroups of mood disorders, anxiety disorders, PTSD, SUDs, and their co-occurrence using LCA; and

3. Integrate concepts from TVEM into LCA to flexibly model age trends in the prevalence of disorder classes.

Methods

Sample and Procedure

The sample was drawn from the National Epidemiologic Survey on Alcohol and Related Conditions III (NESARC-III). NESARC-III is a nationally representative, cross-sectional study of noninstitutionalized US adults (25). Data collection took place from 2012 to 2013 and included in-person structured interviews and random participant callbacks to verify data. Study consent was electronic, and participants received $90 for participation. Hispanic, Black, and Asian participants were oversampled to ensure sufficient ethnic and racial representation. Sample weights were included in all analyses and descriptive statistics so that results better generalize to the US adult population. We restricted our analytic sample to include participants between the ages of 18 and 65 (n = 30,999), given the low rates of SUDs among adults older than 65; 51% of the sample were women; 64% were White, 16% Hispanic/Latinx, 12% Black, 6% Asian/Native Hawaiian/other Pacific Islander, and 2% American Indian/Alaska Native.

Measures

Past-year mental health disorders and SUDs were measured using the Alcohol Use Disorder and Associated Disabilities Interview Schedule-5 (AUDADIS-5) (26, 27). We examined three categories of mental health disorders: mood disorder (major depressive, persistent depressive, bipolar 1 disorder), anxiety disorders (generalized anxiety, specific phobia, agoraphobia, panic disorder, social anxiety disorder), and PTSD. We also examined four categories of SUDs: AUD, cannabis use disorder (CanUD), opioid use disorder (OUD) (if symptoms were related to prescription opioid or heroin use), and stimulant use disorder (STIMUD) (if symptoms were related to cocaine use, misuse of prescription stimulants, or non-prescription stimulants). The disorders that were subsumed within each disorder type—mood, anxiety, post-traumatic, and substance use disorders—were determined based on the DSM-5 classifications (28). PTSD is no longer grouped within anxiety disorders as it was it the DSM-IV, and thus is a departure from how anxiety disorders were operationalized in some earlier epidemiological studies (29, 30). Consistent with the DSM-5, mood and anxiety diagnoses excluded substance- and medical illness-induced disorders.

Statistical Analysis

To address Aim 1, we used data from all participants ages 18–65 to estimate the population-level prevalence of each past-year disorder using intercept-only weighted logistic TVEMs in SAS (19, 20, 31). To select the complexity of each nonparametric age trend (i.e., intercept function, which represents the estimated prevalence of an outcome as a continuous function of age), we compared models with 1 through 5 knots, or splitting points along the continuous function, based on the Akaike information criterion (AIC) and Bayesian information criterion (BIC). For each coefficient function, the number of knots corresponds to the complexity of the curve, where more knots correspond to more complex age trends. TVEM estimates and corresponding 95% CIs, where non-overlapping CIs between disorders provides a conservative statistical test for ages at which the prevalence rates differ, are presented as figures.

To address Aim 2, we restricted our sample to only those with a past-year mood disorder, anxiety disorder, PTSD, or SUD (36.4%; n = 11,279) and used LCA to identify key classes (i.e. “subgroups”) of common diagnosis profiles (21, 32). LCA is appropriate for uncovering underlying subgroups in a population, where subgroup membership is measured based on responses to a set of indicators. This study had seven categorical indicators: one three-level indicator for AUD (no disorder, AUD-mild [2–3 symptoms], and AUD-moderate/severe [>3 symptoms]) and six dichotomous indicators (no disorder, disorder). We differentiated severity of AUD using a three-level indicator given the high prevalence of AUD-mild among US adults, particularly young adults, and because people with AUD-mild are less likely to have psychiatric comorbidity than people with AUD-moderate/severe (33–35).

Identification of each model was assessed using 1000 sets of random starting values. Poorly identified models (operationalized here as <10% of starting value sets arriving at the maximum-likelihood solution) were not considered for final model selection. Additionally, as complex latent class models often yield one or more parameter estimates at boundary values (i.e., conditional probabilities estimated at 0 or 1), an uninformative Bayesian prior was specified in all models to enable the calculation of standard errors for all parameters. Consistent with Collins and Lanza (21), we conducted model selection using the AIC, consistent AIC (CAIC), BIC, and sample-size adjusted BIC (aBIC), where lower values indicate an improved balance between model fit and parsimony (36–39). We also inspected entropy for each model to determine how well the latent classes were differentiated, where entropy ranges from 0 to 1 and values closer to 1 generally preferred. Latent class models were estimated via the expectation-maximum (EM) algorithm using PROC LCA Version 1.3.2. (40). (For additional software information, visit https://www.latentclassanalysis.com/software/).

To address Aim 3 of examining age-varying prevalence of the disorder classes, we modeled latent class membership as a flexible polynomial function of age (i.e., covariates indicating age, age²…age⁵ were added to the latent class model selected in Aim 2 and the measurement model and coefficients for the covariates were simultaneously estimated). This is conceptually similar to the non-parametric spline basis function used in TVEM to address our first aim. We centered age at 40 years to reduce multicollinearity among the polynomial age terms. To explore the plausibility of imposing measurement invariance of the classes across age, we estimated weighted models and conducted model selection separately for three age groups: 18–29, 30–44, and 45–65 (41) and examined the interpretation of optimal models in each age group. This is consistent with procedures used in a similar analysis (24). The interpretation of the five classes was consistent across age groups, supporting our decision to estimate class membership probabilities across continuous age.

Results

Aim 1: TVEM of Individual Mental Health Disorders and SUDs Across Age

Our model selection indicated 1 knot was sufficient for capturing age trends for most disorders (2 knots for SUD overall and AUD specifically) (i.e., intercept function; see Supplementary Table 1). Figure 1, Panel A shows the estimated prevalence of any past-year mental health disorder and any past-year SUD in the overall sample. The estimated (weighted) percentage of US adults with any mental health disorder was highest among young adults and lowest among older adults (26% at age 18; 20% at age 65), with a subtle downward shift in prevalence across ages. In contrast, the estimated percentage of US adults with any SUD varied considerably with age—highest among 23-year-olds (31%) and decreasing sharply with age. SUDs were less prevalent than mental health disorders among adults in their early thirties or older, with a widening difference compared to the prevalence of mental health disorders as age increased. The rate of past-year SUD was constant at about 14% for those in their early forties through fifties, then declined with age to just 5% at age 65.

Figure 1. Mental health disorder and substance use disorder prevalence across adult ages (a) for all types and (b) by specific subtype of disorder.

Note. AUD = alcohol use disorder, CanUD = cannabis use disorder, OUD = opioid use disorder, Mood = any mood disorder, Anxiety = any anxiety disorder, and PTSD = post-traumatic stress disorder. Results presented here are estimates of past year substance use, mood, anxiety, or post-traumatic stress disorders from the full sample (n = 30,999)

The estimated percentage of US adults with each mental health disorder and SUD is presented in Figure 1, Panel B. The rates of PTSD were mostly stable and mood and anxiety disorders had a subtle downward shift across ages, whereas AUD and CanUD prevalence both peaked among younger adults and decreased sharply across age. The prevalence of past-year mental health disorder (Panel A) was largely driven by mood and anxiety disorders; anxiety disorders were somewhat was less common in younger adults but followed an age trend very similar to mood disorders from ages 25–65 (Panel B). Mood disorder was most common at age 18 (18%) and least common among older adults (11% at age 65). Anxiety disorder was most common for those in their mid-twenties (14% at age 26) and least common among older adults (12% at age 65). PTSD prevalence was 5%–6% across all ages.

The prevalence of AUD increased with age from 20% at 18 to 28% at 23 years, then declined with age steeply through about age 35, followed by a continued decrease in prevalence through age 65. Few adults had OUD and/or STIMUD, with highest percentages at age 25 for OUD (1.5%) and 24 for STIMUD (1.0%). CanUD showed a different trend, peaking among the youngest adults at about 11% at age 18 and decreasing consistently across age to 5% at 25, before plateauing at about 1% from ages 45–65. Among adults aged 45–65, the rate of CanUD was similar to that of OUD or STIMUD.

Aim 2: Latent Class Analysis of Mental Health Disorders and SUD Comorbidities

To address Aim 2, LCA was used to examine common comorbidities among 18–65-year-olds with past-year disorders (n = 11,279). We compared models with two through seven latent classes (Supplementary Table 2). The 6- and 7-class models were not considered because only 6% and 7% of random starting values converged to maximum likelihood solution (see “percentage agreement” Supplementary Table 2). We selected the 5-class model because it had considerably higher percentage agreement, as well as better fit indices and interpretability compared to models with fewer classes.

Table 1 presents weighted estimates of the class sizes and the class-specific item-response probabilities for the 5-class model. The largest class, labeled the AUD class, comprised 31% of adults and was characterized by AUD–probabilities of 0.60 for mild AUD and 0.40 for moderate/severe AUD–and very low probabilities of any other disorder. The second largest class was the Mood Disorder class (28%), characterized by a high probability of a mood disorder. The third largest class was the Anxiety Disorder class (17%), characterized by an anxiety disorder but very low probabilities of all other disorders. The two smallest classes, the PTSD with Comorbid Anxiety/Mood and Poly-Disorder classes, each comprised 12% of the population. The former was characterized by PTSD and moderate probabilities of mood (0.50) and anxiety (0.54) disorders. The latter was characterized by a relatively high probability of CanUD (0.54), with moderate probabilities of severe AUD (0.47) and anxiety disorders (0.30) and relatively high probabilities of OUD (0.17) and STIMUD (0.13).

Table 1.

Latent class analysis results for the 5-class model: Overall prevalence of past-year mental health disorders and SUDs and class-specific probabilities of having each disorder

Disorder Class:			AUD (31%)	Mood Disorder (28%)	Anxiety Disorder (17%)	PTSD with Mood/Anxiety (12%)	Poly-Disorder (12%)
Indicators	Disorder Status	Overall Prevalence	Item-Response Probabilities
Substance Use Disorders:
Alcohol Use Disorder	No	.52	0.00	0.79	0.88	0.79	0.42
	Mild	.25	0.60	0.11	0.07	0.09	0.11
	Moderate/Severe	.23	0.40	0.10	0.05	0.12	0.47
Cannabis Use Disorder	No	.92	0.96	1.00	1.00	0.96	0.46
	Yes	.08	0.04	0.00	0.00	0.04	0.54
Opioid Use Disorder	No	.97	1.00	1.00	1.00	0.97	0.83
	Yes	.03	0.00	0.00	0.00	0.03	0.17
Stimulant Use Disorder	No	.98	0.99	0.99	1.00	0.99	0.87
	Yes	.02	0.00	0.00	0.00	0.01	0.13
Mental Health Disorders:
Mood Disorder	No	.61	1.00	0.00	1.00	0.50	0.57
	Yes	.39	0.00	1.00	0.00	0.50	0.43
Anxiety Disorder	No	.62	0.96	0.67	0.00	0.46	0.70
	Yes	.38	0.04	0.33	1.00	0.54	0.30
PTSD	No	.86	1.00	1.00	1.00	0.00	0.85
	Yes	.14	0.00	0.00	0.00	1.00	0.15

Note. Latent class analysis based on those with at least one past year substance use, mood, anxiety, or post-traumatic stress disorders (n = 11,279). AUD = alcohol use disorder, PTSD = post-traumatic stress disorder.

Aim 3: Disorder Latent Class Prevalences Across Continuous Age

Age-varying probabilities of class membership for adults ages 18–65 with any past-year disorder are presented, along with 95% confidence intervals, in Figure 2. The class membership probabilities sum to 1.0 at each age. Among adults with any past-year disorder, the AUD class prevalence increased sharply from ages 18 to 26, peaked at 46%, and decreased with age thereafter, with about one-third of adults in their 40s and just 17% of adults age 65 in the AUD class. The prevalence of the Poly-Disorder class was greatest at age 18 (36%), but declined with age to just 5% at age 65. The age trends in Mood Disorder and Anxiety Disorder class prevalences were nearly identical in shape, with a slight dip in the early twenties then increasing with age through age 63. The Mood Disorder class had higher prevalence across all ages compared to the Anxiety Disorder class; prevalence of both classes peaked among those in their sixties (at age 63, 26% in the Anxiety Disorder class and 37% in the Mood Disorder class). Finally, the prevalence of the PTSD with Comorbid Mood or Anxiety class increased gradually throughout young adulthood to 16% at age 44, then increasing slightly after age 60. Age trends with especially large confidence intervals should be interpreted with caution, in particular estimates beyond about age 60. Note that the overall prevalence of any past-year disorder decreased across age (see Figure 1, Panel A, based on national sample, n = 30,999), whereas, the relative prevalence increased with age for certain latent classes (among those with a past-year disorder, n = 11,279). This difference is because the full sample includes individuals without any past-year disorder and the subsample only includes individual with at least one past-year disorder.

Figure 2. Probability of latent class membership as a polynomial function of age for adults 18–65 years old.

Note. AUD = alcohol use disorder, PTSD = post-traumatic stress disorder. Results presented here are from those with at least one past year substance use, mood, anxiety, or post-traumatic stress disorders. Note. The overall prevalence of any past-year disorder decreased across age (see Figure 1, Panel A, which depicts estimated prevalences in the entire sample), whereas, the relative prevalence increased with age for certain latent classes present here (among only those with a past year disorder, n = 11,279).

Discussion

Implications for SUD

These results demonstrate TVEM, LCA, and their integration, providing insight into age trends in disorder prevalence, the manifestation of distinct classes of individuals defined by single or multimorbid psychiatric disorders, and how the prevalence of those classes vary across ages 18 to 65. We found that SUDs are more prevalent for adults in their twenties, but by late thirties mental health disorders are more common. Further, disorder class prevalences exhibited complex age trends across adult ages.

From our analyses of individual disorders, we found enduring prevalence of mood, anxiety, and PTSD diagnoses and decreasing, non-linear shifts in the prevalence of SUDs across adult ages. These patterns are consistent with heightened early adulthood environmental and neurobiological risks for SUDs (e.g., 4, 6, 42) and possibly the persistent or recurring life problems and neurobiological risks associated with mental health disorders (e.g., 9, 43, 44). A key advantage of LCA was to reduce the number of possible comorbidity patterns (from 3×2⁶=192) to five meaningful classes. Psychiatric comorbidity was prevalent for two classes: the PTSD and Mood or Anxiety disorders class (12%) and the Poly-Disorder class (12%). The most interesting distinction between these classes was that the Poly-Disorder class primarily had co-occurring SUDs, whereas the PTSD and Mood or Anxiety class had clusters of mental health disorders, but no SUDs. The largest class were the AUD class, the Mood Disorders class, and the Anxiety Disorders class, which was expected, given the high prevalence of these disorders in the US adult population (41, 45). Finally, we modeled the age-varying relative prevalence of classes across age, demonstrating the general pattern that classes with primarily one mental health disorder (Mood Disorder class and Anxiety Disorder class) increased slightly in relative prevalence across age, while the AUD class prevalence decreased sharply in relative prevalence across age.

Most striking were the classes with high multimorbidity. The relative prevalence of the PTSD with Comorbid Anxiety or Mood Disorder classes increased slowly across age, whereas the Poly-Disorder class decreased sharply in young adulthood. The high Poly-Disorder prevalence in late teenage years and rapid decline in young adulthood might be because substance use experimentation and availability is greatest in young adulthood (46, 47). However, prevalence rates in the Poly-Disorder class rapidly declined after age 18 and did not appear to coincide with the legality of drinking alcohol at age 21, nor with accepting social norms around substance use during college-aged years (48). The peak at age 18 followed by a rapid decline may reflect that individuals in the Poly-Disorder class initiate substance use and experience greater use-related consequences earlier, leading to greater and earlier rates of treatment seeking, criminal justice system encounters, or mortality (e.g., 49, 50), all of which may contribute to the shrinking prevalence for this subgroup across ages. The PTSD and Comorbid Mood or Anxiety Disorder class slowly increased in prevalence across ages, which may be because PTSD symptoms do not often resolve quickly without treatment and may have bidirectional impacts on other mental health disorders, including the co-occurrence of mood disorder, anxiety disorders, or both. This may play a role in persistent multimorbidity across age (51–53).

In addition to mental health and SUD age-varying insights for single disorders, our approach for revealing heterogeneity from multiple indicators (i.e. mental health and substance use disorder status) across a continuous function (i.e. age) helped more closely capture the heterogeneity that clinical providers routinely encounter. In turn, clinical providers may be better informed to allocate resources and treatment services based on knowing the age-specific manifestations of these multimorbidity patterns. For example, individuals with multimorbid mental health and SUDs patients may require considerable mental health resources and often other social services to attend appointments, including financial assistance and transportation (54, 55). Older adults with multimorbid disorders may require more integrative treatment services that provide both physical and mental healthcare (56, 57). In addition, many older adults with PTSD are veterans and may benefit from veteran health care services (53, 58).

Estimating latent class prevalences as complex functions of age

The current study demonstrated a set of flexible approaches to model single and co-occurring substance use and mental health disorders across adult ages. LCA, TVEM, and their integration together hold substantial potential to advance addiction research. This set of tools is uniquely designed to unpack heterogeneity both within person, where latent classes characterize the intersection of multiple characteristics (in this case, diagnoses) of an individual, and TVEM enables highly flexible modeling of age trends. By integrating concepts from both frameworks, we demonstrated how latent class prevalences – which sum to 1.0 across classes for any given age given that the classes are mutually exclusive and exhaustive – can be estimated as a function of continuous age. Weighted results such as these can shed light on the epidemiology of intersecting conditions, allowing for flexible age trends in those intersections.

Our modeling approach not only helped reveal heterogeneous subgroups, but also offers an approach in uncovering heterogeneity as a function of a continuous variable. We used age as a continuous variable in our demonstration, but time (e.g., historical time, time post treatment) could be equally insightful for uncovering dynamic changes in subgroups longitudinally and/or before and after a treatment intervention. The substantive insights revealed from including a continuous time variable can provide new insights regarding the nature of change in class prevalences, as well as potential changes in latent class measurement, over time. These tools can help addiction researchers more deeply probe heterogeneity—and its myriad sources—both within and between people across time.

Limitations and Future Directions

Our study had several limitations, including a cross-sectional design with data collected between 2012–2013. These features preclude testing within-person changes in disorder status across adulthood and ruling out cohort effects, where disorder prevalence at certain ages may have been driven by social or environmental factors unique to a given age cohort. It will be important to replicate the LCA results in future cohorts of adults to examine potential cohort effects. Perhaps more recent data could have permitted more up-to-date estimates of disorder prevalence. However, our goal was to better characterize the intersection of substance use and other mental health diagnoses to explain relative prevalences and provide general epidemiologic insights across adult ages. Characteristics like race, socioeconomic status, and gender will be important to consider alongside these results—such considerations may be important for offering a greater degree of specificity in informing clinical care. For instance, disparities in access to treatment based on racial or socioeconomic status (e.g., 59) may lengthen the time to mental health disorder and SUD recovery. Further, replication of our results among strictly clinical samples will help validate whether the patterns of comorbidity we found here may be related to important clinical outcomes like treatment seeking and retention.

Important next steps include considering other mental health disorders that can be comorbid with SUDs, including externalizing disorders like conduct disorder and attention-deficit/hyperactivity disorder. Further, consideration of demographic and clinical characteristics may help clarify who is at greatest risk for certain patterns of SUDs and mental health disorders (e.g., the closing gender gap in AUD prevalence as an emergent cohort effect). Future studies might also consider the analytic approach demonstrated here, integrating concepts of LCA and TVEM, to capture age-varying patterns of symptoms to characterize shared etiological pathology across disorders and age. Future software and methodological advancements may include (1) estimating latent class prevalences across age using spline regression, as in traditional TVEM estimation, enabling even more flexible, nonparametric functions, and (2) making the integration of these methods more accessible to addictions researchers.