A latent class analysis of clinical complexity: Secondary analysis of the collaboration leading to addiction treatment and recovery from other stresses (CLARO) randomized trial Authors: Megan S Schuler, Rebecca Weir, Colleen M McCullough, Grace M Hindmarch, Beth Ann Griffin, Kirsten Becker, Lisa S Meredith, Katherine E Watkins Published: 2026 ABSTRACT AbstractBackground and aimsIndividuals with opioid use disorder (OUD) frequently present with co‐occurring mental health conditions such as depression and posttraumatic stress disorder (PTSD), along with other mental health substance use disorders, physical health conditions and social determinants that together comprise ‘clinical complexity.’ Collaborative care (CC), a primary care‐based behavioral health integration model, aims to improve outcomes through coordinated, patient‐centered treatment. This study examined heterogeneity in baseline clinical complexity among participants in the CLARO (Collaboration Leading to Addiction Treatment and Recovery from Other Stresses) trial and assessed whether the effects of CC versus enhanced usual care (EUC) differed across clinical complexity groups.MethodsSecondary analysis of CLARO, a pragmatic randomized clinical trial (n = 797) of CC versus EUC for adults with OUD and depression and/or PTSD, conducted in 18 low‐resourced primary care clinics in New Mexico and California, USA. Latent class analysis of baseline data identified subgroups defined by mental health and substance use as the main characteristics along with physical health and social challenges. Six‐month treatment effects were estimated using one‐step models that jointly estimated latent class membership and class‐specific outcomes, incorporating class × treatment interactions. Outcomes included depression symptom severity, PTSD symptom severity, and buprenorphine utilization and prescribing duration.ResultsThree subgroups were identified: (1) low complexity (43%), with relative clinical and social stability; (2) mental health complexity (34%), with high psychiatric symptoms but lower substance use and moderate social adversity; and (3) high dual complexity (23%). Compared with EUC, CC was associated with lower depression severity at follow‐up for the low‐complexity class, but not for other classes. CC and EUC did not differ statistically significantly with respect to PTSD outcomes in any class. Buprenorphine utilization and prescribing duration outcomes were similar across CC and EUC treatment arms.ConclusionsThe effects of collaborative care may vary according to patients' baseline clinical complexity. Collaborative care appears to be associated with more favorable depression outcomes than enhanced usual care among participants with lower clinical complexity but not among those with higher clinical complexity.Keywords: clinical complexity, collaborative care, depression, latent class analysis, opioid use disorder, posttraumatic stress disorder, primary care INTRODUCTION Despite sustained public health attention and funding, opioid use disorder (OUD) continues to kill nearly 80 000 people each year and is the leading cause of death among young adults in the USA [1, 2]. Just under half of American adults know of or have lost a friend or family member to overdose; one in eight have had their lives disrupted by a drug overdose death [3, 4]. Treatment with medication is highly effective, reducing mortality by 50%–70% [5], yet only one in four adults with OUD receives treatment, and more than half discontinue treatment prematurely, increasing the risk for relapse and mortality [6, 7]. Marked disparities in medication for OUD (MOUD) use and overdose rates exist across racial and gender groups [8, 9], underscoring the need to better understand heterogeneity in treatment access, retention and outcomes. One contributor to treatment heterogeneity is patient variation in clinical complexity, a multi‐dimensional construct spanning co‐occurring psychiatric symptoms, substance use severity, medical comorbidities and social or functional challenges [10]. Clinical complexity emerges when patient needs across personal, social and clinical domains are misaligned with the health care services available, complicating care delivery [11, 12]. In general, individuals with higher complexity require more time, coordination and resources to treat, yet health care reimbursement is typically volume based rather than needs based, resulting in a mismatch between services and demands [13]. Among people with OUD, mental health comorbidity, polysubstance use, chronic pain and social adversity are common, and each factor has been individually linked to challenges in treatment engagement, retention and recovery [14, 15, 16, 17]. However, there is limited empirical understanding of how these factors cluster together and how different clinical complexity profiles and demographic characteristics among those with OUD and co‐occurring mental illness may impact treatment outcomes. Latent class analysis (LCA) provides a robust framework for empirically identifying subgroups, or ‘phenotypes’, within heterogeneous populations [18]. In prior research, LCA has been used to characterize patterns of co‐occurring risk factors for substance use and mental health disorders [19, 20, 21], to identify clinical complexity phenotypes in various clinical contexts [22, 23], as well as to examine treatment heterogeneity across different phenotypes identified by LCA [24]. For individuals with OUD, LCA can identify complexity phenotypes that reflect intersecting psychiatric, medical and social dimensions present at treatment entry, clarifying how these factors may shape treatment response and guiding more targeted interventions. This study uses data from CLARO (Collaboration Leading to Addiction Treatment and Recovery from Other Stresses), a randomized controlled trial of collaborative care versus enhanced usual care for adults with OUD and co‐occurring depression and/or posttraumatic stress disorder (PTSD) in low‐resourced primary care clinics. We used LCA to examine heterogeneity in baseline clinical complexity among participants and assessed whether the treatment effects of collaborative care differed across clinical complexity groups. Findings from this study can inform clinical practice by identifying clinically meaningful patterns of complexity among primary care patients with OUD and co‐occurring mental health disorders and by examining whether these patterns are associated with differential treatment outcomes, thereby guiding future efforts to tailor integrated care models. METHODS Study population This study is a secondary data analysis of the CLARO trial, a pragmatic, multi‐site randomized clinical trial evaluating the effectiveness of collaborative care versus enhanced usual care for individuals with OUD and co‐occurring depression and/or PTSD [25, 26]. A total of 797 participants were enrolled in the trial and included in the analytic sample for this study. The trial was conducted in partnership with four health systems across 18 primary care clinics in New Mexico and California, most of which served low‐income populations and were in federally designated Mental Health Professional Shortage Areas. The trial was approved by the RAND Institutional Review Board and was monitored by an independent data and safety monitoring board. For more details on the original trial, please refer to the article by Meredith et al. [25]. Eligibility criteria Participants were eligible for inclusion if they were adults (aged ≥18 years) receiving primary care at one of the participating clinics who had probable OUD as well as probable depression and/or PTSD. Probable OUD was defined as current use of MOUD for opioid‐related problems within the past 90 days or a score of ≥1 on the opioid‐use items of the National Institute on Drug Abuse Tobacco, Alcohol, Prescription medication, and other Substance use (myTAPS) screening tool [27]. Probable depression was defined as a score of ≥10 on the Patient Health Questionnaire 8 (PHQ‐8; used for eligibility screening), whereas depression symptom severity at follow‐up was assessed using the PHQ‐9 (see Measures) [28]. Probable PTSD was defined as a score of ≥3 on the Primary Care PTSD Screen for Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM‐5) (PC‐PTSD‐5) [29]. Participants were also required to consider the participating clinic as their usual source of care and to speak English or Spanish. Participant recruitment Recruitment occurred between January 2021 and December 2023 using a multi‐pronged approach across 18 participating primary care clinics. Recruitment strategies included universal pre‐screening in clinic waiting rooms, provider referral of potentially eligible patients and proactive outreach based on diagnoses documented in electronic medical records. Interested individuals were then contacted by study interviewers, who conducted eligibility screenings and obtained informed consent. This approach was designed to capture a broad spectrum of patients with OUD and co‐occurring depression and/or PTSD who considered the participating clinic their usual source of care, thereby enhancing the representativeness of the study population. Intervention Participants were randomized at the individual level to either the collaborative care (CC) arm or the enhanced usual care (EUC) arm. Those randomized to CC received a structured, team‐based intervention adapted to address the needs of individuals with OUD and co‐occurring mental health conditions in primary care settings. Trained community health workers served as care managers and were responsible for symptom monitoring, coordination and communication across providers, assessment and management of social determinant needs, and referral to study‐trained behavioral health providers. Social determinants of health were assessed using the WellRx screening tool [30]. Evidence‐based treatments supported within the CC model included: pharmacotherapy for OUD (buprenorphine or methadone), depression and PTSD; motivational interviewing; problem‐solving therapy for depression [31, 32]; and written exposure therapy for PTSD [33, 34]. Care managers met with participants regularly over a period of 6 months, tracked encounters using a caseload management tool [35], and reviewed cases weekly with an addiction psychiatrist and monthly with a broader clinical team. Intervention fidelity was monitored using registry data on care manager encounters, measurement‐based assessments and psychiatric case review; fidelity metrics have been reported previously [26]. In contrast, participants randomized to EUC received standard primary care, which could include access to a community health worker not trained in the CC model. All clinics were also offered optional training for providers on treating co‐occurring disorders, including brief therapies and MOUD prescribing; providers could use this training as they saw fit, regardless of the patient’s study arm. Measures To characterize participants’ clinical complexity via LCA, we selected participant self‐reported baseline measures spanning mental health, substance use, physical health and social determinants. Latent class measures included: Mental health. Depression symptom severity was measured using the nine‐item Patient Health Questionnaire (PHQ‐9; range = 0–27) [28]. PTSD symptom severity was assessed using the PTSD checklist for DSM‐5 (PCL‐5; range = 0–80) [36]. Suicidal ideation in the past 30 days was assessed via the Columbia Suicide Severity Rating Scale (C‐SSRS) [37]. Mental health. Depression symptom severity was measured using the nine‐item Patient Health Questionnaire (PHQ‐9; range = 0–27) [28]. PTSD symptom severity was assessed using the PTSD checklist for DSM‐5 (PCL‐5; range = 0–80) [36]. Suicidal ideation in the past 30 days was assessed via the Columbia Suicide Severity Rating Scale (C‐SSRS) [37]. Substance use. The number of days with opioid and stimulant use in the past 30 days were derived from items adapted from the National Survey on Drug Use and Health (NSDUH) [38]. OUD symptom severity was assessed using the PROMIS Severity of Substance Use Disorder—Opioid Use Disorder short form, scored on a T‐score scale (mean = 50, SD = 10) [39]. Substance use. The number of days with opioid and stimulant use in the past 30 days were derived from items adapted from the National Survey on Drug Use and Health (NSDUH) [38]. OUD symptom severity was assessed using the PROMIS Severity of Substance Use Disorder—Opioid Use Disorder short form, scored on a T‐score scale (mean = 50, SD = 10) [39]. Physical health. Physical health was assessed using the physical health component score from the Veterans RAND 12‐item Health Survey (VR‐12), with higher scores indicating better functioning (range = 0–100) [40]. Pain was measured using the Pain, Enjoyment of Life and General Activities (PEG) scale (range = 0–30) [41]. Functional impairment was assessed using the Sheehan Disability Scale (range = 0–30). Physical health. Physical health was assessed using the physical health component score from the Veterans RAND 12‐item Health Survey (VR‐12), with higher scores indicating better functioning (range = 0–100) [40]. Pain was measured using the Pain, Enjoyment of Life and General Activities (PEG) scale (range = 0–30) [41]. Functional impairment was assessed using the Sheehan Disability Scale (range = 0–30). Social determinants. Indicators included lack of stable housing in the past 3 months and lack of social support, defined as not having an adult with whom they regularly connect and who is supportive of them (i.e. someone they see, call or text at least three times per week). Social determinants. Indicators included lack of stable housing in the past 3 months and lack of social support, defined as not having an adult with whom they regularly connect and who is supportive of them (i.e. someone they see, call or text at least three times per week). The primary outcomes were assessed at 6 months. Measures of mental health were collected via participant self‐report. Buprenorphine receipt and duration were derived from prescription fill records obtained from the New Mexico and the California Prescription Drug Monitoring Programs [42, 43]. Mental health outcomes Depression symptom severity (PHQ‐9, range = 0–27) was assessed among participants with probable depression at baseline; PTSD symptom severity (PCL‐5, range = 0–80) was assessed among participants with probable PTSD at baseline. Buprenorphine outcomes Buprenorphine outcomes were assessed among participants not receiving methadone at baseline. Outcomes included: (i) any buprenorphine use during follow‐up; (ii) total number of days prescribed; and (iii) coverage (i.e. the proportion of follow‐up days with an active prescription). Coverage was calculated from baseline for those entering the study on buprenorphine and from date of initiation for those who began treatment during the study. Other baseline covariate measures include: treatment group assignment (CC or EUC); age; sex (male/female, as specified on current birth certificate); race and ethnicity (non‐Hispanic White, Hispanic any race or non‐Hispanic other/multiple races); education level (less than high school, high school diploma or equivalent, or some college or more); marital status (never married, married/living with partner or widowed/divorced/separated); legal involvement (binary indicator of any current legal problems, based on Addiction Severity Index, ASI, items [44]); opioid overdose events (binary indicator of any non‐fatal opioid overdose in the past 90 days, based on the Overdose Baseline Questionnaire [45]); alcohol use severity (past 3‐month Alcohol Use Disorders Identification Test—Consumption, AUDIT‐C, score [46]); past‐month MOUD status (categorized as: no MOUD, methadone as prescribed, buprenorphine as prescribed, or MOUD received never or sometimes as prescribed); trauma exposure (based on the worst traumatic event anchoring the PCL‐5 [36, 47]: no trauma, trauma without interpersonal violence or trauma with interpersonal violence; and health system (health systems 1–4). We note that while some of these covariates overlap conceptually with constructs included in our LCA model, they were not selected as LCA indicators to avoid collinearity and to maintain model parsimony. Statistical analysis Latent class analysis (LCA) was used to identify subgroups of participants with distinct clinical and social profiles at baseline, reflecting varying levels of clinical complexity. Indicators spanned mental health, substance use, physical health and social determinants; the VR‐12 physical health indicator was reverse coded so that higher scores indicated worse functioning, ensuring comparable directionality across all indicators. Models with between two and five classes were estimated in Stata 17 using generalized structural equation modeling (gsem procedure) with full information maximum likelihood (FIML), which accommodates missing data on indicators under the missing‐at‐random assumption. Model selection considered statistical fit (log‐likelihood; Akaike information criterion, AIC; Bayesian information criterion, BIC), separation (entropy), minimum class size and clinical interpretability. Because outcomes were not necessarily assessed among all participants, reducing the effective sample for class‐stratified analyses, we prioritized solutions with larger class sizes to preserve precision and avoid unstable estimates. The three‐class solution was selected as the primary model; the four‐class model is presented for comparison in the supporting information for comparison (Table S2). For descriptive purposes, each participant was assigned to the latent class for which they had the highest posterior probability (modal assignment). Baseline characteristics were compared across classes using c2 tests for categorical variables and one‐way analysis of variance (ANOVA) for continuous variables. To control the false discovery rate across multiple comparisons, the Benjamini–Hochberg (BH) procedure was applied (q = 0.05). To evaluate whether CC and EUC had differential effects across classes, we estimated 6‐month treatment effects using a one‐step modeling approach, in which latent class membership and outcome models were estimated jointly within the gsem framework. This approach incorporates uncertainty in class assignment directly into the estimation of class‐specific treatment effects. Continuous outcomes were modeled using linear models and binary outcomes were modeled using logistic models. For mental health outcomes (PHQ‐9 and PCL‐5), non‐response weights were applied (the response rates were 65% and 57%, respectively); non‐response weights were not applicable for buprenorphine outcomes (the response rates were 96%–100%) [26]. Marginal means and contrasts were estimated using the margins command in Stata, with standard errors derived from the delta method and Wald F‐tests for inference. As a sensitivity analysis, we re‐estimated the one‐step outcome models including health system indicator variables; treatment effect estimates were substantively unchanged. This secondary analysis was not pre‐specified in the original trial protocol and was not separately pre‐registered; findings should therefore be interpreted as exploratory. All analyses were conducted in Stata 17 with a two‐sided significance level of 0.05. RESULTS Latent classes of clinical complexity We selected the three‐class solution, balancing statistical fit with parsimony, interpretability and adequate class sizes (Table 1). Model fit indices (AIC, BIC) improved steadily as the number of classes increased from two to five, but solutions with higher numbers of classes produced small subgroups (<10% of the sample) that would limit the statistical power for examining potential treatment effect heterogeneity across classes (for model fit statistics and comparison with the four‐class model, see the supporting information Tables S1 and S2). TABLE 1Clinical complexity phenotypes at baseline: latent class analysis results.Class 1 Low complexityClass 2 Mental health complexityClass 3 High dual complexity n = 352 (43%) n = 263 (34%) n = 182 (23%)Lacks stable housing (%)19.645.656.6Lacks social support (%)7.914.118.1PHQ‐9 score, mean (SD)10.3 (4.6)17.4 (4.6)15.4 (5.3)Suicidal ideation, past 30 days (%)14.846.436.8PCL‐5 score, mean (SD)29.3 (14.1)48.0 (14.5)40.8 (16.1)VR‐12 physical health (reverse‐coded), mean (SD)57.1 (11.4)69.7 (12.7)62.9 (13.7)PEG pain score, mean (SD)11.3 (8.0)21.2 (6.6)16.8 (8.6)Sheehan Disability Scale score, mean (SD)5.9 (6.4)18.4 (7.8)18.3 (8.4)OUD severity (PROMIS T‐score), mean (SD)48.7 (6.7)52.9 (7.1)62.3 (4.7)Days of opioid use (past 30 days), mean (SD)0.8 (2.7)1.7 (3.6)27.1 (5.2)Days of stimulant use (past 30 days), mean (SD)1.1 (4.1)4.5 (8.9)16.2 (12.2) Note: Class sizes and descriptive characteristics are based on modal class assignment, in which each participant is assigned to the class for which their posterior membership probability is highest. VR‐12 physical health scores were reverse‐coded so that higher values indicate worse functioning for comparability across measures.Abbreviations: OUD = opioid use disorder; PCL‐5 = PTSD checklist for Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM‐5); PEG = Pain, Enjoyment of Life and General Activities scale; PHQ‐9 = Patient Health Questionnaire 9; VR‐12 = Veterans RAND 12‐Item Health Survey. Clinical complexity phenotypes at baseline: latent class analysis results. Note: Class sizes and descriptive characteristics are based on modal class assignment, in which each participant is assigned to the class for which their posterior membership probability is highest. VR‐12 physical health scores were reverse‐coded so that higher values indicate worse functioning for comparability across measures. Abbreviations: OUD = opioid use disorder; PCL‐5 = PTSD checklist for Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM‐5); PEG = Pain, Enjoyment of Life and General Activities scale; PHQ‐9 = Patient Health Questionnaire 9; VR‐12 = Veterans RAND 12‐Item Health Survey. Class 1 The low complexity class (43% of participants) comprised the most clinically and socially stable subgroup at baseline. Individuals in this class had the lowest levels of substance use and mental health symptoms, with very low opioid use (mean = 0.8 days) and stimulant use (mean = 1.1 days) in the past month. Participants in this class had the lowest depression (PHQ‐9 = 10.3) and PTSD (PCL‐5 = 29.3) scores across groups, although 14.8% endorsed suicidal ideation. Notably, the mean PHQ‐9 still exceeded the cut point for probable depression (≥10), while the mean PCL‐5 remained below the threshold for probable PTSD (>33). Approximately one in five participants lacked stable housing (19.6%) and fewer than one in 10 lacked social support (7.9%). Functional impairment and pain levels were low (Sheehan Disability Scale, SDS = 5.9; PEG = 11.3). Class 2 Mental health complexity (34% of participants) was characterized by the most severe psychiatric burden across classes. Depression and PTSD symptoms were higher than for other classes (PHQ‐9 = 17.4; PCL‐5 = 48.0) and nearly half (46.4%) reported recent suicidal ideation. The mean number of days with substance use was 1.7 (opioid use) and 4.5 (stimulant use). Approximately half experienced housing instability (45.6%) and 14.1% reported a lack of social support. Functional disability and pain were substantial (SDS = 18.4; PEG = 21.2). Class 3 High dual complexity (23% of participants) represented the most clinically complex subgroup. Participants had elevated OUD severity (PROMIS T‐score = 62.3), near‐daily opioid use (mean = 27.1 days) and frequent stimulant use (mean = 16.2 days) in the past month. The mean mental health symptom severity exceeded cut points for probable depression and PTSD (PHQ‐9 = 15.4, cut point ≥10; PCL‐5 = 40.8, cut point >33), with 36.8% reporting suicidal ideation. Pain and disability levels were high (PEG = 16.8; SDS = 18.3). Over half of participants in this subgroup lacked stable housing (56.6%) and 18.1% lacked social support. Class differences in characteristics are visualized in Figure 1, which displays standardized deviations (i.e. z‐scores) across indicators relative to the overall sample mean. As shown, class 1 clustered below the sample average on all measures. Class 2 showed positive deviations on depression, suicidal ideation, PTSD, physical health, pain and disability, but scored near or below the mean for substance use. In contrast, class 3 deviated sharply above the mean on substance use measures, scored the highest on measures of social disadvantage and had elevated mental health measures. This visualization underscores the distinct clinical and social profiles that differentiate the three complexity phenotypes. Clinical complexity phenotypes at baseline: standardized differences across indicators. Values represent standardized (z‐score) means of the indicators reported in Table 1, calculated relative to the full sample mean (0). Positive deviations indicate worse status or greater burden, while negative deviations indicate better status. As shown, class 1 consistently falls below the sample mean across domains, class 2 shows elevated psychiatric symptoms and disability but average substance use, and class 3 exhibits elevated burden for substance use, psychiatric indicators and social disadvantage. Abbreviations: OUD = opioid use disorder; PCL‐5 = PTSD checklist for Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM‐5); PEG = Pain, Enjoyment of Life and General Activities scale; PHQ‐9 = Patient Health Questionnaire 9; VR‐12 = Veterans RAND 12‐Item Health Survey. Baseline characteristics by class Baseline socio‐demographic and clinical characteristics differed meaningfully across the three complexity phenotypes (Table 2). Participants in the CC and EUC treatment arms were similarly distributed across classes, consistent with successful randomization. With respect to socio‐demographics, the classes were similar in sex distribution and racial/ethnic composition; most participants in all classes were female and identified as Hispanic. Significant differences were observed for age and marital status. Participants in the mental health complexity class were slightly older on average (42.5 years) than those in the low complexity (39.5 years) or high dual complexity (38.0 years) classes. The mental health complexity class had the highest proportion of individuals who were widowed, divorced or separated (35.0%), while the low complexity class had the largest proportion who had never married (38.6%). Educational attainment was generally lower in the mental health and high dual complexity classes compared with the low complexity class, though these differences were not statistically significant. Current legal trouble was also similar across classes. TABLE 2Socio‐demographic and baseline clinical characteristics by clinical complexity phenotype.Class 1 Low complexity (n = 352, 43%)Class 2 Mental health complexity (n = 263, 34%)Class 3 High dual complexity (n = 182, 23%)BH‐adjusted P Treatment group assignment, n (%)0.536Collaborative care (CC)168 (47.7%)138 (52.5%)94 (51.7%)Enhanced usual care (EUC)184 (52.3%)125 (47.5%)88 (48.4%) Socio‐demographic characteristics Age in years, mean (SD)39.5 (12.1)42.5 (11.7)38.0 (11.0) <0.001 Sex, n (%)0.655Male157 (44.6%)118 (44.9%)89 (48.9%)Female195 (55.4%)145 (55.1%)93 (51.1%)Race and ethnicity, n (%)0.083White, non‐Hispanic84 (23.9%)55 (20.9%)48 (26.4%)Hispanic, all races240 (68.2%)192 (73.0%)111 (61.0%)Other/more than one race, non‐Hispanic28 (8.0%)16 (6.1%)23 (12.6%)Education, n (%)0.232Less than high school94 (26.7%)87 (33.1%)62 (34.1%)High school or equivalent99 (28.1%)80 (30.4%)50 (27.5%)Some college or more159 (45.2%)96 (36.5%)70 (38.5%)Marital status, n (%) 0.014 Never married136 (38.6%)80 (30.4%)72 (39.6%)Married/living with partner137 (38.9%)91 (34.6%)66 (36.3%)Widowed/divorced/separated79 (22.4%)92 (35.0%)44 (24.2%)Any current legal trouble, n (%)67 (19.0%)44 (16.7%)43 (23.6%)0.243 Clinical characteristics Any opioid overdose events (past 3 months), n (%)9 (2.6%)10 (3.8%)21 (11.5%) <0.001 AUDIT‐C (past 3 months), mean (SD)3.2 (6.2)5.4 (9.4)3.6 (6.6) 0.003 MOUD status (past 30 days), n (%) <0.001 No MOUD42 (11.9%)45 (17.1%)65 (35.7%)On methadone as prescribed58 (16.5%)41 (15.6%)48 (26.4%)On buprenorphine as prescribed241 (68.5%)158 (60.1%)43 (23.6%)On MOUD, never/sometimes prescribed11 (3.1%)19 (7.2%)26 (14.3%)Any history of trauma, n (%)0.927No trauma31 (8.8%)21 (8.0%)12 (6.6%)Trauma, no interpersonal violence191 (54.3%)141 (53.6%)100 (55.0%)Trauma with interpersonal violence130 (36.9%)101 (38.4%)70 (38.5%)VR‐12 mental health component score, mean (SD)39.3 (11.3)27.8 (10.8)29.9 (13.0) <0.001 PEG score, mean (SD)11.3 (8.0)21.2 (6.6)16.8 (8.6) <0.001 Health system, n (%) 0.005 Health system 1158 (44.9%)143 (54.4%)80 (44.0%)Health system 216 (4.6%)13 (4.9%)6 (3.3%)Health system 3158 (44.9%)92 (35.0%)71 (39.0%)Health system 420 (5.7%)15 (5.7%)25 (13.7%) Note: Class sizes and descriptive characteristics are based on modal class assignment. Entries in bold indicate significance after BH correction (q ≤ 0.05). BH‐adjusted P‐values are for omnibus comparisons across classes.Abbreviations: AUDIT = Alcohol Use Disorders Identification Test; BH = Benjamini–Hochberg; MOUD = medication for opioid use disorder; PCL‐5 = PTSD checklist for Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM‐5); PEG = Pain, Enjoyment of Life and General Activities scale; PHQ‐9 = Patient Health Questionnaire 9; VR‐12 = Veterans RAND 12‐Item Health Survey. Socio‐demographic and baseline clinical characteristics by clinical complexity phenotype. Note: Class sizes and descriptive characteristics are based on modal class assignment. Entries in bold indicate significance after BH correction (q ≤ 0.05). BH‐adjusted P‐values are for omnibus comparisons across classes. Abbreviations: AUDIT = Alcohol Use Disorders Identification Test; BH = Benjamini–Hochberg; MOUD = medication for opioid use disorder; PCL‐5 = PTSD checklist for Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM‐5); PEG = Pain, Enjoyment of Life and General Activities scale; PHQ‐9 = Patient Health Questionnaire 9; VR‐12 = Veterans RAND 12‐Item Health Survey. Clinical characteristics showed sharper contrasts. Statistically significant differences across classes were observed for recent opioid overdose, current MOUD status, AUDIT‐C scores, VR‐12 mental health component scores and health system distribution. The high dual complexity class demonstrated the greatest substance use severity, with 11.5% reporting a recent opioid overdose and with the lowest proportion taking prescribed buprenorphine at baseline (23.6%). In contrast, nearly 70% of the low complexity class entered the study on prescribed buprenorphine. The mental health complexity class reported the poorest mental health functioning, with the lowest VR‐12 mental health scores (mean = 27.8) and the greatest alcohol use severity, and with the highest AUDIT‐C scores (mean = 5.4), relative to other classes. Trauma exposure, particularly involving interpersonal violence, was common across all groups, affecting over one‐third of participants, but the differences were not statistically significant. Treatment effects by class At baseline, no significant differences in PHQ‐9, PCL‐5, or MOUD status were observed across treatment arms within each class (see Table S3). However, at 6 months, CC was associated with lower depression severity compared with EUC in the low complexity class (mean PHQ‐9 score = 7.8 vs 9.7; Δ = −1.88; P = 0.029) (Table 3). Differences between CC and EUC in the mental health complexity class (13.6 vs 14.7; Δ = −1.19; P = 0.319) and in the high dual complexity class (12.5 vs 12.3; Δ = 0.27; P = 0.867) were smaller and were not statistically significant. TABLE 3Depression and PTSD outcomes by clinical complexity phenotype and treatment arm.Mental health outcomesTreatment groupClass 1 Low complexityClass 2 Mental health complexityClass 3 High dual complexityPHQ‐9 (Depression)CC7.8113.5512.52EUC9.6914.7412.26 Contrast (95% CI) P −1.88 (−3.58 to −0.19) 0.029 −1.19 (−3.52 to 1.15) 0.319 0.27 (−2.85 to 3.38) 0.867 PCL‐5 (PTSD)CC22.0941.4334.31EUC24.4841.0934.31 Contrast (95% CI) P −2.39 (−7.20 to 2.42) 0.330 0.34 (−5.56 to 6.24) 0.910 0.00 (−8.61 to 8.62) 1.000 Note: Treatment effect estimates were obtained using a one‐step latent class model that jointly estimates class membership and outcome models, integrating classification uncertainty probabilistically. Total analytic samples were n = 405 for PHQ‐9 and n = 368 for PCL‐5. Contrasts are estimated mean differences between CC and EUC with 95% confidence intervals (95% CIs) and P‐values. Entries in bold denote contrast estimates that are significant at the 0.05 level.Abbreviations: CC = collaborative care; EUC = enhanced usual care; PCL‐5 = PTSD Checklist for Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM‐5); PHQ‐9 = Patient Health Questionnaire 9; PTSD = posttraumatic stress disorder. Depression and PTSD outcomes by clinical complexity phenotype and treatment arm. Contrast (95% CI) P −1.88 (−3.58 to −0.19) 0.029 −1.19 (−3.52 to 1.15) 0.319 0.27 (−2.85 to 3.38) 0.867 Contrast (95% CI) P −2.39 (−7.20 to 2.42) 0.330 0.34 (−5.56 to 6.24) 0.910 0.00 (−8.61 to 8.62) 1.000 Note: Treatment effect estimates were obtained using a one‐step latent class model that jointly estimates class membership and outcome models, integrating classification uncertainty probabilistically. Total analytic samples were n = 405 for PHQ‐9 and n = 368 for PCL‐5. Contrasts are estimated mean differences between CC and EUC with 95% confidence intervals (95% CIs) and P‐values. Entries in bold denote contrast estimates that are significant at the 0.05 level. Abbreviations: CC = collaborative care; EUC = enhanced usual care; PCL‐5 = PTSD Checklist for Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM‐5); PHQ‐9 = Patient Health Questionnaire 9; PTSD = posttraumatic stress disorder. For PTSD symptoms, no significant between‐arm differences were observed in any class. Mean PCL‐5 scores for CC versus EUC were 22.1 versus 24.5 in the low complexity class (Δ = −2.39; P = 0.330), 41.4 versus 41.1 in the mental health complexity class (Δ = 0.34; P = 0.910), and 34.3 versus 34.3 in the high dual complexity class (Δ = 0.00; P = 1.000). Buprenorphine outcomes were broadly similar across treatment arms within each clinical complexity class (Table 4). No significant differences were observed between CC and EUC in any class for the proportion of participants with any buprenorphine use during follow‐up, total number of days prescribed buprenorphine or percentage of study days that a participant had a buprenorphine prescription (‘coverage’). Across all classes, effect estimates were small and confidence intervals spanned the null. TABLE 4Buprenorphine outcomes by clinical complexity phenotype and treatment arm.Buprenorphine outcomesTreatment groupClass 1 Low complexityClass 2 Mental health complexityClass 3 High dual complexityAny utilizationCC74.6%76.6%61.1%EUC78.3%78.5%70.2% Contrast (95% CI) P −0.03 (−0.57 to 0.51) 0.905 −0.02 (−0.65 to 0.61) 0.955 −0.23 (−0.86 to 0.41) 0.484 Days prescribedCC121.389.823.6EUC119.682.627.2 Contrast (95% CI) P 1.77 (−15.58 to 19.12) 0.841 7.19 (−12.95 to 27.33) 0.484 −3.58 (−27.81 to 20.65) 0.772 % coverageCC67.5%51.1%16.8%EUC66.8%45.8%17.6% Contrast (95% CI) P 0.01 (−0.09 to 0.10) 0.886 0.05 (−0.06 to 0.17) 0.354 −0.01 (−0.14 to 0.13) 0.908 Note: Treatment effect estimates were obtained using a one‐step latent class model that jointly estimates class membership and outcome models, integrating classification uncertainty probabilistically. Any buprenorphine utilization was examined among all randomized participants (n = 797). Days prescribed and percentage coverage were examined among participants with any buprenorphine utilization during follow‐up (n = 604). Contrasts are estimated mean differences between collaborative care (CC) and enhanced usual care (EUC) with 95% confidence intervals (95% CIs) and P‐values. Buprenorphine outcomes by clinical complexity phenotype and treatment arm. Contrast (95% CI) P −0.03 (−0.57 to 0.51) 0.905 −0.02 (−0.65 to 0.61) 0.955 −0.23 (−0.86 to 0.41) 0.484 Contrast (95% CI) P 1.77 (−15.58 to 19.12) 0.841 7.19 (−12.95 to 27.33) 0.484 −3.58 (−27.81 to 20.65) 0.772 Contrast (95% CI) P 0.01 (−0.09 to 0.10) 0.886 0.05 (−0.06 to 0.17) 0.354 −0.01 (−0.14 to 0.13) 0.908 Note: Treatment effect estimates were obtained using a one‐step latent class model that jointly estimates class membership and outcome models, integrating classification uncertainty probabilistically. Any buprenorphine utilization was examined among all randomized participants (n = 797). Days prescribed and percentage coverage were examined among participants with any buprenorphine utilization during follow‐up (n = 604). Contrasts are estimated mean differences between collaborative care (CC) and enhanced usual care (EUC) with 95% confidence intervals (95% CIs) and P‐values. DISCUSSION This study demonstrates substantial heterogeneity in baseline clinical and social complexity among participants in CLARO, a community‐based randomized trial of CC for individuals with co‐occurring OUD and mental health conditions. In the primary CLARO analysis, no significant differences were observed between CC and EUC at 6 months across primary outcomes [26]. Using LCA, we identified three distinct subgroups of patients who differed meaningfully in their clinical, functional and social profiles. Nearly half of participants were classified as ‘low complexity’, while the remainder were divided between the ‘mental health complexity’ class, with severe psychiatric symptoms but limited substance use, and the ‘high dual complexity’ class, with high substance use and psychiatric symptoms coupled with social challenges. These findings highlight that clinical complexity is heterogeneous within OUD populations and that latent class methods can identify clinically interpretable baseline risk profiles associated with variation in treatment response. The heterogeneity in clinical complexity that we observed is consistent with prior research documenting facets of clinical comorbidity among individuals with OUD. A recent systematic review found that over one‐third of individuals with OUD in the USA meet criteria for depression, nearly one‐fifth meet criteria for PTSD, and nearly two‐thirds have another non‐opioid drug use disorder [48]. Prior studies have found that these co‐occurring conditions are associated with greater risk of suicide, overdose and functional impairment compared with OUD alone [15, 16, 17, 49, 50, 51]. Physical health problems and chronic pain are also prevalent, along with social adversity such as housing instability, trauma exposure and legal involvement [14]. Such complexity may attenuate the average treatment effects when outcomes are evaluated across the full sample and may shape the degree to which patients are able to engage with, and benefit from, structured interventions. Among participants in the low complexity class, CC was associated with modest improvements in depression severity relative to EUC. This finding is consistent with extensive evidence demonstrating the effectiveness of CC for depression management in primary care settings [52, 53]. Our findings add to the prior literature by highlighting that the relative advantage of CC may be concentrated among patients with fewer barriers to engagement, such as stable housing and social support, and those with less severe depression [54]. However, CC did not significantly improve depression symptoms above and beyond EUC for participants in the mental health complexity or high dual complexity classes. Participants in the low complexity class may have been more able to engage consistently with psychotherapy, care management visits and follow‐up, thereby receiving a higher effective dose of CC. In contrast, individuals with greater psychiatric burden, active substance use or social instability may have faced competing demands that limited sustained participation and/or made their depression more difficult to treat. For these patients, standard CC workflows may not be sufficient to achieve symptom improvement within 6 months. Collaborative care (CC) did not demonstrate differential effects on PTSD symptoms across any complexity class, in contrast to the finding for depression in the low complexity class. This pattern indicates that the modest benefit for depression observed among patients with lower complexity did not extend to PTSD. This may be in part because of the lower effectiveness of medications for the treatment of PTSD. Qualitative findings from the trial also indicated that some providers expressed discomfort in delivering trauma‐focused psychotherapy and concerns about exacerbating symptoms, and that the uptake of PTSD‐focused treatment was variable. These implementation and engagement challenges may have reduced effective exposure to trauma‐focused care and attenuated potential intervention effects for PTSD. Collaborative care also did not improve buprenorphine‐related outcomes above and beyond EUC for any clinical complexity class. This is consistent with two previous studies of CC for drug use disorders; a third study showed small positive effects in a population with lower OUD severity at baseline [55, 56, 57, 58]. The absence of buprenorphine benefits across all subgroups suggests that CC, as implemented here, did not confer additional advantage over EUC for improving OUD‐related outcomes in primary care. It remains an open question whether adaptations to the CC model (e.g. more intensive case management, peer recovery support or direct prescribing authority for care managers) could improve OUD outcomes. Additionally, because randomization occurred at the individual level within shared clinic sites, potential spillover across treatment arms may have contributed to the null buprenorphine findings. As providers cared for participants in both arms simultaneously, access to psychiatric case review and consultation provided as part of the CC model may have enhanced MOUD‐related clinical practice more broadly within participating clinics, attenuating observable differences between arms. These findings suggest several directions for future research. The present study focused on depression, PTSD and buprenorphine outcomes, but CC may have effects on broader domains of patient wellbeing that were not captured here. Pain severity, housing instability, legal involvement and social functioning all contributed to the definition of the latent classes at baseline, and may themselves represent meaningful targets of CC. Future work incorporating a wider range of outcomes, designed prospectively to examine treatment effect heterogeneity across multi‐dimensional complexity profiles, could more comprehensively characterize for whom CC is most effective and through which mechanisms. Second, given the absence of differential effects on buprenorphine outcomes, future studies should determine whether adaptations to CC can meaningfully improve OUD care or whether resources are better directed toward strengthening either EUC in primary care or pathways to specialty addiction treatment. Limitations This study has several limitations. First, the primary trial outcomes focused on substance use and mental health symptoms and did not capture broader domains of patient wellbeing. Clinical complexity in OUD is multifaceted and includes social determinants such as housing instability, legal involvement and limited social support. Collaborative care may have influenced functioning in these domains, but such effects were not directly assessed. Future studies should incorporate outcome measures of social functioning and recovery to more comprehensively evaluate intervention impact. Second, this secondary analysis was not pre‐specified in the original trial protocol, and the CLARO trial was not designed or powered to detect treatment effect heterogeneity across latent classes. Subgroup‐specific findings should therefore be interpreted as exploratory. The observed improvement in depression in the low complexity class was modest and below the minimum detectable effect size specified in the original trial power calculation, warranting cautious interpretation. Third, substance use measures were based on self‐report without objective toxicology confirmation, and other drug use beyond opioids, alcohol and stimulants was not comprehensively characterized, which may limit precision in assessing baseline complexity and treatment response. Fourth, findings may not generalize to OUD populations in other settings, as the CLARO trial enrolled patients from low‐resource primary care clinics in New Mexico and California. Finally, the 6‐month follow‐up period may not have been sufficient to capture the longer‐term effects of CC, particularly for patients with high complexity whose needs may require sustained intervention. CONCLUSION Overall, this study characterizes the heterogeneity of clinical complexity among patients with OUD and co‐occurring mental health disorders and suggests that baseline complexity may be associated with variation in response to CC. Latent class analysis identified three distinct profiles characterized by varying levels of psychiatric burden, substance use, physical health and social determinants. Across all groups, CC was not more effective than EUC for OUD outcomes, indicating that the model as implemented did not improve buprenorphine utilization, prescribing duration or coverage. CC was associated with modest relative improvements in depression only among patients with the lowest complexity. Taken together, these findings suggest that CC, as implemented in CLARO, did not provide meaningful additional benefit over EUC for those with greater clinical and social complexity. Future work should consider whether adaptations to the model or alternative care strategies are needed to improve outcomes for patients with OUD and co‐occurring mental health conditions. The complexity profiles identified here, empirically derived from routinely collected clinical and social measures, may offer a pragmatic framework for stratifying patients and evaluating differential intervention strategies in future research. AUTHOR CONTRIBUTIONS Megan S. Schuler: Conceptualization; formal analysis; writing—original draft; methodology. Rebecca Weir: Methodology; formal analysis; data curation; writing—review and editing. Colleen M. McCullough: Methodology; writing—review and editing; formal analysis; data curation. Grace M. Hindmarch: Writing—review and editing; project administration. Beth Ann Griffin: Writing—review and editing. Kirsten Becker: Writing—review and editing; project administration. Lisa S. Meredith: Writing—review and editing. Katherine E. Watkins: Writing—review and editing; funding acquisition; conceptualization; supervision. DECLARATION OF INTERESTS The authors have no conflicts of interest to declare. CLINICAL TRIAL REGISTRATION ClinicalTrials.gov: NCT04559893 and NCT04634279. Supporting information Table S1. Fit statistics for latent class models. Table S2. Clinical complexity phenotypes at baseline: four‐class latent class model. Table S3. Comparison of mental health outcomes and medication for opioid use disorder (MOUD) status at baseline, by latent class analysis (LCA) subgroup and treatment group ACKNOWLEDGEMENTS The authors would like to acknowledge our community health system and patient partners, without whom this work would not have been possible. This includes the CLARO care managers as well as clinical champions: John Bettler (University of New Mexico Department of Family & Community Medicine), Seth Williams (University of New Mexico Family Health Clinic, Southeast Heights), Kyle Ragins (Los Angeles County Department of Health Services), Hrishikesh Belani (Los Angeles County Department of Health Services) and Scott Jeansonne (First Choice Community Healthcare). We would also like to thank Virginia Chitwood‐Sedore (University of New Mexico) for community health worker supervision, Denise Sloan (Boston University Chobanian and Avedisian School of Medicine) and Patric Raue (University of Washington) for the training and supervision of therapists for evidence‐based psychotherapies, and Laura Brown (University of New Mexico) for assistance with implementation. We further thank Vanessa Miller and Keisha McDonald from RAND's Survey Research Group for their assistance with data collection and Daniel Wang (RAND) for assistance with data preparation; as well as Jess Anderson, Mary Carmody and Lina Tarhuni (all from University of New Mexico), and Cristina Murray‐Krezan (University of Pittsburgh), for assistance with data collection and preparation. Finally, from RAND, we thank: Maya Rabinowitz and Isabel Leamon for administrative and research support; Daniel Wang for programming and quality assurance support; Liisa Ecola for administrative support and project management; Graciela Arzac for assistance with manuscript preparation; and Tiffany Hruby for project administrative support. Schuler MS, Weir R, McCullough CM, Hindmarch GM, Griffin BA, Becker K, et al. A latent class analysis of clinical complexity: Secondary analysis of the collaboration leading to addiction treatment and recovery from other stresses (CLARO) randomized trial. Addiction. 2026;121(10):2833–2844. 10.1111/add.70503 Funding information This study was funded by the National Institute of Mental Health (grant U01MH121954 to K.E.W. and Miriam Komaromy, with multiple principal investigators). The National Institute of Mental Health had no role in the: design and conduct of the study; collection, management, analysis and interpretation of the data; preparation, review or approval of the article; or decision to submit the article for publication. DATA AVAILABILITY STATEMENT Deidentified participant data will be available beginning April 1, 2026 through the NIMH HEAL repository (https://heal.nih.gov/data. REFERENCES Associated Data Supplementary Materials Table S1. Fit statistics for latent class models. Table S2. Clinical complexity phenotypes at baseline: four‐class latent class model. Table S3. Comparison of mental health outcomes and medication for opioid use disorder (MOUD) status at baseline, by latent class analysis (LCA) subgroup and treatment group Data Availability Statement Deidentified participant data will be available beginning April 1, 2026 through the NIMH HEAL repository (https://heal.nih.gov/data.