Trajectories of Treatment Response and Non-Response in Youth at High Risk for Suicide RH = Treatment Response-Youth with Self Harm

Abstract

Objective:

To examine trajectories of treatment response in suicidal youth who participated in a randomized controlled trial (RCT) comparing dialectical behavior therapy (DBT) and individual and group supportive therapy (IGST).

Method:

Using latent class analysis across both treatment conditions, we conducted secondary analyses of data from a multisite RCT with N = 173 youth, ages 12–18, with repetitive self-harm (SH, including ≥ 1 lifetime suicide attempt) and elevated suicidal ideation (SI). The sample was 95% female, 56.4% White and 27.49% Latina. Participants received 6 months of DBT or IGST and 6 months of follow-up. Primary outcomes were SH and SI.

Results:

63% of the sample were members of latent classes that showed improvement in SI and 74% showed improvement in SH. 13% were “total non-responders,” with no improvement in SI or SH. SH non-response emerged at the mid-point of treatment (3 months), with non-responders showing a sharp increase in SH over the remainder of treatment and follow-up. Youth in DBT were significantly more likely to be an SH responder v. non-responder than those in IGST (χ² [1] =6.53, p=0.01). An optimal threshold cut point using multivariate predictors of total non-response (White, externalizing symptoms, total SH, and SI) predicted total non-responders to DBT with 100% accuracy.

Conclusion:

This is the first study to identify trajectories of both SI and SH response to treatment in a sample of adolescents at risk of suicide. Results may inform personalized treatment approaches.

Introduction

Suicide is the second leading cause of death among 10–24 year-olds in the United States.^1,2 Both suicide attempts (SA) and non-suicidal self-injury (NSSI) are risk factors for death by suicide.^3,4 Effective treatments for reducing SA and NSSI (jointly labeled self-harm [SH])⁵ in adolescence are urgently needed. Dialectical behavior therapy (DBT) is the only treatment that has shown reductions in suicidal ideation (SI) and SH that has been replicated in two separate RCTs.^6,7 However, there is considerable room for improvement in DBT outcomes. In the largest RCT, the Collaborative Adolescent Research on Emotions and Suicide (CARES) Study,⁶ which included youth with the most severe history of SH, 54% of youth in the DBT condition had at least one SH episode during treatment and 50% had at least one SH between treatment completion and 6-month post-treatment follow-up.⁶ At post-treatment, youth who received DBT had significant reductions in SA, SH, NSSI, and SI, as compared to an individual and group supportive therapy (IGST) control condition. However, at 6-month follow-up, the only significant difference between DBT and IGST was SH remission, a post-hoc variable indexing the absence of any SH during the follow-up interval.^7,8

Understanding individual differences in trajectories of treatment response and non-response is needed to personalize treatments and match adolescents at risk to the treatments most likely to be effective for them, as well as to increase scalability. Two prior studies have examined adolescents’ trajectories of response to suicide-focused treatments for SI. In an RCT comparing attachment-based family therapy (ABFT) and family-enhanced non-directive dupportive therapy (FE-NST), growth mixture modeling yielded three classes of treatment response, defined in terms of improvements in SI and depressive symptoms over 16 weeks of treatment: responders (57.5%), partial responders (26.7%), and non-responders (15.8%). Youth with the lowest baseline SI and depression showed the most improvement and history of NSSI at baseline was a predictor of non-response.⁹ In an inpatient sample who received either the youth nominated support team II (YST-II) intervention plus treatment-as-usual (TAU) or TAU alone, latent class growth modeling identified three trajectories over 12-months: one with subclinical SI at baseline SI (31.6%) that continued to decline, one with high baseline SI that sharply decreased by 3-month follow-up (57.4%), and one with high baseline SI that was persistent over time (10.9%)¹⁰. Membership in the persistent SI group was predictive of SA at 12-month follow-up.

The present study used latent class analyses to examine longitudinal trajectories of response and non-response to two suicide-focused psychotherapies (DBT and IGST) in youth who took part in the CARES Study. This is the first study that we are aware of to examine trajectories of treatment response using both SI and SH as outcomes. The CARES sample consisted of youth with at least one lifetime SA, repetitive SH within the past 3 months and elevated SI within the past month. Hence, we were able to examine if there are unique subtypes of treatment response and non-response in both SI and SH, in a sample of youth at high risk for suicide, across multiple outcome assessments (3, 6, 9 and 12-months from baseline). Aims of these secondary analyses of the CARES Study data were to identify: 1) latent classes of treatment response and non-response over multiple time points for SI, SH and SI + SH combined, both across the entire sample and by treatment condition; 2) univariate and multivariate predictors of class membership; and 3) an optimal threshold cut-point that predicts if a given youth is likely to be a treatment responder or non-responder, that can be used to personalize treatment approach for optimal reduction of suicide risk.

Method

For a detailed description of the study design, see McCauley et al. (2018) and other prior publications from the CARES Study.^6,11,12 Here, we describe only the measures and procedures relevant to this report. The study was reviewed by each site’s local IRB. All youth and parents gave informed assent/consent to participate in the study.

Participants

Participants were recruited from January 2012 to August 2014 at four sites: University of Washington; Seattle Children’s Hospital; Harbor-UCLA Medical Center; and UCLA Medical Center. The sample included N = 173 adolescents, ages 12–18, who were randomized to 6 months of either DBT or IGST, with assessments conducted at baseline (pre-treatment) and 3 (mid-treatment), 6 (end of treatment), 9 and 12-month follow-ups. Inclusion criteria were: ≥ 3 lifetime SH episodes, one within 12-weeks before screening; ≥ 1 lifetime SA; ≥ 24 on the Suicidal Ideation Questionnaire Junior (SIQ-Jr)¹³, and ≥ 3 borderline personality disorder (BPD) symptoms. Exclusion criteria were: IQ less than 70; court-ordered to treatment; primary psychosis, mania, anorexia, or other life-threatening condition; and youth not fluent in English or parent not fluent in English or Spanish. Mean age was 14.89 years (SD=1.47). The sample was 95% female, 56.4% White and 27.49% Latina (all Latina participants were female). Sixty-five percent of families reported an annual income of $50,000 or higher. Participants met criteria for the following DSM IV TR diagnoses: 83.81% depressive disorders, 54.1% anxiety disorder, 53% BPD. For a detailed description of sample demographics, as published previously,^6,11 see Table S1, available online.

Given that the study inclusion criteria selected for youth at increased risk of SH, both treatment conditions had detailed safety protocols, including the use of emergency services, hospitalization and removal from the study and referrals to other types of treatment if needed. The study was closely monitored by a Data Safety and Monitoring Board and a study ombudsman was designated at each site who was available to independently evaluate if a subject needed to be removed from the study protocol. For a detailed description of study safety protocols see Berk et al., 2014¹⁴.

Randomization

Adaptive randomization balanced participants across conditions within sites based on age, number of prior SA, number of prior SH and psychotropic medication use. Both DBT and IGST were manualized, used similar training and adherence protocols, and offered equivalent treatment components (e.g., both individual and group therapy) to maximize internal validity. For a detailed description of the treatment approaches see Berk et al., (2014)¹⁴ and McCauley et al. (2018)⁶.

Procedures

Assessments were standardized self-report questionnaires and structured interviews with youth and parents and were conducted by independent evaluators who were blind to condition.

Measures

Primary outcome measures were: 1) number of SH episodes using the Suicide Attempt and Self Injury Interview (SASII; inter-rater reliability was maintained throughout the study at ICC ≥.80 at the item level^8,5,15 and 2) past-month SI on the Suicidal Ideation Questionnaire-Junior (SIQ-Jr., cut-off score for clinical concern ≥ 31).¹³ We focused on global self-harm (SH) as the primary behavioral outcome measure of treatment response, as consistent with prior CARES Study publications examining mediators and moderators of treatment outcome^11,12 and recent meta-analyses of existing therapeutic interventions for suicidal youth.^16,17

Measures of predictors of class membership were: demographics (race, ethnicity, age, family income and presence of youth medical conditions) assessed by parent report; Axis I diagnoses (Schedule for Affective Disorders and Schizophrenia for School-Aged Children [KSADS]);¹⁸ borderline personality disorder (BPD) diagnosis (Structured Clinical Interview for the DSM-IV, Axis II [SCID-II]);¹⁹ depressive symptoms (Center for Epidemiologic Studies Depression Scale Revised [CESD-R]);²⁰ internalizing symptoms, externalizing symptoms, total problem score, hypersomnia, nightmares, insomnia and total sleep problems (Child Behavior Checklist [CBCL] and Youth Self-Report [YSR]);²¹ impairment related to substance use (Drug Use Screening Inventory-Revised [DUSI-R] problem density score);²² PTSD symptoms (UCLA PTSD Reaction Index for DSM IV);²³ parent psychopathology (General Severity Index, Positive Symptom total and Positive Symptom Distress Index from the Brief Symptom Inventory [BSI]);^24,25 emotion dysregulation (Difficulties in Emotion Regulation Scale [DERS]);²⁶ family conflict (Conflict Behavior Questionnaire-20 [CBQ-20]);²⁷ and school, family and peer functioning (Social Adjustment Scale Self-Report [SAS-SRA]).²⁸

Statistical Analyses

Latent class analysis (LCA) was used to describe distinct patterns of treatment response and non-response for SI and SH across the entire sample over the 6-month active treatment phase and 6-month follow-up period. Due to positive skew in SH frequency, in order to reduce spurious results due to outliers, we limited extreme values using winsorization at the upper 2.5 percentile (80 reported SH episodes acts in the 3 month period).²⁹ All LCAs were intent-to-treat (ITT). The latent class model estimates the conditional probability of SI and SH at each time point and the posterior probabilities of membership in each class for each subject. Subjects are assigned to the class with the highest posterior probability of membership. We determined the number of latent classes per observation period using the Bayesian information criterion (BIC), entropy and the rule that each class must exceed 5% of the effective sample size.^30–32 Next, we descriptively characterized each class in terms of their pattern of improvement or worsening. All analyses were conducted in SAS 9.4, with the latent class modeling performed with the add-on LCA procedure developed at Penn State University (PROC LCA³³). Differences in latent class membership by treatment arm were examined using Chi-square tests of independence. Chi-square tests were replaced by Fisher’s exact tests in the presence of small cells. We combined the observed SI and SH latent classes into two broader groups of “responders” versus “non-responders” for each outcome and into a group of “total non-responders” which included youth with lack of improvement or worsening for SI + SH. Although we used LCA, there are alternative analytic approaches. Another option is latent growth curve modeling (LGCM); however, we decided against this approach because the limited number of assessments in this study could limit the model convergence of the growth curve model.

Predictors of Class Membership

We used baseline predictors to predict membership in responder versus non-responder groups for SI, SH, and SI + SH combined. All predictors were included as univariate predictors. Multivariate analyses were based on the identification of significant univariate predictors. For each continuous covariate, we pre-specified the use of polynomial relationships, to avoid assumptions of linear associations while achieving interpretable results. Model development began with the inclusion of all significant univariate predictors in a logistic regression model with baseline SI and SH included as covariates. We additionally considered interaction terms and then eliminated potential interactions and main effects in the full model based on likelihood ratio tests of nested models to yield a parsimonious model which still yielded significant discrimination.

Optimal Threshold Cut-Point for Determining Response versus Non-Response to Treatment

An optimal threshold cut-point can be determined from a diagnostic model using the receiver operating characteristic (ROC),^34–40 which illustrates the discrimination over all possible observed combinations of the components of the multivariate prediction model. This optimal threshold, the optimal operating point (OOP), yields an optimal cut-off score from our multivariate prediction model for predicting an individual patient’s likely response or non-response to treatment. The OOP provides an algebraic and geometrical tool for the classification of the predicted outcome at the corresponding sensitivity and specificity levels of the OOP. Based on clinical significance, we focused the optimal threshold analyses on prediction of total non-responders, both for the sample as a whole and by treatment condition.

Results

Latent Class Analyses

Suicidal ideation.

As shown in Figure 1, for SI, LCA yielded five classes falling within two subgroups: 1) SI responders (63.1% of sample), composed of three classes showing improvement over the 12-month observation period and 2) SI non-responders (36.9% of sample), composed of two classes characterized by lack of improvement. As shown in Figure 3, SI responders showed steady decreases in baseline SI over time, with an average reduction of 32.05 points on the SIQ-Jr. during active treatment and an average reduction of an addition 7.08 points during the follow-up phase of the study. In contrast, SI non-responders had higher baseline SIQ Jr. scores than SI responders that decreased slightly during the treatment phase by an average of 8.74 points and increased an average of 1.45 points during the follow-up period. Scores remained well over the SIQ-Jr. clinical cut-off score of ≥ 31 over the course of treatment and 12-month follow-up for non-responders, while scores for responders generally declined below the clinical range at post-treatment and remained low through follow-up.

Figure 1: Five Latent Classes for Suicidal Ideation.

Note: Each graph shows the individual trajectories for each youth within the identified latent classes along with the smoothing spline average trajectory within each class, where the smoothing spline represents the average continuous trajectory over time using the individual trajectories rather than the descriptives of the available data at each point. SIQ = Suicidal Ideation Questionnaire.

Self-harm.

As shown in Figure 2, the LCA for SH also yielded five classes; three SH responder classes that showed decreases in SH episodes over time (74% of sample) and two SH non-responder classes that showed increases in the number of SH episodes over time (26% of sample). As shown in Figure 3, from baseline through mid-treatment (3-months), both SH responders and non-responders had a similar average reduction in the mean number of SH episodes (20.11 for SH responders; 22.89 for SH non-responders); however, from 3 to 12-months/final follow-up, responders continued to improve with an on-average reduction of 3.1, whereas non-responders showed a steady increase in SH by an average of 61.3 episodes.

Figure 2: Five Latent Classes for Self-Harm.

Note: Each graph shows the individual trajectories for each youth within the identified latent classes along with the smoothing spline average trajectory within each class, where the smoothing spline represents the average continuous trajectory over time using the individual trajectories rather than the descriptives of the available data at each point.

Total non-responders.

To capture variation using a combined risk index, we created a group that combined non-responders on both SI and SH (13% of the sample) that we labeled total non-responders. The remaining group (87%) consisted of youth with any positive response (i.e., positive response for SI alone, SH alone, and/or both SI and SH). Average trajectories for total non-responders closely mirrored the patterns seen for SI non-responders and SH non-responders examined separately. As shown in Figure 3, for SI, during the active treatment, all participants showed a decrease in SIQ-Jr. scores, with a much smaller decrease for total non-responders (average of 4.7 points) than any-positive responders (average of 24.6 points). However, from the end of treatment (6 months) through the end of the 12- month follow-up, any-positive responders continued to show a small average improvement of 4.2; whereas total non-responders showed an average increase of 16.7. Similarly, as shown in Figure 3, for SH, both total non-responders and any-positive responders showed a decrease in the number of SH episodes in the first 3 months of treatment (25.1 episodes for non-responders and 20.5 episodes for responders). However, total non-responders and any-positive responders showed very different trajectories from 3 months through the end of the 12- month follow-up period, with any-positive responders continuing to show average reductions in SH episodes of 3.0 and total non-responders showing a sharp increase in mean number of SH episodes of an average of 51.8. The proportion of youth that showed trajectories of improvement in both SI and SH (i.e., total responders) was 50.3% (87/173).

Figure 3:

Trajectories of Response and Non-Response for Suicidal Ideation (SI), Self-Harm (SH) and SI + SH Groupings

Class Membership Differences by Treatment Condition

No significant differences between the two treatment arms (DBT versus IGST) were observed for the five SI latent classes (χ² [4] =3.50, p=0.48). For the five SH latent classes, there was a significant difference between treatments (χ² [4] =11.42, p=0.022). In the DBT condition, 52.3% of the subjects were in the largest of the three decreasing classes (with the lowest SH at baseline), whereas only 36.8% of youth in the IGST condition were in this class (χ² [1] =4.23, p<.04). Similarly, only 3.5% of DBT patients were in the two increasing SH classes versus 18.4% of those in IGST (χ² [1] =9.82, p<.002). When we combined the five class models into SI responders/non-responders and SH responders/non-responders groups, there was no association with treatment condition for SI (χ² [1] =0.79, p=0.38; 66.3% responders in DBT vs 59.8% responders in IGST), but there was a significant association for SH (χ² [1] =6.53, p=0.01; 82.6% responders in DBT vs 65.5% responders in IGST) as well for any-positive responders versus total non-responders (χ² [1] =3.94, p=0.047; 91.9% responders in DBT vs 81.6% responders in IGST). Youth who were total responders (SI + SH) were more likely to have received DBT (57.0%) than IGST (43.7%); however, this difference was only marginally significant (χ² [1] =3.06, p=0.08).

Predictors of Group Membership

Each baseline term was added into a logistic regression for the three non-responder groups: SI non-responders, SH non-responders, and total non-responders. Significant univariate and multivariate predictors are shown for each of the three groups in Table 1. Odds ratios larger than 1 indicate that increases in the predictor correspond to a higher likelihood of non-response. Significant univariate predictors of SI non-response were: White race, youth and parent report of nightmares on the YSR and CBCL, higher parent-reported youth sleep problems on the CBCL, higher number of past SH episodes reported in the 3 months prior to baseline, higher internalizing symptom scores on YSR and CBCL, higher levels of youth reported emotion dysregulation on the DERS and higher SI. Significant univariate predictors of SH non-response were parent-report of youth nightmares on the CBCL, higher youth self-reported depression on the CESD-R, and higher parent-reported youth externalizing symptom scores on the CBCL. Significant univariate predictors of total non-response were White race, higher parent psychopathology on the BSI General Severity index, higher number of past SH episodes at baseline, and higher parent-reported youth externalizing symptom scores on the CBCL.

Table 1.

Univariate and Multivariate Predictors of Non-Responder Group Membership

	SI Non-Responder
	Univariatea		Multivariateb
Predictor	Estimate (se, p-value)	Odds Ratioc (95% LB, 95% CI)	Estimate (se, p-value)	Odds Ratioc (95% LB, 95% CI)
Race: White	1.045 (0.339, p=0.002)	2.843 (1.457, 5.547)	1.149 (0.380,p=0.003)	3.157 (1.500,6.645)
Nightmare-YSR	0.637 (0.229, p=0.006)	1.891 (1.203, 2.971)	0.331 (0.286, p=0.247)	1.201 (0.659,2.186)
YSR – Internalizing	0.047 (0.017, p=0.006)	1.048 (1.014, 1.084)	0.034 (0.026, p=0.181)	1.035 (0.984,1.088)
Sleep Problems – CBCL	0.219 (0.102, p=0.033)	1.244 (1.018, 1.520)	0.260 (0.156, p=0.095)	1.297 (0.955,1.761)
Nightmare – CBCL	0.467 (0.219, p=0.034)	1.599 (1.037, 2.467)	0.059 (0.345, p=0.865)	1.061 (0.539,2.087)
CBCL-Internalizing	0.031 (0.015, p=0.037)	1.031 (1.002, 1.062)	−0.011 (0.021, p=0.597)	0.989 (0.950,1.030)
DERS – Total	0.016 (0.008, p=0.048)	1.016 (1.000, 1.033)	−0.001 (0.011, p=0.971)	1.000 (0.979,1.021)
Baseline Self Harm	0.008 (0.004, p=0.039)	1.008 (1.000, 1.016)	0.029 (0.004, p=0.491)	1.003 (0.995,1.030)
Baseline Ideation	0.035 (0.011,p=0.013)	1.036 (1.014,1.058)	0.0279 (0.012,p=0.018)	1.028 (1.005,1.052)
	SH Non-Responder
Predictor	Estimate (se, p-value)	Odds Ratio (95% LB, 95% CI)	Estimate (se, p-value)	Odds Ratio (95% LB, 95% CI)
CBCL-Externalizing	0.038 (0.016, p=0.021)	1.039 (1.006, 1.073)	0.039 (0.025, p=0.115)	1.039 (0.991,1.091)
CES-D	0.038 (0.018, p=0.033)	1.039 (1.003, 1.076)	0.038 (0.018, p=0.031)	1.039 (1.004,1.076)
Nightmares-CBCL	0.477 (0.239, p=0.048)	1.611 (1.004, 2.583)	0.295 (0.253, p=0.244)	1.344 (0.817, 2.209)
	Total Non-Responder (SI AND SH)
Predictor	Estimate (se, p-value)	Odds Ratio (95% LB, 95% CI)	Estimate (se, p-value)	Odds Ratio (95% LB, 95% CI)
CBCL-Externalizing	0.048 (0.020, p=0.021)	1.049 (1.007, 1.092)	0.106 (0.040, p=0.008)	1.112 (1.028, 1.203)
Parent Global Severity Index	0.797 (0.368, p=0.031)	2.220 (1.075, 4.586)	0.519 (0.425, p=0.221)	1.681 (0.732, 3.863)
Race: White	1.117 (0.534, p=0.038)	3.056 (1.064. 8.775)	1.647 (0.697, p=0.018)	5.192 (1.324, 20.362)
Baseline Self Harm	0.007 (0.004, p=0.049)	1.007 (1.000. 1.015)	0.007 (0.004,p=0.047)	1.007 (1.000, 1.015)
Baseline Ideation	0.042 (0.017, p=0.012)	1.043 (1.009, 1.077)	0.046 (0.018, p=0.012)	1.047 (1.010, 1.085)

Note: CBCL = Child Behavior Checklist; CES-D, Center for Epidemiologic Studies-Depression Scale; DERS = Difficulties in Emotion Regulation Scale; YSR = Youth Self-Report.

^aOnly the significant individual univariate predictors of outcome are shown in this table. All other predictors not shown are non-significant (p>0.05).

^bSignificance of the multivariate models are adjusted for all significant univariate predictors.

^cOdds ratios represent the multiplicative impact on non-responder group membership for a unit increase per predictor.

Significant multivariate predictors of SI non-response were White race and baseline SI. The only significant multivariate predictor of SH non-response was baseline depressive symptoms. Multivariate predictors of total non-response were White race, parent report of youth externalizing symptoms on the CBCL, number of prior SH episodes at baseline and SI at baseline.

Optimal Threshold Analysis of Responder vs. Total Non-Responder Groups

Using multivariate predictors of total non-response, we built a prognostic parsimonious third-degree polynomial model including two-way interaction terms that predicted membership as a total non-responder on SI + SH⁴¹. This analysis yielded the following equation:

$$\text{ F(White, CBCL-Externalizing, Ideation, TotalActs) }=-418.203+{2.706}^{\ast }\text{ White }+{0.056}^{\ast }\text{ Ideation }+{19.07}^{\ast }\text{ CBCL-Externalizing - }{8.067}^{\ast }\text{ZT otacts }+{0.12115}^{\ast }\text{CBCL}-{\text{ Externalizing }}^{\ast }\text{ ZTotal Acts- }{0.292}^{\ast }{\text{CBCL- Externalizing }}^2+{0.0015}^{\ast }{\text{CBCL- Externalizing }}^3$$

Function components are youth race (White versus non-White), parent report of externalizing symptoms on the CBCL, total SH, and SI. These predictions are made with 85% sensitivity and 80% specificity. As shown in Figure 4, by entering the baseline score on each variable in the equation, it can be determined if the adolescent is likely to be a responder for at least one outcome (SI, SH and/or SI + SH) or a total non-responder. We also examined the performance of the equation within each intervention arm. For IGST the model yields 76.9% sensitivity and 84.4% specificity, whereas, for DBT the model yields 100% sensitivity and 89.6% specificity; therefore, using baseline, pre-treatment data, we can correctly predict 3 out of 4 adolescents who fall in the total non-response category for IGST and identify all total non-response for DBT.

Figure 4: Prediction of the Optimal Threshold for the Total Sample.

Note: The black curve represents the optimal operating point (OOP). Subjects are divided into high self-harm (SH) (≥ 25) and low SH (≤ 24) at baseline, with the Child Behavior Checklist (CBCL) Externalizing t-score on the x-axis and the Suicidal Ideation Questionnaire, junior (SIQ-Jr.) score on the y-axis. Adolescents predicted as total non-responders are indicated by the red dots, where the blue dots represent responders. Total non-responder status can be predicted by plotting the ordered pair. If the ordered pair lies on the black curve or in the same region as the red dots, then the patient is predicted to be a total non-responder. For example, an adolescent with a CBCL Externalizing t-score = 60, SIQ-Jr. = 40 and high SH is in the responder region.

Discussion

To our knowledge, this is the first study to examine trajectories of SH and SI + SH response to suicide-focused treatments in a sample of adolescents at high risk of suicide. Results showed heterogeneity in trajectories of treatment response and non-response that varied between SI, SH, and SI + SH combined. Taken together, results indicated that even in a high-risk sample of youth characterized by a history of at least one lifetime suicide attempt and recent repetitive SH, a large proportion of youth showed trajectories of improvement in SI, SH, and SI + SH, with only 13% of the sample showing total non-response for both SI and SH. DBT was associated with significantly greater likelihood of membership in an SH responder class than IGST; however, did not significantly impact class membership for SI. This is consistent with DBT’s focus on prioritizing SH behavior change and underscores the benefits of DBT for youth with repetitive SH. Given that repeat SH may increase the “acquired capability” of SAs and death by suicide, reduction of SH is a critical suicide prevention target.^{4, 42}

Examination of the trajectories of SH non-response indicated that worsening SH during the second half of treatment (3–6 months) may indicate the need to modify the treatment approach. Self-reported depressive symptoms at baseline predicted SH non-response, suggesting that additional intervention for depression may be needed, particularly given that SH often functions to regulate negative mood states.⁴³ Youth were more likely to show improvement in SH in DBT versus IGST, suggesting that a directive, skills-based approach may be more protective against worsening SH than a non-directive, supportive treatment. However, without a no-treatment comparison group, we are unable to rule out the possibility that worsening was due to some aspect of the treatments themselves, or to the experience of participating in treatment and not improving. Additional research examining specific medication algorithms for depression in combination with DBT is also needed, particularly given findings in the adolescent depression literature supporting the superiority of combined medication and psychotherapy treatment^44,45.

Consistent with prior research,^9,10 trajectories of SI response over time were associated with severity of baseline SI. Membership in a non-responder class for SI in this study (36.9%) was higher than in prior work examining SI trajectories over 12-months in a sample of youth recruited during an inpatient hospitalization (10.9% ),¹⁰ which likely reflects the severity of the CARES Study sample. Nightmares emerged as a univariate predictor of both SI and SH non-response and sleep problems were also a univariate predictor of SI non-response. Although these variables did not remain significant in multivariate analyses, the well-documented association between sleep problems and suicidality¹² suggests that addition sleep interventions may improve response rates, although the extent to which sleep problems can be accounted for by internalizing symptoms, which were also a significant univariate predictor of SI, needs further exploration.. PTSD symptoms were not a significant predictor of response for any of the outcome variables examined in this study; however, further consideration is warranted given the association between PTSD and nightmares, and the potential for additional PTSD intervention to augment treatment response.

Results provided some clues regarding early identification of youth likely to be total non-responders. We identified a formula that predicted the likelihood of total non-response using a combination of variables measured prior to starting treatment. While this formula requires replication in an independent sample, the approach has potential for guiding efforts to personalize treatment strategies, and matching youth with high risk for suicide/suicide attempts to treatment approaches that are most likely to be effective. Greater baseline rates of SI and SH were predictors of total non-response. Hence, total non-responders may constitute a subgroup of suicidal youth with a more severe illness trajectory that may require novel therapeutic targets and/or require higher levels of care to reduce suicide risk; however, the relative effectiveness of different levels of care remains unknown. Parent-reported youth externalizing symptoms were also a predictor of total non-response. This is consistent with prior work showing that externalizing symptoms are associated with elevated suicide risk and may characterize a subtype of suicidal youth characterized by impulsivity and aggression versus depression^46,47. Research on borderline personality disorder in youth has also identified an externalizing subtype associated with increased anger, impulsivity, oppositionality and deficits in executive functioning.^48,49 Hence, additional focus on externalizing symptoms may improve treatment response. Reports of youth externalizing symptoms by parents may also represent a need for additional behavioral parenting intervention.⁵⁰ Neither sleep nor depression were univariate or multivariate predictors of total non-response. Further research is needed to determine alternatives or additions to DBT for youth who are likely to be total non-responders, given that DBT is the only replicated treatment for reducing SH in this patient population.^5–7,16

White race was a significant multivariate predictor of both SI non-response and total non-response. The CARES sample was 56% White and 27% Latina, with the remaining 17% being Black, Asian, “other,” and Native American⁶. The finding that ethnic minority youth were more likely to benefit from the study interventions is consistent with accumulating data indicating stronger intervention responses among minority populations in studies of depressed youth and adults,^51,52as well as with other studies on suicidal youth. Prior secondary analyses of the CARES Study data found that Latina ethnicity was a moderator of treatment response, with Latina youth who received DBT having a lower likelihood of having a SA during treatment than non-Latina youth^1,2. In addition, a study of family-focused treatments for suicidal youth also found greater reductions in SI among ethnic minority and/or low-income youth (who were predominately Black) as compared to White and/or high-income youth.⁵³ These findings require cautious interpretation, as not all studies have found an advantage for ethnic and racial minority youth.⁵⁴ One possible explanation for the results of the present study is that Latina participants were predominately recruited from a county hospital setting serving underserved families and may not have had prior access to evidence-based treatments, which may account for their more robust response.⁵³ Of note, although no cultural adaptations were made to the content of DBT or IGST, Latina therapists who were fluent in Spanish provided treatment to Spanish-speaking families. Ethnic matching of therapist and patient has been shown to improve treatment engagement and adherence.⁵⁵ Further research is needed to determine if these findings generalize to other samples of suicidal adolescents and are unique to Latina youth versus other ethnic minority groups. These results suggest the potential value of evidence-based treatments for enhancing health equity and reducing risk of repeat SH among ethnic and racial minority youth.

Limitations to generalizability include the predominately female and high-risk sample and the absence of a cross-validation sample for the optimal threshold analyses. We were unable to measure trajectories of response for SA alone, given that the rate of SAs decreased consistently over the follow-up period and the latent class model for SA yielded only one class containing most of the sample, with subsequent classes being too small to meaningfully discriminate between. The SH variable used primarily included NSSI. Future work with a larger sample of youth with SA is needed to examine trajectories of response for SA alone.

In conclusion, this research extends prior work by examining trajectories of treatment response and non-response for SH over the course of 6-months of treatment and 6-months of follow-up. This is the first study that we are aware of to examine adolescents’ trajectories of response to DBT, the only replicated treatment for reducing SH in youth and a treatment that has already been widely disseminated into the community. Across both study treatments (DBT and IGST), a substantial percentage of youth, including those with high levels of baseline SI and SH, showed trajectories of improvement. Youth who received DBT were more likely to show trajectories of improvement in SH than youth who received IGST, consistent with prior CARES Study findings and further supporting the benefits of DBT for this patient population. Further research is needed to examine the efficacy of DBT as compared to other evidence-based treatments for youth suicide risk. Results also highlighted a subset of youth with high suicide risk that do not appear to respond to treatment and may worsen while in treatment. Additional research is needed to determine effective treatment approaches and therapeutic targets for this subset of youth, who remain at high risk for suicide despite receiving 6-months of treatment. Possible future areas of study for increasing the efficacy of DBT for adolescents at high suicide risk include additional behavioral parent training, examining the impact of combined DBT and specific medication algorithms, sleep-focused interventions, and longer-term follow-up care to address the persistence of suicide risk over time.