Factors Affecting Antidepressant Response Trajectories: A Veterans Affairs Augmentation and Switching Treatments for Improving Depression Outcomes Trial Report

Major depressive disorder (MDD) accounts for the greatest number of disability‐adjusted life years among psychiatric disorders (1). Thus, optimizing pharmacological interventions for the management of MDD is a critical goal. Attempts at characterization of antidepressant treatment response have increasingly focused on analysis of response trajectories (2, 3). Using the antidepressant agent venlafaxine XR, six response trajectory groups were observed (2). That study corroborated the frequently documented finding that over one‐half of patients will have limited improvement with an antidepressant trial, and identified that high baseline depression and anxiety scores predicted being in the least responsive trajectories. One recent attempt to identify improved strategies for antidepressant use was the Veterans Affairs (VA) Augmentation and Switching Treatments for Improving Depression Outcomes (VAST‐D) clinical trial, which addressed whether there was an advantage to switching antidepressants rather than augmenting with a second antidepressant or an atypical antipsychotic (4). VAST‐D's large sample size provides an excellent opportunity for careful characterization of treatment response patterns.

In this secondary analysis of the VAST‐D data, we used group‐based trajectory modeling (GBTM) to evaluate the response trajectories for VAST‐D participants. GBTM is a semiparametric technique that identifies a finite number of groups (trajectories) whose members follow similar patterns of response (5, 6, 7). Although GBTM does not make any a priori assumptions about the existence of trajectories in the population, it allows the identification of early and late responders, reduces the variability of parameter estimates, and accounts for uncertainty in individual group assignments. Using GBTM with the VAST‐D data, we attempted to (1) identify unique trajectories of primary and secondary outcomes during acute phase treatment, (2) characterize the response trajectories of symptom clusters, and (3) determine whether any specific VAST‐D interventions or other clinical/demographic factors would influence the likelihood of inclusion in specific trajectories, either for overall response or for symptom clusters.

METHODS

Compliance

All procedures involving human subjects/patients were approved by the VA Office of Research and Development and the VA Central Institutional Review Board, and a Certificate of Confidentiality was obtained from the National Institutes of Health. Annual reviews were conducted by the VA Central Institutional Review Board, and a Data Monitoring Committee reviewed the study biannually. Adverse events were reviewed by the VA Central Institutional Review Board and Data Monitoring Committee throughout the study. All participants provided written informed consent and privacy authorization after the procedures had been fully explained.

Study Design

VAST‐D was a multisite, randomized, single‐blind, parallel‐assignment, next‐step trial in veterans whose MDD was inadequately responsive to at least one course of antidepressant treatment with a selective serotonin reuptake inhibitor (SSRI), serotonin and norepinephrine reuptake inhibitor (SNRI), or mirtazapine that met or exceeded minimal treatment standards for dose and duration (2). Inadequate response was defined as a Quick Inventory of Depressive Symptomatology‐Clinician Rated (QIDS‐C; 8) score ≥16 (severe depression) after at least 6 weeks of treatment or a score ≥11 (moderate depression) after at least 8 weeks of treatment with the 3 most recent weeks at a stable dose. A full description of the overall design (including the CONSORT statement and flow diagram) was given in earlier manuscripts (4, 9).

Participants

Veterans Health Administration (VHA) patients with an MDD diagnosis were included in the study if they were at least 18 years old and were referred by a VHA clinician. Before enrollment, study clinicians confirmed the MDD diagnosis, and research staff reconfirmed diagnostic eligibility using criteria from the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition, Text Revision (DSM‐IV‐TR) (10). Exclusion criteria included: pregnancy or breast‐feeding; currently using contraindicated medications, including either study drug; a clear history of non‐response or intolerance to bupropion‐SR or aripiprazole; a primary diagnosis of bipolar, psychotic, obsessive‐compulsive, dementia, or eating disorders; general medical conditions contraindicating the use of bupropion‐SR or aripiprazole; serious, unstable medical conditions requiring acute treatment; meeting criteria for substance dependence that required inpatient detoxification; or in need of acute treatment because of suicide risk.

Interventions

This report addresses the acute phase of treatment from the VAST–D study, in which 1522 veterans with MDD were randomized to one of three treatment groups: (1) augmenting an SSRI/SNRI/mirtazapine with bupropion SR (Aug‐BUP), (2) augmenting an SSRI/SNRI/mirtazapine with aripiprazole (Aug‐ARI), or (3) switching to another antidepressant, that is, bupropion‐SR (Switch‐BUP) (4, 9). Treatments included titration (cross‐titration for the switching arm)—from standard starting daily doses of either 150 mg bupropion‐SR with titration up to 400 or 2 mg aripiprazole with titration up 15 mg—until depressive symptoms remitted or side effects were intolerable. Dose adjustments were guided using the Patient Health Questionnaire (PHQ‐9) (11) and Frequency, Intensity and Burden of Side Effects Rating (12) obtained at each visit (baseline and at the end of weeks 1, 2, 4, 6, 8, 10, and 12).

Assessments

Baseline Measures

The baseline measures of our analysis included demographic factors (age, education, employment status, marital status, and race/ethnicity) and clinical factors or assessments (duration of index episode, presence of a substance or alcohol abuse diagnosis by the Mini‐International Neuropsychiatric Interview [M.I.N.I.] (13), Adverse Childhood Experiences Survey (14), Beck Anxiety Inventory [BAI] (15), Columbia Suicide Severity Rating Scale‐Lifetime Suicidal Ideation [C‐SSRS] (16), 9‐item adaptation of the Brief Grief Questionnaire documenting the participants' responses to the death of a close relationship [as applicable] (17), self‐rated Mixed Features Scale based on the DSM‐5 (18), Cumulative Illness Rating Scale (19), PHQ‐9, Quality of Life Enjoyment and Satisfaction Questionnaire‐Short Form [Q‐LES‐Q‐SF] (20), and QIDS‐C (8)). The QIDS‐C evaluates the symptoms of sleep, mood, appetite, concentration, guilt, acute suicidal ideation, interest, fatigue, and psychomotor function. Three QIDS‐C symptom clusters have been characterized: core emotional cluster (energy/fatigability, concentration/decision making, loss of interest, mood, and feelings of worthlessness), sleep cluster (mid‐nocturnal insomnia, sleep‐onset insomnia, and early morning insomnia), and atypical cluster (psychomotor agitation, psychomotor slowing, suicidal ideation, and hypersomnia).

Outcome Measures

The primary outcome measure, QIDS‐C, was collected by an independent evaluator who was blind to the treatment assignment at baseline and each visit following randomization. The PHQ‐9 was collected as a secondary measure. We used standard definitions of “response” (≥50% decrease in the baseline symptom score at the end of Week 12) and “remission” (symptom score scores ≤5 on two consecutive evaluations) (2). Early improvement was defined as a ≥20% drop from baseline QIDS‐C score by the end of week 2.

Statistical analysis

Trajectory Analysis

We assumed a censored normal distribution of the outcome measures (QIDS‐C or PHQ‐9) (21). GBTM, performed using Proc TRAJ from SAS 9.4.2 (22), uses maximum likelihood estimation to determine group sizes, the polynomial order and drop‐pattern of each trajectory, and groups of individuals following similar response pathways. Groups were added to the model in a step‐wise fashion, thereby assessing each group's contribution to the overall fit of the model at each step. For every subsequent addition of a group, the log Bayes factor was calculated to assess whether the addition of that group provided a better model fit. The log Bayes factor was obtained by multiplying by two the difference in Bayesian Information Criterion (subtracting a less complex model from a more complex model) for the two models under comparison. A log Bayes factor >10 was used as a benchmark to favor the more complex model. The polynomial order for each trajectory was also obtained using the log Bayes factor as a criterion for each added order. Four a priori criteria were used to assess the adequacy of the performance of the trajectory groups identified by GBTM: (1) the average estimated posterior probabilities of group membership are at least 70% (Mean Posterior Probability of Group Membership; MPP); (2) the odds of correct classification (OCC) into a group in comparison to the odds of group membership by random assignment is ≥5; (3) The differences between estimated and the actual group proportions (DEAP) for each group are expected to be <10%; and (4) the minimum group size should be ≥5% of the total population (6).

Weighted Multinomial Logistic Regression and Exploratory Analyses

To identify factors influencing assignment to trajectory groups, we performed a weighted multinomial logistic regression analysis using the posterior probability of group membership as weight. For the weighted logistic regression analysis of both the QIDS‐C and PHQ‐9, we implemented a Bonferroni correction for comparison of the two measures total QIDS‐C and PHQ‐9 such that the acceptable type I error rate in the multinomial logistic regression analysis was set to p < 0.025. We did not apply a Bonferroni correction for the multinomial logistic regression analyses of QIDS‐C clusters because they were considered exploratory. In other exploratory analyses, we identified additional factors associated with trajectory group membership by performing either chi‐square analysis (categorical data) or an analysis of variance (continuous data). For these analyses, Bonferroni corrections were not applied. All covariates included in this analysis are listed in the Baseline Measures section. This analytic approach was repeated using PHQ‐9 data. In addition, the same analyses were performed on QIDS‐C clusters (core emotional, sleep, and atypical). The QIDS‐C Core Emotional Cluster scores are based upon a sum of QIDS‐C scores for the five following items: energy/fatigability, concentration/decision making, loss of interest, mood, and feelings of worthlessness. Each item is rated on a scale from 0 to 3, with 3 representing greater severity. Scores range from 0 to 15, with higher scores indicating greater severity of symptoms. The QIDS‐C Sleep Cluster scores are based upon a sum of QIDS‐C scores for the following three items: mid‐nocturnal insomnia, sleep‐onset insomnia, and early morning insomnia. Each item is rated on a scale from 0 to 3, with 3 representing greater severity. Scores range from 0 to 9, with higher scores indicating greater severity of symptoms. The QIDS‐C Atypical Cluster scores are based upon a sum of QIDS‐C scores for the 4 following items: psychomotor agitation, psychomotor slowing, suicidal ideation, and hypersomnia. Each item is rated on a scale from 0 to 3, with 3 representing greater severity. The total range is from 0 to 12, with higher scores indicating greater severity of symptoms.

RESULTS

Trajectory Analysis

The optimal number of trajectory groups for the VAST‐D QIDS‐C or PHQ‐9 data is six (Supporting Information S1: Supplement B) (6). This number is based upon the criterion that the log Bayes' factor associated with the addition of a group must be >10 for the group to qualify as a significant addition to the model. The addition of a seventh group to the model also produced a Bayes' factor >10, but resulted in the proportion present in trajectory 7 being <5%. Therefore, the model with seven trajectory groups was not chosen for either analysis. The only predetermined criterion violated in the analysis using the QIDS‐C was that the difference in the actual and the estimated proportion for trajectory 3 was >10% (22.5%) (Supporting Information S1: Supplement C). In the analysis performed on the PHQ‐9, we observed that there was a difference in the actual and estimated proportions for trajectory 6 (14.3%). However, the confidence intervals of the estimates of group membership probabilities were reasonably tight for QIDS‐C and PHQ‐9, indicating a good fit of the model. We used the log Bayes factor criteria for all combinations of quadratic and linear trajectories (5) and optimized our model to two quadratic and four linear trajectories.

Figure 1 illustrates average QIDS‐C and PHQ‐9 scores over the 12‐week acute phase for each of the observed group trajectories. Similar patterns of response were seen for each QIDS‐C cluster (Supporting Information S1: Supplement D). The QIDS‐C and PHQ‐9 trajectories showed similar patterns (Tables 1 and 2). Trajectories 1–3 included nearly all remitters (99.0% and 89.1% for QIDS‐C and PHQ‐9 scores, respectively) but only a small percentage of non‐responders (11.9% and 19.6% for QIDS‐C and PHQ‐9 scores, respectively). In contrast, trajectories 4–6 included most of the non‐responders (88.1% and 80.4% for QIDS‐C and PHQ‐9 scores, respectively) but only a small proportion of remitters (1.0% and 10.9% for QIDS‐C and PHQ‐9 scores, respectively). Patients included in trajectories 1–3 were the least likely to withdraw because of a lack of treatment response or worsening of symptoms (20.0% and 14.5% of all withdrawing for lack of treatment response; see Supporting Information S1: Supplement E).

Weighted Multinomial Logistic Regression Analysis

The odds of inclusion into specific trajectory groups versus inclusion into the least responsive group (trajectory 6) was estimated for each of the baseline measures (Tables 3, 4, 5, 6). Elevated baseline total QIDS‐C and PHQ‐9 scores were more likely to be associated with less responsive trajectories. The baseline QIDS‐C severity finding was also present in all QIDS‐C cluster analyses. In contrast, early improvement increased inclusion into responsive trajectories, an effect that was evident for all QIDS‐C clusters. Higher baseline BAI scores were more likely to be present in the least responsive trajectories for all QIDS‐C clusters. Higher baseline Q‐LES‐Q‐SF scores marginally increased inclusion into responsive trajectories for both the total QIDS‐C and PHQ‐9 scores, but that effect was not observed wth any of the cluster analyses. For the QIDS‐C sleep cluster, the benefits of employment were strong but advanced age and decreased duration of the index episode provided only marginal benefits. For the QIDS‐C atypical cluster, a modest benefit was seen for younger age or lower lifetime suicidal ideation. Greater severity of health impairment marginally increased inclusion in the least responsive trajectories. In the atypical cluster analysis, being married or cohabiting greatly increased the likelihood of inclusion in the more responsive trajectories. In these weighted multinomial logistic regression analyses, treatment allocation had no influence on trajectory inclusion for the PHQ‐9, QIDS‐C, or any QIDS‐C cluster.

Exploratory Analysis of Clinical and Demographic Factors

Table 7 shows the statistical significance of the influence of clinical and demographic factors on inclusion into specific trajectory groups, and Supporting Information S1: Supplement F shows the actual numbers and percentages of individuals in each trajectory group. Several clinical and demographic factors increased the likelihood of inclusion into more responsive trajectories of the QIDS‐C, including being employed, female, or Caucasian, endorsing three or fewer grief items, receiving treatment allocation Aug‐ARI, experiencing a shorter index episode, having lower baseline anxiety or depression scores, fewer mixed features, a higher baseline quality of life, and lower lifetime suicidal ideation. The benefits of five of these factors were driven by contributions from all three QIDS‐C clusters: being employed, having lower baseline anxiety or depression severity scores, experiencing a shorter index episode, and reporting a higher baseline quality of life. However, the influence of some factors was seen to be uniquely affected by certain QIDS‐C clusters. For example, the sleep cluster influenced the race‐based findings, whereas the gender‐based findings had contributions from the sleep and atypical clusters. Grief endorsement effects were influenced by the sleep and core emotional clusters. The benefit of Aug‐ARI treatment allocation was entirely dependent on changes in the core emotional cluster. The benefit of decreased lifetime suicidal ideation resulted from the influence on the atypical cluster, which includes the QIDS‐C item on suicide ideation. Although an overall benefit was not demonstrated, a beneficial effect of having fewer mixed features was associated with outcomes of the sleep and atypical clusters.

Factors that influenced inclusion into more responsive clusters according to the PHQ‐9 included being employed, being female, endorsing three or fewer grief items, demonstrating early improvement, having lower baseline anxiety or depression severity scores, a shorter index episode, the presence of fewer mixed features, and a higher baseline quality of life.

DISCUSSION

Understanding antidepressant response patterns helps shape the decision‐making process in the clinical management of depressed patients. Using GBTM, six response trajectories were identified, similar to findings by other investigators (2, 3). The same patterns were found using the QIDS‐C and the PHQ‐9. However, the QIDS‐C provides additional utility because its response clusters can be analyzed to tease out more subtle findings regarding factors influencing outcomes. The presence of these patterns across multiple studies derived from different patient populations using diverse treatment interventions suggests these response patterns could be representative of all antidepressant trials. Previous studies focused solely on the initial acute phase of treatment, but VAST‐D studied participants that were ready for a “next‐step” intervention. The similarity of response patterns despite the different phases of treatment suggests that antidepressant response trajectories are comparable across treatment stages.

Identifying the factors that affect antidepressant response helps inform clinical decisions. Many clinical and demographic factors in this study were found to influence antidepressant response. Weighted multinomial logistic regression analysis for the QIDS‐C, QIDS‐C clusters, and the PHQ‐9 shows a strong role of the severity of baseline anxiety and depression, as well as early improvement in determining responsiveness. These findings are consistent with the findings of the analysis of moderators of treatment effect for the VAST‐D study (23) and with a body of literature highlighting the negative influence of a higher baseline severity of depression (23, 24, 25, 26). However, some studies have found the degree of depression severity positively correlates with response/remission rates (27, 28) or is unrelated to inclusion into specific response trajectories (3). Clinically, it makes sense that the more severely depressed patients would be the least responsive to treatment. In addition, multiple studies, in agreement with the current findings, have demonstrated the negative effects of higher baseline anxiety levels (23, 29, 30, 31). Finally, in contrast to the findings of Uher et al. (3), most studies, including the present study, have demonstrated that the lack of early improvement predicts non‐response and non‐remission (32, 33, 34, 35, 36).

In our exploratory analyses, being employed increased the likelihood of responsiveness in the QIDS‐C, all QIDS‐C clusters, and the PHQ‐9. Other studies have also documented the positive effects of employment (23, 25, 31, 37). Although our data does not allow us to identify all related factors, it is possible that employment could be associated with unmeasured factors (e.g. the degree of life engagement or having a sense of purpose). This finding also raises the question of whether encouraging and supporting employment could be a positive therapeutic intervention. The presence of a life partner would be expected to provide stability and improve psychological health (26). It is, therefore, not surprising that marital/co‐habitation benefits were found, but they were associated only with the sleep and atypical clusters. Despite that, life quality at baseline had only a marginal positive influence on trajectory assignment.

By evaluating the influence of various factors on the trajectories of QIDS‐C clusters, it is apparent that clusters are often differentially driven by specific demographic and clinical factors. For example, as would be predicted by the greater likelihood of response and remission from treatment allocation to Aug‐ARI in the initial acute phase intervention VAST‐D analysis (4), Aug‐ARI treatment allocation did increase the probability of inclusion into more responsive trajectories. The present analysis demonstrates that this important effect is primarily driven through the influence of aripiprazole on the QIDS‐C core emotional cluster, which is the primary focus of treatment interventions. A similar treatment allocation effect could not be detected from PHQ‐9 trajectories. In contrast, the role of mixed features and gender effects do not involve the core emotional cluster; instead, they are mediated by the sleep and atypical clusters. Also, grief endorsement effects on trajectories are driven by the core emotional and sleep clusters, not the atypical cluster. The apparent benefit of Caucasian race was driven solely by a signal from the sleep cluster and should be cautiously interpreted because of the predominance of Caucasians in this analytical sample. Finally, the presence of lifetime suicidal ideation expectedly appears to influence the magnitude of the QIDS‐C atypical symptom cluster, which includes rating of acute suicidal ideation.

Strengths and Limitations

The present analysis of the VAST‐D data has several strengths. First, because of the large patient population, this study was ideally suited to perform trajectory analysis and obtain precise estimates of the effect of various factors on trajectory membership. Second, the present analysis extracted trajectories based on outcomes over time, without regard to treatment allocation, minimizing any associated bias. Third, we explored the role of novel factors such as duration of index episode, childhood adversity, and quality of life in predicting trajectory group membership. Finally, we used commonly used tools of psychometric measurement for depression scores, which improves the generalizability of the findings and facilitates comparisons with other studies.

This study does have its limitations. First, even though the VAST‐D trial was conducted in a diverse sample with regard to many demographics and historical features (38), the gender of the patient population was predominantly male (approximately 85%), and the race of participants was predominantly Caucasian (approximately 70%). Hence, the findings of effects of gender and race on trajectory group membership in the exploratory analysis should be interpreted with caution. A second limitation is that trajectories are only approximations of possible patterns in the population. Some argue that GBTM may “over‐extract” trajectories, resulting in too many complicated patterns and leading to uncertain clinical conclusions (4, 39, 40, 41). Third, because individuals are assigned trajectory groups based on their posterior probability of membership, some uncertainty in trajectory membership may exist. In this analysis, we attempted to address this uncertainty in the multinomial regression by using posterior probability of trajectory group membership as a weight in multinomial logistic regression analysis. Finally, it is also possible that we may have undervalued the role of factors with potential to influence treatment response (e.g. familial history of depression or number of life stressors) that were not studied.

Importance of Findings

In this study, we used the VAST‐D trial database to explore differential trajectories of improvement in depression scores using a relatively new statistical procedure, GBTM (3, 4, 5, 40). Our characterization of the response trajectories helps to establish reasonable short‐term clinical expectations of an antidepressant trial. Although we identified several factors that were associated with specific patterns of response, across multiple statistical probes, we consistently recognized three factors: baseline depression, baseline anxiety, and early improvement. Thus, GTBM analysis based on the large population of the VAST‐D data set findings (N = 1522), substantiated a previous GBTM study with a much smaller number of advanced age participants (N = 453) (21) identifying that baseline depression and anxiety severity are very important factors in determining outcome. In addition, our weighted multinomial logistic regression analysis following GTBM corroborates another VAST‐D analysis highlighting the role of baseline depression and anxiety in determining response (23). Our GTBM analyses also confirmed the role of early improvement (by the end of Week 2) as a predictor of response (33). Finally, the PHQ‐9, as a self‐report measure, reproduced the key findings of the QIDS‐C, reinforcing the benefit of using it to follow antidepressant response in clinical settings.