Introduction
Roughly 20–30% of patients with major depression develop a chronic course lasting two years or longer [
1,
2]. In the fifth edition of the Diagnostic and Statistical Manual (DSM-5), this condition was first introduced as a distinct clinical category labelled as Persistent Depressive Disorder (PDD) [
3]. Psychotherapy and pharmacotherapy, delivered as monotherapies or in combination, represent, in addition to brain stimulation, two main pillars of treatment for PDD, with the Cognitive Behavioral Analysis System of Psychotherapy (CBASP) [
4] being the only psychotherapy model specifically developed to target PDD. CBASP has proven to be overall effective and has been thus recommended as first line psychotherapeutic treatment for PDD [
5]. Nevertheless, various findings have shown that CBASP may not be the most effective treatment for all patients with PDD [
6‐
9]. Similarly, the effect of pharmacotherapy appears to be limited for certain patients with PDD. For instance, a review by Kocsis [
10] showed that the average rate of complete remission for patients with dysthymia and double depression was below 50% in several 6- to 12-week short-term studies.
Furthermore, pharmacotherapy was shown to be generally more effective (
d = −0.31; 95% CI: − 0.53 to − 0.09) than psychotherapy in an earlier meta-analysis by Cuijpers and colleagues from 2010 [
11]. However, this former result was exclusively due to patients with dysthymia included in the analysed studies, leaving open the question of which type of treatment works better for patients with other subtypes of PDD. In a more recent meta-regression by Furukawa and colleagues from 2018 [
8], psychotherapy and pharmacotherapy showed similar results when delivered as monotherapies. However, this result was only valid for patients with characteristics near the population averages (e.g., low or moderate baseline depression and anxiety), with both monotherapies displaying different effects depending on the severity of baseline depression and anxiety, previous history of pharmacotherapy, age at baseline, and PDD subtypes. These results suggest that for some subgroups of patients, either CBASP or pharmacotherapy alone is a more effective treatment option and highlight the need for further investigations.
To date, there is little empirical evidence guiding clinicians to select the most effective treatment option for an individual patient with PDD [
12,
13], and it is likely that the lack of personally selected and tailored treatment strategies is one of the main contributors to the overall low treatment success of patients with PDD [
14]. In clinical practice, treatments for PDD are commonly selected in an unsystematic matter, often based on subjective clinical experience, treatment-preference of patients or trial-and-error approaches [
12,
15,
16]. Over the last two decades, only a modest number of studies have aimed to gain a better understanding of which subgroups of patients with PDD are most likely to benefit from a particular psychotherapeutic or pharmacotherapeutic treatment. With regard to the choice between psychotherapy and pharmacotherapy, studies have analysed the impact of baseline severity of depression and anxiety as well as patients age [
8], self-reported traumatic childhood experiences [
17,
18], dysfunctional attitudes [
19] and patients treatment preference [
7,
20]. The results of the meta-regression by Furukawa and colleagues from 2018 [
8] showed that for patients with elevated initial depression and anxiety scores, combination treatment of psychotherapy and pharmacotherapy was generally more effective than pharmacotherapy alone, which in turn was more effective than treatment with CBASP alone. In contrast, patients with moderate baseline depression scores and mild anxiety scores benefited equally well from combination treatment and treatment with CBASP alone, but less from treatment with pharmacotherapy alone. In addition, this study reported that monotherapy with antidepressants was more likely to be tolerated by younger patients with PDD, who had a lower dropout rate when treated with antidepressants compared to CBASP. Regarding the moderating effect of childhood traumatic experiences, different results were reported across studies: In a secondary analysis by Nemeroff and colleagues from 2003 [
18], monotherapy with CBASP was shown to be superior to monotherapy with nefazodone in patients who reported childhood trauma. However, these findings could not be replicated in a later study by Bausch and colleagues from 2017 [
17], who concluded that CBASP and escitalopram—a modern antidepressant delivered in combination with clinical management—were equally effective in treating patients with PDD and childhood trauma, with CBASP possibly having a longer treatment latency in these patients. Furthermore, conflicting results were also reported for the role of patients' treatment preference: While Kocsis and colleagues [
7] reported in 2017 that patients preferring CBASP had better treatment outcomes when receiving CBASP than when receiving nefazodone, and vice versa, Steidtmann and colleagues [
20] could not replicate this association in their study published in 2012. Finally, another study by Shankman and colleagues from 2013 [
19] showed that higher baseline scores of dysfunctional attitudes were associated with a better response to pharmacotherapy compared to psychotherapy.
Despite yielding first interesting although not always replicable results, this previous evidence base comes along with several limitations: First, the individually examined baseline variables of these studies do not reflect the full individuality of a patient, who will have many other variables which are potentially critical to its treatment response that are not taken into account when focusing on a single variable. Second, individual baseline variables often have little effect size as moderators, which limits their practical relevance with regard to treatment selection in clinical practice [
21]. Third, the composition of other unconsidered baseline variables in a clinical sample may influence the results of a stratifying predictor or moderator analysis, which may partly explain why previous secondary analyses have repeatedly produced contradictory results. Fourth, the existing evidence base may result in conflicting treatment recommendations for clinical practice. For example, to a patient reporting childhood trauma and a preference for antidepressant medication, one would recommend antidepressants over CBASP based on the findings by Kocsis and colleagues [
7] and at the same time CBASP over antidepressants based on the findings by Nemeroff and colleagues [
18]. Taken together, these factors complicate the evidence-based treatment selection for clinicians, and necessitate newer statistical approaches that capture the integral individuality of patients and use it to predict outcomes under different psychotherapeutic and pharmacological treatments.
The overarching aim of the present study was to address the question of which outpatients with PDD are, based on their multivariable baseline profile, more likely to benefit from psychotherapy with CBASP than from pharmacotherapy with escitalopram during the first eight weeks of treatment, and vice versa, thereby adding new findings to the existing body of evidence. In contrast to the studies summarized before, rather than examining single moderating baseline variables, we sought to exploratory identify subgroups of PDD patients with different treatment benefits using a modern composite moderator method together with machine learning that enable to simultaneously consider the treatment effect moderating role of multiple baseline variables. Our analyses were based on the data of a bi-centric randomized controlled trial (RCT) by Schramm et al. [
22], who compared the effectiveness of CBASP to escitalopram, a well-tolerated standard selective serotonin reuptake inhibitor, combined with clinical management (ESC/CM) over 28 weeks in a sample of outpatients with PDD. The general findings showed that the clinician-rated depression scores decreased significantly after both eight and 28 weeks, however with no significant differences between the two treatment groups. Furthermore, in the original RCT, in case of non-improvement (defined as < 20.0% reduction in depression severity) after the 8-week acute treatment phase, the other treatment condition was augmented for the following 20 weeks of the extended treatment phase. Non-improvers to the initial treatment caught up with the initial improvers in terms of depression severity by the end of the extended treatment phase after being augmented with the respective other condition [
22]. In conclusion, CBASP and ESC/CM appeared to be equally effective treatment options for chronically depressed outpatients in both the acute and extended treatment phase, whereas for patients who did not respond to their first treatment in the acute phase, augmentation with the other condition during the extended phase appeared to be effective in reducing depression severity.
The secondary analysis of this RCT presented in this paper was conducted to revisit this conclusion by examining whether, despite the reported general equivalence of the two treatments, there were in fact ‘hidden’ subgroups of patients who were likely to benefit more from CBASP than from ESC/CM and vice versa during the acute 8-week treatment phase. In addition, we investigated whether the initial lack of response in those patients augmented with the other treatment condition at week eight was because they did not receive their likely more effective treatment during the first eight weeks, and whether the observed improvement at week 28 was due to the augmentation with the treatment condition from which they would have likely benefitted more from the beginning of the treatment. By considering multiple baseline variables, our analysis methodologically extends an earlier, previously cited secondary analysis of this RCT by Bausch et al. [
17] from 2017, who found that ESC/CM outperformed CBASP in patients with childhood trauma after the 8-week acute treatment phase and that both therapies were equally effective in patients with childhood trauma after the extended treatment phase. By addressing the research question of
‘what works for whom’ with regard to the choice between CBASP and pharmacotherapy with escitalopram for PDD, we pursued the overarching aim of adding more sustainable evidence that can help clinicians in choosing between psychotherapeutic and pharmacological treatments, thereby addressing the urgent need to advance personalized medicine for PDD [
13].
Materials and methods
The original study by Schramm et al. [
22], on which this secondary analysis is based, was an evaluator-blind, parallel-design, 2-armed RCT conducted between 2008 and 2013 at two university medical centers in Germany. The study was approved by the institutional review boards and ethics committees at each site. Written informed consent was obtained from all participants before study enrollment. Study registration was performed at the University Register of Clinical Studies (No. 2007-006914-41) and at
www.clinicaltrials.gov (No. NCT00837564). For details on participant inclusion and exclusion criteria, please refer to the main publication of the original trial [
22].
Study interventions
CBASP: The CBASP is a highly structured interpersonal learning approach that integrates behavioral, cognitive and, most importantly, interpersonal treatment strategies with personal disciplined involvement [
9,
17]. Based on the assumption that early interpersonal trauma has led to dysfunctional mechanisms of derailed affective and motivational regulation and a reduction in perceived functioning, the overarching goal of CBASP is to help the patient to recognize the consequences of one’s own behavior for others, to reduce social fear in the interpersonal hot spot area, and to develop social problem-solving skills and empathy [
23]. In the original RCT, in the initial acute treatment phase of eight weeks, two weekly CBASP sessions were conducted during the first four weeks and one weekly CBASP session during the last four weeks to ensure a minimum of 12 CBASP sessions. Study therapists were experienced and trained professionals, and all sessions were videotaped and viewed regularly by the supervisors.
ESC/CM: The second treatment condition in this RCT was escitalopram as a well-tolerated standard selective serotonin reuptake inhibitor with an excellent benefit/side-effect ratio [
24]. A minimum of a 2-week washout period of the previous antidepressant medication was required for study participation, if indicated. The initial dose of escitalopram was 10 mg/day for the first week, which was increased to 20 mg/day in the following weeks. For patients experiencing dose-related side effects, their dose could be reduced to a minimum of 10 mg/day later. Clinical Management is a psychoeducative, supportive and empathic intervention including symptom management, monitoring of the medication and its possible side effects, providing hope, encouragement, and simple advice. The guideline based visits were conducted by senior psychiatrists or advanced psychiatric residents, taking place weekly during the acute phase, and were limited to 20 min [
22].
Study sample and analyzed baseline variables
Sixty patients including one non-starter who was excluded from the analyses were randomly assigned to receive treatment with CBASP (n = 29) or ESC/CM (n = 31) over a total period of 28 weeks, including an acute treatment phase within the first eight weeks. Of the n = 59 patients who began treatment, n = 6 discontinued it before the end of the acute treatment phase, resulting in n = 53 completers (n = 27 CBASP; n = 26 ESC/CM), who were included in the present moderator analysis. Notably, study discontinuation occurred due to motivational or logistical reasons (e.g. move to another city, start of another therapy). We did not detect any patients in the ESC/CM group who discontinued study participation due to side effects from taking escitalopram.
The utilized statistical procedure [
25] of this analysis preselects and combines multiple individual baseline variables into one optimal composite moderator (
M*) to detect possible subgroup effects. Baseline variables were assessed by evaluators blinded to the treatment condition. For being preselected for the compilation of
M*, a baseline variable had to contain at least n = 50 valid cases or no more than three missing cases so as not to reduce the sample size relevant for the final regression analysis. Our set of preselected initial baseline variables thus comprised 11 baseline variables from a wide range of domains (see Table
1).
Table 1
List of initially considered baseline variables
Socio-demographic characteristics |
1. Female gender | Nominal | Yes/no |
2. Age | Metric | Years |
Clinical characteristics |
3. Early illness onset | Nominal | Defined as an onset of PDD before the age of 21; yes/ no |
4. Depression severity | Metric | Clinician-rated MADRS total score at baseline |
5. History of suicidality | Metric | Self-reported number of previous suicide attempts |
6. Comorbidity of ≥ 1 Axis-I disorder | Nominal | Yes/no; diagnosed with the SCID-I by clinician |
7. Comorbidity of ≥ 1 Axis-II disorder | Nominal | Yes/no; diagnosed with the SCID-II by clinician |
Childhood and life trauma |
8. Childhood trauma | Nominal | Self-reported moderate-to-severe childhood trauma that occurred before the age of 18 in at least one of the five dimensions of the CTQ; yes/ no |
9. Adverse life events | Metric | Item assessing the number of self-reported major psychosocial stressors over the lifetime |
Previous treatments |
10. Previous psychotherapies | Ordinal | Self-reported number of previous psychotherapies, provided in categories (0 = none, 1 = 1, 2 = 2, 3 = 3, 4 = 4, 5 = more than 5) |
11. Previous medication | Ordinal | Self-reported number of previous treatments with antidepressants, provided in categories (0 = none, 1 = 1, 2 = 2, 3 = 3, 4 = 4, 5 = more than 5) |
Main outcome
The main outcome in this secondary analysis was the percentage change in MADRS scores from baseline to week eight (corresponding to the end of the acute treatment phase) calculated according to the following equation:
$$percentchang{\text{e}}_{MADRS} = \frac{{MADRS_{week 8 } - MADRS_{baseline} }}{{MADRS_{baseline} }} \times { 1}00\%$$
Based on this equation, negative values of this outcome reflect a reduction in depression severity, a score of zero reflects no change and positive scores indicate an increase in depression severity from baseline to week eight. The MADRS ratings were performed by trained and experienced evaluators. All n = 53 completers had valid MADRS scores at week eight.
Statistical analyses
All analyses described in the following were performed in the sample of treatment completers (n = 53) at week eight using STATA version 15.1 (StataCorp 2017). To ensure that the results of the analyses were not driven by possible outliers, both the outcome variable as well as all analyzed baseline variables were tested for outliers and skewness before calculating the moderator effect sizes. We detected no outliers.
Calculating individual moderator effect sizes
By using the method described by Kraemer [
25], we first computed moderator effect sizes for the 11 preselected baseline variables. For this, we paired each patient assigned to CBASP to each patient assigned to ESC/CM. Next, for each pair of this dataset, we calculated the difference in outcome (i.e., the percentage change in MADRS scores) and the average value of each of the 11 baseline variables. Next, for obtaining moderator effect sizes, non-parametric Spearman correlations between the difference in outcome and each average were calculated together with their 95% bootstrap confidence intervals based on 100 replications. In principle, moderator effect sizes based on this method are invariant over linear transformations of the baseline variable or the outcome, and vary between −1 and + 1, with higher magnitudes indicating a stronger moderation and zero indicating the absence of a moderation effect [
25]. Baseline variables were preselected to be included in the model for complying
M* when their effect size was ≥ |0.20|. This cutoff is more rigorous than others used in previously published applications of Kraemer’s composite moderator method [
9,
26,
27], and was chosen as such in order to select as few meaningful moderator variables as possible to account for the modest sample size. We abstained from calculating and including statistical significance of interaction effects between the treatment variable and the baseline variables as a further selection criterion for a baseline variable to be used for the compilation of
M* [
28].
Model selection of the composite moderator
Next, we determined the statistical weights of those baseline variables with effect sizes ≥ |0.20| for inclusion in the composite moderator. For this, in the paired dataset, the weights of the single moderators were estimated by a multivariable regression model, in which the difference in outcome was predicted by the averages of all preselected variables. Analogous to previous applications of the composite moderator approach [
9,
27,
29], we performed a least absolute shrinkage and selection operator (lasso) regression [
30] for the multivariable model. In principal, lasso regression selects the most useful independent variables and shrinks the regression weights of the least useful variables with little predictive power or correlated with other predictors to zero, thereby removing them from the model [
30]. We chose to apply this method in order to circumvent subjective and arbitrary decisions by the researchers about which variables to remove from the model.
In addition and in line with previous applications of the composite moderator method [
9,
31,
32], for further optimizing the model’s predictive performance and avoiding overfitting, we combined lasso regression with
k-fold cross-validation [
33]. The methodological advantages of combining lasso regression with
k-fold cross-validation have been explained before [e.g., 14]. Briefly, empirically based statistical methods can sometimes utilize chance associations within a single data set, making it difficult to replicate results across various studies. To protect against the exploitation of random associations, we used the
k-fold cross-validation method. In
k-fold cross-validation, the data is randomly sampled into
k folds, whereby (
k-1) folds are used as the training dataset, and the
kth fold constitutes the validation dataset. The model is estimated within the training dataset, and its predictive performance is assessed within the held-out validation dataset [
33]. The entire procedure is repeated
k times so that each fold is used for validation once. When applied to lasso regression,
k-fold cross-validation can be used to identify the value of the tuning parameter (λ) that minimizes the estimated mean-squared prediction error (MSPE) in the validation dataset. Thus,
k-fold cross-validation enables the researcher to select a model that is more likely to have a good predictive performance in future new data, than a model that was trained and tested within the same data. Given the modest sample size and the associated need to replicate the results provided in our study, we decided to use this method in order to provide more reproducible results.
Concretely, in our analysis, for defining the optimal tuning parameter that yields the smallest MSPE, we applied 10-folds cross-validation as described by Ahrens et al. [
34] and implemented in their package
“lassopack” developed for use in STATA. Within the paired dataset, we ran the 10-folds cross-validation by using the command “
cvlasso”, which internally repeats lasso regression and finally selects the model with the optimal tuning parameter (λ
opt) that yields the smallest MSPE.
Identification and characterization of subgroups
After selecting the optimal model based on the procedure described before, weights from each of the moderators selected by this model were extracted in order to calculate the value of
M* for each patient as described by Kraemer [
34]. Thereafter, in the unpaired dataset, we performed a regression analysis predicting the outcome (i.e., percentage change in MADRS scores) from the composite moderator
M*, the treatment group, and their interaction. We furthermore computed the moderator effect size of the composite moderator
M* together with its 95% bootstrap confidence interval. We then calculated the value of
M* at which the predicted outcomes for the CBASP and ESC/CM group intersected and divided the sample into two subgroups, one below and one above this cross-point. Each of these subgroups is consequently associated with a likely more beneficial outcome for one of the two treatments compared to the other. For characterizing and comparing the two identified subgroups, we analyzed and compared relevant baseline characteristics and calculated between-group treatment effect sizes (Cohen’s
d).
Subgroup and treatment interaction effects
We next analyzed whether those patients who received their likely more beneficial treatment condition had higher response and remission rates than those who received their likely less beneficial condition. For this, we stratified the sample of completers in four clusters: 1. Patients randomized to CBASP and likely to respond better to CBASP; 2. Patients randomized to CBASP and likely to respond better to ESC/CM; 3. Patients randomized to ESC/CM and likely to respond better to ESC/CM; 4. Patients randomized to ESC/CM and likely to respond better to CBASP. In these four clusters, we compared rates of response (defined as ≥ 50.0% reduction in MADRS scores from baseline to week eight) and remission (defined as a MADRS score of ≤ 9 at week eight). We also analyzed between-cluster differences in MADRS scores at week eight as well as values of the percentage change of the MADRS scores from baseline to week eight. Finally, we examined whether those patients who did not experience a change of at least 20.0% after the acute treatment phase and who received augmentation with the other treatment condition were, in majority, those who did not receive their likely more beneficial treatment condition during the acute treatment phase.