Short-term psychodynamic psychotherapy (STPP) is an empirically supported treatment for depression (Driessen et al., Reference Driessen, Hegelmaier, Abbass, Barber, Dekker, Van and Cuijpers2015) that is frequently applied in clinical practice. Reviews have concluded that combined treatment of antidepressants and STPP is more efficacious than antidepressants alone (Driessen et al., Reference Driessen, Dekker, Peen, Van, Maina, Rosso and Cuijpers2020; Fonagy, Reference Fonagy2015). However, this effect – on average – is small at treatment completion (Driessen et al., Reference Driessen, Dekker, Peen, Van, Maina, Rosso and Cuijpers2020) and it is unclear which patients benefit specifically. Due to scarcity of treatment resources and because adding STPP to antidepressants requires a financial investment, it is of considerable clinical importance to know for whom the addition is beneficial and for whom this might not be the case.
A main reason why it is unclear which pre-treatment patient characteristics (so-called moderators) are associated with differential response to combined treatment of antidepressants and STPP, is lack of statistical power in individual clinical trials, which have sample sizes aimed at identifying intervention effects rather than moderators. Similarly, ‘conventional’ meta-analyses, which are based on study-level characteristics extracted from published trial reports, have not been able to examine moderators in this regard because the number of available studies was too small (Driessen et al., Reference Driessen, Hegelmaier, Abbass, Barber, Dekker, Van and Cuijpers2015). Moreover, such analyses would have been prone to ecological bias, such that the association between the study-level characteristics may not be representative of the true relationships in the data at the individual level.
‘Individual participant data’ (IPD) meta-analysis is an alternative technique to examine treatment effects by combining patient-level data from multiple clinical trials, which increases the statistical power to examine moderators of treatment efficacy (Lambert, Sutton, Abrams, & Jones, Reference Lambert, Sutton, Abrams and Jones2002). Furthermore, because moderators are studied at the patient level, ecological bias can be circumvented. We, therefore, conducted a systematic review and IPD meta-analysis to examine efficacy moderators of acute-phase combined treatment (antidepressants + STPP) v. antidepressants [with/without brief supportive psychotherapy (BSP)] as compared on depressive symptom measures in randomized clinical trials for adults with depression.
Methods
Design
This study is part of a systematic review and IPD meta-analysis project aimed at examining different aspects of STPP for depression efficacy. This larger project was registered (PROSPERO registration number: CRD42017056029) and its study protocol was published (Driessen et al., Reference Driessen, Abbass, Barber, Connolly Gibbons, Dekker, Fokkema and Cuijpers2018). This report complies with the Preferred Reporting Items for Systematic Review and Meta-Analyses of Individual Participant Data (PRISMA-IPD) statement (Stewart et al., Reference Stewart, Clarke, Rovers, Riley, Simmonds, Stewart and Tierney2015).
Study selection
We searched five bibliographic databases (PubMed, PsycINFO, Embase.com, Web of Science, and Cochrane's Central Register of Controlled Trials), two gray-literature databases (GLIN, UMI ProQuest), and a prospective trial register (http://www.controlled-trials.com) from inception to 19 June 2017, without applying language or date restrictions (for the full electronic search strategy see Driessen et al., Reference Driessen, Abbass, Barber, Connolly Gibbons, Dekker, Fokkema and Cuijpers2018). We also searched references of psychodynamic therapy meta-analyses and contacted psychodynamic therapy researchers. Finally, in order to identify recent studies, we searched a database of randomized depression psychotherapy trials (www.metapsy.org) from inception to 1 January 2022. This METAPSY database is developed through comprehensive literature searches in PubMed, PsycINFO, Embase.com, and the Cochrane Library and is updated annually. In each review phase (screening, eligibility, and inclusion), records were screened by two raters independently. Disagreements were resolved through consensus.
We included studies if they reported outcomes on standardized measures of depressed adult participants receiving STPP. Participants were considered depressed if they met specified criteria for major depressive disorder or another unipolar mood disorder as assessed by means of a semi-structured interview or clinicians' assessment, or if they presented an elevated score above the ‘no depression’ cut-off on a standardized measure of depression. We included studies in which STPP (a) was based on psychoanalytic theories and practices, (b) was time-limited from the onset (i.e. not a therapy that was brief only in retrospect), and (c) applied verbal techniques (e.g. therapies applying art as expression form were not considered STPP). Inclusion criteria were assessed at the study level. From the resulting set of studies, we identified randomized comparisons of combined treatment (antidepressants + STPP, provided individually in an outpatient setting) v. antidepressants (with/without BSP to control for non-specific treatment effects).
Data collection
Authors of the included studies were contacted and invited to contribute their studies' IPD. We decided a priori to request anonymized participant-level datasets, including all variables assessed before treatment start (potential moderators) and all outcome variables assessed during and after treatment. We checked whether the IPD received matched the data reported in the publication and whether outcome and moderator variables had out-of-range, invalid, or inconsistent scores (Driessen et al., Reference Driessen, Abbass, Barber, Connolly Gibbons, Dekker, Fokkema and Cuijpers2018). We listed multiple study characteristics (Driessen et al., Reference Driessen, Hegelmaier, Abbass, Barber, Dekker, Van and Cuijpers2015). We rated Cochrane risk of bias items for selection and detection bias based on information in the publications, and attrition bias based on the IPD (Higgins et al., Reference Higgins, Altman, Gøtzsche, Jüni, Moher, Oxman and Sterne2011). If information was not reported in the publications, we requested this from the authors. We did not rate performance bias as it is considered impossible to blind participants and therapists to treatment in psychotherapy research. Selective reporting bias was considered not applicable, as we requested all outcome measures assessed.
Measures
Depression symptom level at post-treatment constituted the pre-specified primary outcome measure. Follow-up was the additional time point. For each trial, we identified the primary continuous depression outcome and post-treatment/follow-up end points as defined by the authors. Because different depression measures were used, we standardized outcomes by converting depression scores into z-scores within end point within each study. We conducted sensitivity analyses using unstandardized 17-item Hamilton Depression Rating Scale (HAMD) scores as outcome.
Potential moderators were pre-specified as all pre-treatment demographic, clinical, and psychological participant characteristics that were assessed in more than one study. Based on definitions provided in data dictionaries and/or publications we decided whether variables could be pooled across studies. If constructs were operationalized differently in individual studies, they were standardized too, by converting scores into z-scores (within study) for continuous variables or by recoding variables into similar categories for categorical variables (see online Supplementary Table ST1 for an overview). Standardization occurred before data-analysis started and no changes were made afterward.
Data analysis
We conducted one-stage IPD meta-analyses using mixed-effects models with a three-level structure (study, participant, repeated measures) and restricted maximum likelihood estimation. Analyses were based on intent-to-treat samples. Follow-up data were excluded from post-treatment analyses, due to the fact that additional help-seeking could not be controlled. We estimated heterogeneity with the I 2 statistic.
For each potential moderator, we estimated a model including time and the moderator, a time-by-treatment interaction, and a time-by-moderator-by-treatment three-way interaction. This approach is recommended by Twisk et al. (Reference Twisk, Bosman, Hoekstra, Rijnhart, Welten and Heymans2018) because it adequately accounts for baseline depression values and has favorable properties concerning missing data (i.e. participants with a baseline value but missing post-treatment and/or follow-up assessments are still included in the analyses). Time was treated as a categorical variable to facilitate treatment comparison at the two end points and continuous moderator variables were grand-mean centered. To account for the clustering of participants within studies, we estimated a random intercept with respect to study. We also estimated a random intercept with respect to participants, and fixed slopes. Based on a −2-log likelihood change evaluation in a basic model (time and time-by-treatment interaction), we added a random slope for the time-by-treatment interaction at the study level in the sensitivity analyses with unstandardized 17-item HAMD scores.
Using this approach, regression coefficients of the time-by-treatment interactions at post-treatment and follow-up represent the treatment comparisons at these assessment moments (Twisk et al., Reference Twisk, Bosman, Hoekstra, Rijnhart, Welten and Heymans2018). For analyses with z-scores as outcome measure, these regression coefficients are standardized mean differences, which can be interpreted in the same way as a Cohen's d effect size. For the analyses with unstandardized 17-item HAMD scores as outcome measure, regression coefficients are mean differences. For categorical moderator variables, we varied the reference category to obtain treatment comparisons in each moderator category. The three-way interaction's regression coefficient reflects the difference between time-by-treatment interactions for different moderator levels. A Bonferroni correction for multiple testing would yield an α level of 0.00125 but would also increase the risk of type-II errors given the low power of three-way interactions. We thus considered a p value of <0.05 for the three-way interaction indicating a statistically significant moderator effect but interpreted p values >0.00125 with caution. We visually inspected histograms of (standardized) residuals after each analysis. The normality assumption was always met.
Next, we conducted three pre-specified sensitivity analyses. First, we repeated the analyses in studies that administrated the 17-item HAMD, using unstandardized scores as outcome measure. Second, to examine the impact of risk of bias in the primary studies, we added risk of bias items as dichotomous covariates to the mixed-effects models (Higgins, Thompson, & Spiegelhalter, Reference Higgins, Thompson and Spiegelhalter2009). Third, we conducted analyses including only studies that scored negative on all risk of bias criteria assessed. Finally, we modeled all variables with significant three-way interactions simultaneously.
We then conducted two post-hoc sensitivity analyses. To examine the representativeness of the findings of those eligible for STPP for depression, we examined whether significant moderator effects were still present when excluding studies that only included participants with specific comorbidities. Second, we examined the potential subgroup difference of studies with and without BSP in the comparison condition (Driessen et al., Reference Driessen, Dekker, Peen, Van, Maina, Rosso and Cuijpers2020) by examining significant moderator effects at follow-up in both subgroups. Mixed-effects model analyses were performed with MLwiN (version 2.26). Data availability bias was assessed by examining differences in study characteristics and effect sizes between studies that contributed IPD and studies that did not (Driessen et al., Reference Driessen, Abbass, Barber, Connolly Gibbons, Dekker, Fokkema and Cuijpers2018). We examined publication bias by assessing funnel plot asymmetry for meta-analyses including ⩾10 studies (Sterne et al., Reference Sterne, Sutton, Ioannidis, Terrin, Jones, Lau and Higgins2011).
We also conducted pre-specified exploratory machine learning analyses with the generalized linear mixed-effects model (GLMM) tree algorithm (Fokkema, Smits, Zeileis, Hothorn, & Kelderman, Reference Fokkema, Smits, Zeileis, Hothorn and Kelderman2018) and cross-validation techniques (Kim, Reference Kim2009), which are described in detail in Appendix 1 (online Supplementary material). These analyses offer several advantages over traditional mixed-effects models: they allow for detecting non-linear and higher-order interaction effects in nested data, involve less stringent assumptions about the distribution of the data, and result in decision trees that may be easier to interpret and apply in clinical decision making.
Results
Literature search results are described in Fig. 1. Seven studies were identified to meet the inclusion criteria for this work (Burnand, Andreoli, Kolatte, Venturini, & Rosset, Reference Burnand, Andreoli, Kolatte, Venturini and Rosset2002; de Jonghe, Kool, van Aalst, Dekker, & Peen, Reference de Jonghe, Kool, Van Aalst, Dekker and Peen2001; López Rodríguez, López Butrón, Vargas Terrez, & Villamil Salcedo, Reference López Rodríguez, López Butrón, Vargas Terrez and Villamil Salcedo2004; Maina, Rosso, Crespi, & Bogetto, Reference Maina, Rosso, Crespi and Bogetto2007; Maina, Rosso, Rigardetto, Chiadò Piat, & Bogetto, Reference Maina, Rosso, Rigardetto, Chiadò Piat and Bogetto2010; Martini, Rosso, Chiodelli, De Cori, & Maina, Reference Martini, Rosso, Chiodelli, De Cori and Maina2011; Vitriol, Ballesteros, Florenzano, Weil, & Benadorf, Reference Vitriol, Ballesteros, Florenzano, Weil and Benadof2009). IPD were obtained from all these studies and totaled 482 participants (combined: n = 238, 49%; antidepressants: n = 244, 51%). Data integrity checks identified inconsistent items for one study and discrepancies between the dataset received and published article for three studies. These were resolved with the authors.
Fig. 1. PRISMA-IPD flow diagram. © Reproduced with permission from the PRISMA-IPD Group, which encourages sharing and reuse for non-commercial purposes.
Study characteristics are described in Table 1. Depression inclusion criteria typically consisted of a DSM (Burnand et al., Reference Burnand, Andreoli, Kolatte, Venturini and Rosset2002; de Jonghe et al., Reference de Jonghe, Kool, Van Aalst, Dekker and Peen2001; López Rodríguez et al., Reference López Rodríguez, López Butrón, Vargas Terrez and Villamil Salcedo2004; Maina et al., Reference Maina, Rosso, Crespi and Bogetto2007, Reference Maina, Rosso, Rigardetto, Chiadò Piat and Bogetto2010) or ICD-10 (Vitriol et al., Reference Vitriol, Ballesteros, Florenzano, Weil and Benadof2009) depression diagnosis with an elevated HAMD score (Burnand et al., Reference Burnand, Andreoli, Kolatte, Venturini and Rosset2002; de Jonghe et al., Reference de Jonghe, Kool, Van Aalst, Dekker and Peen2001; Maina et al., Reference Maina, Rosso, Crespi and Bogetto2007, Reference Maina, Rosso, Rigardetto, Chiadò Piat and Bogetto2010; Vitriol et al., Reference Vitriol, Ballesteros, Florenzano, Weil and Benadof2009), though in one study (Martini et al., Reference Martini, Rosso, Chiodelli, De Cori and Maina2011) HAMD score constituted the sole depression inclusion criterion. Four studies (4/7, 57%) included adults with depression in general (Burnand et al., Reference Burnand, Andreoli, Kolatte, Venturini and Rosset2002; de Jonghe et al., Reference de Jonghe, Kool, Van Aalst, Dekker and Peen2001; López Rodríguez et al., Reference López Rodríguez, López Butrón, Vargas Terrez and Villamil Salcedo2004; Maina et al., Reference Maina, Rosso, Crespi and Bogetto2007). Two studies (2/7, 29%) only included participants with specific anxiety disorder comorbidities (Maina et al., Reference Maina, Rosso, Rigardetto, Chiadò Piat and Bogetto2010; Martini et al., Reference Martini, Rosso, Chiodelli, De Cori and Maina2011) and one (1/7, 14%) included women with childhood trauma (Vitriol et al., Reference Vitriol, Ballesteros, Florenzano, Weil and Benadof2009). The antidepressant types and dosages in these three studies did not differ from the standard first-line treatment for depression. STPP was based on various models (Andreoli, Reference Andreoli, De Clercq, Andreoli, Lamarre and Foster1999; Bellak, Reference Bellak1993, Reference Bellak1994; de Jonghe, Rijnierse, & Janssen, Reference de Jonghe, Rijnierse and Janssen1994; Malan, Reference Malan1963, Reference Malan1976; Safran & Muran, Reference Safran and Muran2000; Vitriol, Reference Vitriol, Florenzano, Weil, Carvajal and Cruz2005). Four studies (4/7, 57%) included BSP in the comparison condition (Burnand et al., Reference Burnand, Andreoli, Kolatte, Venturini and Rosset2002; Maina et al., Reference Maina, Rosso, Crespi and Bogetto2007; Martini et al., Reference Martini, Rosso, Chiodelli, De Cori and Maina2011; Vitriol et al., Reference Vitriol, Ballesteros, Florenzano, Weil and Benadof2009). Six studies (6/7, 86%) used the 17-item HAMD as primary outcome measure (Burnand et al., Reference Burnand, Andreoli, Kolatte, Venturini and Rosset2002; De Jonghe et al., Reference de Jonghe, Kool, Van Aalst, Dekker and Peen2001; Maina et al., Reference Maina, Rosso, Crespi and Bogetto2007, Reference Maina, Rosso, Rigardetto, Chiadò Piat and Bogetto2010; Martini et al., Reference Martini, Rosso, Chiodelli, De Cori and Maina2011; Vitriol et al., Reference Vitriol, Ballesteros, Florenzano, Weil and Benadof2009). Four studies (4/7, 57%) scored negative on all four risk of bias criteria assessed (de Jonghe et al., Reference de Jonghe, Kool, Van Aalst, Dekker and Peen2001; Maina et al., Reference Maina, Rosso, Crespi and Bogetto2007, Reference Maina, Rosso, Rigardetto, Chiadò Piat and Bogetto2010; Martini et al., Reference Martini, Rosso, Chiodelli, De Cori and Maina2011; Vitriol et al., Reference Vitriol, Ballesteros, Florenzano, Weil and Benadof2009).
Table 1. Characteristics of the included studies
AD, anxiety disorder; All Con, allocation concealment; AP, second-generation antipsychotic; Blind, blinding of HAMD assessment; BSP, brief supportive psychotherapy in the comparison condition; HAMD-17, 17-item Hamilton Depression Rating Scale; ITT, complete outcome data (full intention-to-treat IPD available); MS, mood stabilizer; RIMA, reversible inhibitor of monoamine-oxidase A; Seq Gen, random sequence generation; SNRI, serotonin noradrenaline reuptake inhibitor; SSRI, selective serotonin reuptake inhibitor; STPP, short-term psychodynamic psychotherapy; TCA, tricyclic antidepressant.
Note. –, data not available.
a Numbers represent successive steps in the pharmacotherapy protocol.
Treatment comparisons at post-treatment and follow-up for the different moderator levels are reported in Table 2 (main analyses) and Table ST2 (sensitivity analyses; online Supplementary material). Relevant results for individual studies are reported in Table 3. At post-treatment, baseline HAMD, education, and episode duration were found to moderate treatment effects (Table 2). The effect of adding STPP to antidepressants was larger for participants with high rather than low baseline HAMD scores [B = −0.49, 95% confidence interval (CI) −0.61 to −0.37, p < 0.0001], for participants with ⩽8 rather than 13–15 education years (B = −0.66, 95% CI −1.05 to −0.27, p < 0.0009), and for participants with a depressive episode duration of >2 years rather than <1 year (B = −0.68, 95% CI −1.31 to −0.05, p = 0.03) and than 1–2 years (B = −0.86, 95% CI −1.66 to −0.06, p = 0.04). No heterogeneity was present in these analyses (I 2 = 0). For baseline HAMD and episode duration, result patterns were replicated across all pre-specified sensitivity analyses (online Supplementary Table ST2). Regarding education, the three-way interaction was no longer statistically significant when considering low risk of bias studies only, nor when modeling all significant moderators simultaneously (p = 0.27). The latter model did show significant three-way interactions for baseline HAMD (p < 0.001) and episode duration (<1 v. >2 years: p = 0.003, 1–2 v. >2 years: p < 0.001). In the post-hoc sensitivity analysis excluding studies that only included participants with specific comorbidities (Maina et al., Reference Maina, Rosso, Rigardetto, Chiadò Piat and Bogetto2010; Martini et al., Reference Martini, Rosso, Chiodelli, De Cori and Maina2011; Vitriol et al., Reference Vitriol, Ballesteros, Florenzano, Weil and Benadof2009), significant moderator effects were also found for baseline HAMD (p < 0.0001) and episode duration (<1 v. >2 years: p = 0.046), although the 1–2 v. >2 episode duration year contrast was no longer statistically significant (p = 0.0507).
Table 2. Cohen's d effect sizes of combined treatment of antidepressants and STPP v. antidepressants at post-treatment and follow-up for the different moderator levels
CGI-S, Clinical Global Impression subscale ‘Severity of Illness’; GAF, Global Assessment of Functioning Scale; HAMD, Hamilton Depression Rating Scale; STPP, short-term psychodynamic psychotherapy.
Note. Negative signs indicate lower depressive symptom levels (i.e. better outcomes) in the combined antidepressants and STPP treatment condition than in the comparison condition. Numbers printed in bold indicate statistically significant time-by-moderator-by-treatment three-way interactions (p < 0.05). For categorical variables, this indicates a significant difference between the treatment effect in this category and another (see below). For continuous variables, the first effect size (‘Average’) reflects the treatment comparison for participants with baseline scores at the average of the study sample, while the second number (‘Per … increase’) reflects the additional effect for each unit increase in baseline score.
a ⩽8 years v. 13–15 years (post-treatment): B = −0.66, 95% CI −1.05 to −0.27, p < 0.001; ⩽8 years v. 13–15 years (follow-up): B = −0.54, 95% CI −0.91 to −0.17, p = 0.005; 13–15 years v. >15 years (follow-up): B = −0.53, 95% CI −1.02 to −0.04, p = 0.03.
b <1 year v. >2 years: B = −0.68, 95% CI −1.31 to −0.05, p = 0.03; 1–2 years v. >2 years: B = −0.86, 95% CI −1.66 to −0.06, p = 0.04.
c B = −0.48, 95% CI −0.87 to −0.09, p = 0.01.
Table 3. Results of individual studies
–, data not available; HAMD, Hamilton Depression Rating Scale.
Note. Negative signs indicate lower depressive symptom levels (i.e. better outcomes) in the combined antidepressants and STPP treatment condition than in the comparison condition.
At follow-up, baseline HAMD, education, anxiety disorder comorbidity, and baseline anxiety symptom level were found to moderate treatment effects. The effect of adding STPP to antidepressants was larger for participants with high rather than low baseline HAMD scores (B = −0.60, 95% CI −0.74 to −0.46, p < 0.0001), for participants with ⩽8 (B = −0.54, 95% CI −0.91 to −0.17, p = 0.005) and >15 (B = −0.53, 95% CI −1.02 to −0.04, p = 0.03) rather than 13–15 education years, for participants without rather than with a comorbid anxiety disorder (B = −0.48, 95% CI −0.87 to −0.09, p = 0.01), and for participants with high rather than low baseline anxiety symptom levels (B = −0.23, 95% CI −0.39 to −0.08, p = 0.004). No heterogeneity was present in these analyses (I 2 = 0). For baseline HAMD and anxiety disorder comorbidity, result patterns were replicated across all pre-specified sensitivity analyses. Regarding education, statistically significant three-way interactions were apparent in all pre-specified sensitivity analyses, though the specific contrasts that were significant differed between the analyses. Regarding baseline anxiety symptom level, the three-way interaction was no longer statistically significant when modeling all significant moderators simultaneously (p = 0.37). This final model did include significant three-way interactions for baseline HAMD (p < 0.001), anxiety disorder comorbidity (p = 0.02), and education (⩽8 v. 9–12 years: p = 0.03; 9–12 v. >15 years p = 0.03). The significant moderator effect for baseline HAMD was also observed in both post-hoc sensitivity analyses (p < 0.0001). The effect for anxiety disorder comorbidity was no longer statistically significant when excluding the three studies that only included participants with specific comorbidities (p = 0.75), nor in the subgroup of studies with BSP in the comparison condition (p = 0.12, no moderator effect could be estimated in the subgroup of studies without BSP).
GLMM tree analyses' results are presented in Appendix 1 (online Supplementary material) and also identified baseline HAMD, episode duration, and anxiety disorder comorbidity as moderators. Correlations between predicted and observed outcomes for the trees were indicative of small to medium-to-large effects.
Discussion
We conducted a systematic review and IPD meta-analysis to examine which patients benefit from adding STPP to antidepressants in the treatment of depression. Across the seven studies that were identified by a thorough literature search, we found baseline HAMD and episode duration moderating post-treatment efficacy and baseline HAMD and anxiety disorder comorbidity moderating efficacy at follow-up. Adding STPP to antidepressants was more efficacious for participants with high baseline depression levels, for participants with episode durations of >2 years, and for participants without a comorbid anxiety disorder.
Regarding baseline depression, our findings are in line with work reporting treatment effects to increase with symptom severity for antidepressants (Stone, Kalaria, Richardville, & Miller, Reference Stone, Kalaria, Richardville and Miller2018), psychotherapy (Driessen, Cuijpers, Hollon, & Dekker, Reference Driessen, Cuijpers, Hollon and Dekker2010), and the addition of the cognitive behavioral analysis system of psychotherapy to antidepressants (Furukawa et al., Reference Furukawa, Efthimiou, Weitz, Cipriani, Keller, Kocsis and Schramm2018). Thus, baseline severity appears to moderate depression treatment efficacy in general rather than applying to STPP specifically. These findings have been taken to imply that relative to low-severity patients, high-severity patients are in more need of treatments with specific effects in order to get well (Driessen et al., Reference Driessen, Cuijpers, Hollon and Dekker2010).
Concerning episode duration, our findings are in line with studies demonstrating the effects of psychodynamic therapy for patients with treatment resistant depression (Fonagy et al., Reference Fonagy, Rost, Carlyle, McPherson, Thomas, Pasco Fearon and Taylor2015; Town, Abbass, Stride, & Bernier, Reference Town, Abbass, Stride and Bernier2017) who typically suffer from long-duration episodes. Episode duration has also been observed to moderate the effect of antidepressants combined with STPP v. antidepressants combined with cognitive behavioral therapy (CBT; Driessen et al., Reference Driessen, Smits, Dekker, Peen, Don, Kool and Van2016), such that combined treatment with STPP was more effective for patients with episode durations of ⩾1 year. It has been speculated in this regard that individuals with longer episode durations have depressive symptoms that are more influenced by their personality structure resulting in more complex working alliances and transference feelings; psychodynamic therapists are trained to elaborate on these therapeutic relational aspects if necessary (Driessen et al., Reference Driessen, Smits, Dekker, Peen, Don, Kool and Van2016). However, the strength of evidence for episode duration as a moderator is limited by the small number of participants reporting chronic episodes.
Concerning anxiety disorder comorbidity, we also found some indication for a moderating effect at post-treatment, but this was only significant in one sensitivity analysis. The strength of evidence for anxiety disorder comorbidity as a moderator at follow-up is limited by the lack of variability in three of the four studies assessing this variable, suggesting that this finding might be driven by between-study effects, and the p values exceeding the Bonferroni correction. Finally, education was associated with treatment efficacy in the mixed-effects models. The specific contrasts between the four education levels, however, reached statistical significance in some analyses but not in others, and education did not appear as a moderator in the GLMM trees.
Strengths and limitations
This study has a number of strengths. First, to the best of our knowledge, this is the first study that examines moderators across randomized clinical trials that assess the efficacy of adding STPP to antidepressants. Second, this meta-analysis did not suffer from data availability bias as IPD were obtained for all included studies. Risk of selection bias in the primary studies was also low. Third, the included studies shared similarities in depression inclusion criteria and primary outcome measure. Fourth, we used mixed-effects models as well as a novel tree-based machine learning technique to examine moderators at both post-treatment and follow-up. This allowed for identifying both short- and long-term moderator effects, as well as potential higher-order interactions. The GLMM trees (Figs. SA1 and SA2, Appendix 1, online Supplementary material) may be easier to apply in clinical practice. Fifth, using IPD meta-analytic methods increased the statistical power to examine moderators (Lambert et al., Reference Lambert, Sutton, Abrams and Jones2002).
This study also has a number of limitations. First, even though IPD could be obtained for all studies, the total number of participants included in this meta-analysis is modest. Relatively few studies examined the efficacy of adding STPP to antidepressants and all were conducted more than 10 years ago. We think this reflects that psychodynamic therapy in general has been studied less extensively than other forms of psychotherapy for depression, such as CBT (e.g. Cuijpers et al., Reference Cuijpers, Quero, Noma, Ciharova, Miguel, Karyotaki and Furukawa2021). Although this study was adequately powered to identify certain moderators, it might have lacked statistical power to identify relatively small moderator relationships or higher-order interactions. Related, not all moderator variables were assessed in all studies. Thus, the individual moderator models can relate to different subgroups of studies. Second, not all studies were free from detection and attrition bias, though the moderator findings discussed previously appeared robust against controls for these risks of bias. The number of studies included in this meta-analysis was too small to assess publication bias. Third, although the studies shared similarities, they also differed with regard to the STPP model used, the antidepressant type, and follow-up length, for instance. Regardless of these differences, moderator effects could be identified in the combined studies' data. Fourth, we were not able to examine every potential moderator variable of interest (e.g. personality structure), because they were not assessed in the primary studies. Fifth, and most important, these findings are observational and need validation in independent samples. Although we applied cross validation to estimate predictive accuracy for the GLMM trees, further replication of our findings would strengthen the basis for their use in guiding treatment selection.
Clinical and research implications
The findings of this study suggest that adding STPP to antidepressants might be particularly efficacious for individuals with relatively high baseline HAMD scores, for individuals with episode durations of >2 years (at post-treatment), and for those without a comorbid anxiety disorder (at follow-up). For individuals with relatively low baseline HAMD scores, for individuals with episode durations of ⩽2 years, and for individuals with a comorbid anxiety disorder, adding STPP might not result in superior treatment effects. However, the findings of this study cannot be taken to imply that antidepressants only (or combined with BSP) should be considered the first-choice treatment for these latter individuals, as this study does not speak to the comparative efficacy of antidepressants v. other depression treatments (e.g. CBT). In addition, this study does not speak to which patients benefit from adding antidepressants to STPP, a question that is as relevant, but is not possible to address with an IPD meta-analysis yet due to a lack of randomized controlled trials.
Given the clinical importance of the research question and the limitations of the current study in terms of sample size and moderator variables assessed, further study of which patients benefit from adding STPP to antidepressants in the treatment of depression is warranted. Specifically, the field would benefit from additional large-scale rigorously conducted randomized clinical trials that include BSP in the comparison condition. Comparisons of combined treatment with STPP only are also needed to address the (reversed) question of which patients benefit from adding antidepressants to STPP. Future trials should assess a range of potential moderators, including baseline depression severity, episode duration, and anxiety disorder comorbidity, but also education level and personality structure. Data from such future trials would provide an important validation set for the current findings. If validated, the findings of this study would suggest that depression severity, episode duration, and anxiety disorder comorbidity are important factors to consider when adding STPP to antidepressants. Such knowledge can be used to facilitate evidence-based personalized treatment selection for depression and making more efficient use of existing depression treatment resources.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0033291722003270.
Data
The collective de-identified individual participant database developed for this study, as well as a data dictionary and relevant related documents (e.g. study protocol) are available for use by other researchers with publication of this manuscript. Requests can be made with the corresponding author (ellen.driessen@ru.nl). Access (with limited investigator support) will be granted after approval of a study proposal by all authors and a signed data access agreement.
Author contributions
ED, JJMD, HLV, JWRT, and PC designed the study and wrote the protocol. ED, FJW, and PC conducted literature searches. JJMD, JP, HLV, GM, GR, SR, FC, VGV, AA, YB, JLR, and VVS collected the individual participant data. ED and MF conducted the statistical analyses. ED wrote the first draft of the manuscript. All authors contributed to and have approved the final manuscript.
Financial support
Recruitment of individual participant data for this work was supported by a Fund for Psychoanalytic Research of the American Psychoanalytic Association. This study was further financed by a research grant from the Dutch Psychoanalytic Funds and by the Netherlands Organisation of Scientific Research (NWO): 016.Veni.195.215 6806 to ED. The funders had no role in the study design; collection, analysis, or interpretation of data; writing the manuscript; or the decision to submit the manuscript for publication.
Conflict of interest
ED reports grants from the Dutch Psychoanalytic Funds and Netherlands Organisation of Scientific Research (NWO) during the conduct of the study. MF, JJMD, JP, HLV, GM, GR, SR, FC, VGV, AA, YB, JLR, VVS, JWRT, FJW, and PC have no conflicts of interest to declare.
Ethical standards
Institutional Review Board (IRB) approval was not required for this project, because we worked with anonymized data from treatment studies that had already been completed. IRB approval might have been required for the investigators to share their IPD depending on their institution's policies. It was the responsibility of the investigators to obtain IRB approval if their institution's policies required them to do so. By signing the data sharing agreement, the authors who shared their IPD declared that those data were collected and transferred according to all applicable local and international laws and regulations, including but not limited to local IRB approval.
Open access
