Psychologic Treatment of Depression Compared With Pharmacotherapy and Combined Treatment in Primary Care

A Network Meta-Analysis

Pim Cuijpers, PhD; Matthijs Oud, MA; Eirini Karyotaki, PhD; Hisashi Noma, PhD; Soledad Quero, PhD; Andrea Cipriani, MD, PhD; Bruce Arroll, PhD; Toshi A. Furukawa, MD, PhD


Ann Fam Med. 2021;19(3):262-270. 

In This Article


Identification and Selection of Studies

The protocol for this meta-analysis was registered at the Open Science Foundation.[22] We identified studies via an existing database of randomized trials on psychotherapy for depression.[23] The database is continuously updated, and every year in January searches for the prior year are conducted (from 1966 to January 1, 2019). Four bibliographic databases (PubMed, PsycInfo, Embase, Cochrane Library) were searched by combining the index and text words of depression and psychotherapy, with filters for randomized controlled trials (see Supplemental Appendix 1,, for full search string for PubMed). After importing the references in Endnote and removing duplicates, 2 independent researchers (E.K., P.C.) screened all records and full texts. Discrepancies were resolved via discussion.

We included studies in which a psychologic treatment for adult patients (aged ≥18 years) with depression who were recruited from primary care was compared with antidepressant medication, combined treatment, care as usual (care usually delivered by GPs), waitlist, or pill placebo. Depression could be established with a diagnostic interview or a score greater than a cutoff point on a self-report measure. We included any type of psychotherapy.[24,5] Therapy could be applied in individual, group, telephone-supported, or guided self-help (Internet-based or not) format because these formats have been found to have comparable effects across different settings.[26] Unguided interventions without human support were excluded because these have been found to be less effective.[26] Comorbid mental or somatic disorders were not excluded.[27] We included studies published in English, Spanish, Dutch, or German.

When a study contained ≥2 arms to be included in the same node (eg, when a study compared 2 types of psychotherapy with 1 pharmacotherapy condition), we considered them as separate comparisons. These were subdivided appropriately to avoid double counting.

Risk of Bias and Data Extraction

Risk of bias (RoB) was assessed using the following 4 criteria of the Cochrane tool:[28] adequate generation of allocation sequence, concealment of allocation to conditions, prevention of knowledge of the allocated intervention (masking of assessors), and dealing with incomplete outcome data (this was assessed as positive when intention-to-treat analyses were conducted). Items were rated by 2 independent assessors (P.C., E.K.), and disagreements were resolved via discussion. Studies were judged as low RoB when all 4 items were rated as positive.

Pharmacotherapy studies were assessed regarding the use of therapeutic dose and titration schedule (ie, therapeutic dose achieved within 3 weeks). Pharmacotherapy was deemed adequate if both criteria were met. Psychotherapy studies were assessed on (1) use of a treatment manual, (2) use of specially trained therapists, and (3) verification of treatment integrity.[29,30] We also coded participant and intervention characteristics (Supplemental Appendix 2,

Outcome Measures

Treatment response (50% decrease in depressive symptomatology) was selected as the primary outcome. When not reported, we imputed response rate using a validated method using mean depression score at baseline and mean, SD, and number of patients at post test.[31] Patients randomized but not included in the analyses of responders were assumed to be nonresponders and were included in the analyses according to the intention-to-treat principle. The time point for the primary outcome was the end of psychotherapy. When multiple scales were used, we selected a single instrument using an algorithm prioritizing clinician-rated instruments over self-rated instruments and according to how often the instruments (Hamilton Depression Rating Scale, Montgomery-Åsberg Depression Rating Scale, another clinician-rated instrument, Beck Depression Inventory I or II, another self-rated instrument) are used in depression trials.

Remission was defined as a depression score less than a specific cut-off on a validated rating scale. The standardized mean difference (SMD) between each of the contrasts was based on mean, SD, and number of patients for the conditions at post test. Acceptability of treatment was operationalized as study dropout for any reason.

Pairwise Meta-analyses

We conducted pairwise meta-analyses for all comparisons using a random effects model. To quantify heterogeneity, we calculated the I2 statistic using the non-central χ2-based approach within the heterogi module for Stata (StataCorp LLC).[32] We tested for small-study effects with the Egger test.[33]

Network Meta-analyses

Comparative effectiveness was evaluated using NMA methodology. First, we summarized the geometry of the network of evidence using network plots for response.[34] We then conducted contrast-based NMA to assess comparative efficacy and acceptability.[35] Random effect models were used in all analyses.[36] Relative risks (RRs) and SMDs were reported with 95% CIs. The ranking of treatments was estimated with the surface under the cumulative ranking curve, based on the estimated multivariate random effects models.[34]

The transitivity assumption was examined via a table of study characteristics. We verified if potential effect modifiers were similarly distributed across the comparisons in the network. We checked the consistency of the network using the local test of inconsistency (with the loop-specific approach, which estimates in each triangular and quadrangular loop whether the direct and indirect effects are consistent)[37] and the global test of inconsistency (the design-by-treatment interaction test).[38] Potential influences of small-study effects were examined with metaregression analyses, adjusting the study-specific variance as a covariate.[39]

Heterogeneity, Metaregression, and Sensitivity Analyses

We evaluated the heterogeneity in the network with τ2 compared with empirically derived values.[40,41] To explore possible sources of heterogeneity, we conducted a multivariate metaregression analysis with the characteristics that were also used to examine transitivity. We conducted sensitivity analyses in which we limited the analysis to (1) studies with low RoB, (2) studies of cognitive behavior therapy (CBT) alone, or (3) studies with pill placebo excluded (in these studies, patients receiving drugs are blinded, and patients in psychotherapy are not).[42]

The main analyses were conducted with Stata/SE 14.2 for Mac (StataCorp LLC). The metaregression analyses examining small-sample bias were conducted with OpenBUGS-3.2.3 (OpenBUGS Foundation). In addition, the main analyses examining the effect sizes of all comparisons for response and acceptability were conducted independently by one of the authors (H.N.) with the Bayesian framework using the gemtc package in R (R Foundation). The analyses were conducted in November 2019.