Effectiveness of cognitive–behavioural therapies of varying complexity in
reducing depression in adults: a systematic review and network meta-analysis
Abstract
Background
Cognitive–behavioural
therapy (CBT) is frequently used as an umbrella term to include a variety of
psychological interventions. It remains unclear whether more complex CBT
contributes to greater depression reduction.
Aims
To (a) compare
the effectiveness of core, complex and ultra-complex CBT against other
psychological interventions, medication, treatment-as-usual and no treatment in
reducing depression at post-treatment and in the long term and (b) explore
important factors that could moderate the effectiveness of these interventions.
Method
MEDLINE,
PsycInfo, Embase, Web of Science and the Cochrane Register of Controlled Trials
were searched to November 2021. Only randomised controlled trials were eligible
for the subsequent network meta-analysis.
Results
We included 107
studies based on 15 248 participants. Core (s.m.d. = −1.14, 95% credible
interval (CrI) −1.72 to −0.55 [m.d. = −8.44]), complex (s.m.d. = −1.24, 95% CrI
−1.85 to −0.64 [m.d. = −9.18]) and ultra-complex CBT (s.m.d. = −1.45, 95% CrI
−1.88 to −1.02 [m.d. = −10.73]) were all significant in reducing depression up
to 6 months from treatment onset. The significant benefits of the ultra-complex
(s.m.d. = −1.09, 95% CrI −1.61 to −0.56 [m.d. = −8.07]) and complex CBT (s.m.d.
= −0.73, 95% CrI −1.36 to −0.11 [m.d. = −5.40]) extended beyond 6 months.
Ultra-complex CBT was most effective in individuals presenting comorbid mental
health problems and when delivered by non-mental health specialists.
Ultra-complex and complex CBT were more effective for people younger than 59
years.
Conclusions
For people
without comorbid conditions healthcare and policy organisations should invest
in core CBT. For people <59 years of age with comorbid conditions
investments should focus on ultra-complex and complex CBT delivered without the
help of mental health professionals.
Major depressive disorder (MDD) is the most common mental
health condition, with more than 264 million people being affected worldwide. It
has a negative impact on people's quality of life and is very costly for health
and care systems. Cognitive–behavioural therapy (CBT) is an evidence-based
psychological intervention that is widely used for treating MDD. There are
large variations in terms of the components delivered as part of CBT protocols.
The two main elements of CBT for depression are: (a) behavioural activation
aiming to understand the potentially reciprocal relationships between negative
mood states and behaviours and (b) cognitive restructuring aiming to
increase behaviours associated with positive moods and identify, critically
evaluate and challenge maladaptive automatic thoughts. Recent evidence
examining the effectiveness of each of these components alone or in combination
suggested that they were equally effective in reducing depression in adults
compared with treatment as usual or no treatment.
Social skills training (including non-verbal and
communication skills as well as assertiveness training), relaxation techniques and
psychoeducation may be used to supplement the effectiveness of core CBT.
Techniques such as problem-solving skills, self-management skills and
relapse prevention may also be used, most likely because patients with
depression may present with multiple comorbidities, including anxieties, cancer and
diabetes. It is standard practice for studies focusing on psychological
treatments for depression to use core, complex containing behavioural
activation and/or cognitive restructuring with either psychoeducation, skills
training modules or relaxation techniques, and/or ultra-complex CBT protocols that
include core CBT with at least two or more additional therapeutic components as
mentioned above.
To date no study has examined the differential effectiveness
of each of these CBT protocols in reducing depression in adult populations. For
example, ample is the evidence garnered from several meta-analyses
demonstrating the effectiveness of CBT in significantly reducing symptoms of
depression in comparison with no intervention or treatment as usual (TAU). Despite
this, evidence on differences in the effectiveness as well as scope (e.g.
subgroups of people with MDD) of the core and multicomponent CBT protocols at
post-treatment and in the long term is lacking. There are also uncertainties
whether specific characteristics of participants, intervention, therapists or
context can influence the effectiveness of core and multicomponent CBT.
Understanding whether multicomponent CBT protocols are more effective than core
CBT, for which patient groups and how they should be delivered has important
practice and policy implications. These include improved access for underserved
patient groups and reducing healthcare inequalities, time and training
requirements for therapists and overall healthcare service delivery costs.
To address this important gap in the literature, this
systematic review and network meta-analysis aimed to (a) comparatively assess
the effectiveness of core, complex and ultra-complex CBT protocols in reducing
symptoms of depression in adults with depression at post-treatment and in the
long term and (b) examine moderators of these protocols, including
characteristics of participants, interventions, therapists and context.
Method
Eligibility criteria
Studies involving participants aged 17 years and older with
depression were eligible. Depression could have been verified through the use
of either validated self-report measures or clinical interviews. Studies that
recruited participants with comorbid mental health problems such as anxiety
were also included. However, studies focusing on COVID-19-related depression
were excluded.
We focused on studies involving core CBT protocols using
either behavioural activation, or cognitive restructuring or both as the
primary treatment components and based on any type of delivery mode, including
face-to-face or online individual/group sessions. Studies that incorporated
additional cognitive–behavioural therapeutic components to the two core
components above, including problem-solving and self-management techniques,
relaxation and social skills training, relapse prevention and psychoeducation,
were also included, if they were based on behavioural activation and/or
cognitive restructuring modules. CBT protocols were grouped into three discrete
categories according to the complexity of the included components: core CBT
(comprising behavioural activation and cognitive restructuring), complex CBT
(core CBT plus one additional component that included psychoeducation or
training in particular skills, including social skills training and relaxation
techniques) and ultra-complex CBT (core CBT plus two or more additional components
that included problem-solving skills, self-management skills, relaxation
techniques, psychoeducation and/or relapse prevention; interventions are
defined in Table 1 and therapeutic components of the included studies
are described in supplementary Appendix 2, available at https://doi.org/10.1192/bjp.2022.35).
Studies comparing the effectiveness of the core components of CBT with each
other (e.g. behavioural activation versus cognitive restructuring) or comparing
the effectiveness of the different formats of delivery of CBT (e.g.
face-to-face CBT versus telephone-delivered CBT or remote online CBT) were
excluded.
Table 1 Description of intervention models
The control groups involved a mixture of individuals on
waiting lists, receiving TAU (participants allocated to the ‘no treatment’
condition were also listed in this category because they are prone to seek
treatment while a study is being conducted), or any other psychological or
pharmacological treatment.
Primary outcomes were validated self-reported measures of
depression at baseline, post-treatment and/or additional follow-up points.
Studies that did not report their outcomes regarding depressive symptoms at
either baseline or post-treatment and/or provided insufficient data for a
meta-analysis were excluded.
We included randomised controlled trials (RCTs) evaluating
the effectiveness of CBT protocols. We excluded observational, cross-sectional
and qualitative studies. Studies that were not in English were also excluded.
Search methods
We searched the bibliographic databases MEDLINE, PsycInfo,
Embase, Web of Science and the Cochrane Register of Controlled Trials from 1
January 1990 to 30 November 2021. Two authors (I.A., C.H.) independently screened
the titles/abstracts and the full texts of potentially eligible studies and
extracted data. Interrater reliability was high (κ = 0.91) for
title/abstract screening and high (κ = 0.94) for full-text screening.
Disagreements were resolved through discussion. The reference lists of the
identified studies and those of previous reviews were examined to ensure that
all relevant studies were included. We also contacted experts in the field to
enquire about unpublished studies and searched trials registers (ClinicalTrials.gov,
ISRCTN, the World Health Organization's International Clinical Trials Registry
Platform (ICTRP) and OpenTrials.net) to identify any unpublished or ongoing
trials. The full search strategy is available in supplementary Appendix 1.
Data collection and extraction
A data extraction sheet was constructed and pilot tested on
six randomly selected papers. Data were extracted on: (a) study/context
characteristics: authors, geographic region where the study was conducted and
method of measuring depressive symptoms; (b) participant characteristics: age,
gender identity and presence of comorbidities; (c) CBT characteristics: number
of and/or components of CBT used, number of sessions, length of sessions and
delivery format; (d) therapist characteristics: background in mental health
services and supervision received; (e) active control/control group
characteristics: no treatment, waiting list, TAU, or other psychological and/or
pharmacological interventions; and (f) outcomes: measures of depression.
We also extracted arm-level data including information about
sample sizes, means and standard deviations for both intervention and control
conditions at baseline (when reported), post-treatment and follow-up.
Standardised mean difference (s.m.d.) effects and the corresponding standard
error were computed using the Comprehensive Meta-Analysis Version 3 (Windows)
software.
Quality assessment and risk of bias
The quality of the RCTs was assessed by two independent
raters (I.A., C.H.) using the Cochrane Risk of Bias 2 (RoB 2.0) tool. Additionally,
we applied the confidence in network meta-analysis (CINeMA) framework to
assess the certainty of evidence covering the six key domains: within-study
bias, reporting bias, indirectness, imprecision, heterogeneity and incoherence.
Data synthesis
We first did the pairwise meta-analyses using
DerSimonian–Laird random effects. We calculated standardised mean differences
using Hedges’ g and interpreted them according to Cohen's
criteria. A negative s.m.d. indicated that the reduction in depression
scores was in favour of the CBT protocols. We presented pooled effect results
with 95% confidence intervals and used forest plots with I (with
test-based 95% confidence intervals) to display statistical heterogeneity. Because
the s.m.d. is based on standardised means and not a specific scale (i.e. is
unit-less), we back-transformed the s.m.d. pooled effects to the mean
difference using the method explained in the Cochrane Handbook.
We then synthesised the study effect sizes by using a
network meta-analysis which allowed for the simultaneous evaluation of our
seven interventions while preserving the within-study randomisation. To
ensure transitivity within the network, we compared the distribution of the
clinical variables (i.e. age, gender and baseline depression score) by grouping
the different CBT protocols, other psychological treatments, medication, TAU
and ‘no treatment’ groups into nodes. A Bayesian random-effects network
meta-analysis model was used with a normal likelihood for the post-treatment
outcome analysis. The 95% credible interval (CrI) displayed uncertainty in the
posterior effects and multivariate distributions were used to account for the
correlations induced by multigroup studies. We considered the I statistic
and the (heterogeneity) variance in the random effects distribution (τ) to
measure the extent of the influence of variability across and within studies on
treatment effects. The surface under the cumulative ranking curve (SUCRA) was
used to rank the treatments’ performance, as well as the P-score, a frequentist
analogue to SUCRA. We statistically evaluated consistency by separating out
direct evidence from indirect evidence using node splitting. Cochrane's Q statistic
was used to calculate consistency throughout the entire network. The
CINeMA judgements were included in the league table of results and forest
plots.
Meta-regression analyses were conducted on the
post-treatment outcomes only because the exact same studies were included in
the long-term analysis, meaning that the variables would be the same. The study
and participant characteristics that were included in the analyses were:
geographical continent where the study was conducted (1, North and South
America; 2, Europe; 3, Africa; 4, Asia; 5, Oceania); socioeconomic status (high
versus low); age (≤30 years, 31–59 years, ≥60 years); gender (males versus
females); diagnosis (self-reported versus formal/interview); recruitment
(community versus in-patient); and comorbidities (none, mental health,
physical). The intervention characteristics included intensity of the treatment
(low: 1–8 sessions; medium: 9–15 sessions; high: 16+ sessions); delivery by a
mental health specialist (yes, no, not applicable); CBT format (individual
face-to-face/telephone, face-to-face group, online/face-to-face self-help with
some therapist support, online self-help with no therapist support); and
measure of depression used (Patient Health Questionnaire (PHQ-9); Beck
Depression Inventory (BDI); Hamilton Rating Scale for Depression (HRSD); Center
for Epidemiology Depression (CES-D); other). The influence of the quality
appraisal scores (low, medium, high) on the effects of different CBT protocols
in reducing depressive symptoms was also examined. We assessed goodness of fit
for each model by comparing total residual deviance and deviance information
criterion.
The models of the post-treatment outcome analysis were
fitted in OpenBUGS (version 3.2.3 for Windows) using uninformative prior
distributions for the treatment effects and a minimally informative prior
distribution for common heterogeneity standard deviation. Uninformative priors
(that is, N(0,1000)) were assumed for all meta-regression
coefficients. Model convergence was established by visual inspection of three
Monte Carlo Markov chains after considering the Brooks–Gelman–Rubin diagnostic.
Statistical evaluation of inconsistency and production of network graphs and
results figures was done using the ‘netmeta’ package in R version 4.0.5
(Windows) (R Foundation for Statistical Computing). Network meta-analysis
of the post-treatment outcome was duplicated in a frequentist environment by
using the same package in R.
A time adjusted analysis of the network involving the
outcome data from the 54 studies replacing the post-treatment data was done at
26 weeks to assess the long-term effectiveness of the interventions. This
analysis was done using the frequentist approach in ‘netmeta’. Our definition
of the long-term effects of CBT interventions as being 6 months (26 weeks)
follows the assumptions made from a previous study that this was when the
effects appeared to start to wane.
To examine the presence of bias due to small-study effects,
we used a comparison-adjusted network funnel plot to visually scrutinise the
criterion of symmetry. We also statistically compared the effect sizes
between short- and long-term outcomes (i.e. <26 weeks versus ≥26 weeks)
using the ratio of means (ROM) formula. All statistical codes used to
perform the network models are available in supplementary Appendix 13.
This study was conducted in accordance with the Cochrane
Handbook and was registered with PROSPERO (registration number
CRD42021237846). Reporting was consistent with the Preferred Reporting Items
for Systematic Reviews and Meta-Analyses (PRISMA) extension statement for
network meta-analysis. See supplementary Appendix 14 for the completed
checklist.
Patient and public involvement
This study was guided by two patients with lived experience
of depression who had received CBT in the past. They contributed to the
refinement of the research questions, classifications of the treatment protocols
and interpretation of the results. They will also support the dissemination of
the findings of the study.
Results
A total of 3975 articles were retrieved. After full-text
screening of 511 studies, 107 RCTs (involving 15 248 participants) met our
inclusion criteria (Fig. 1). Of those, 54 (50%) studies (involving 6383
participants) provided follow-up data. Supplementary Appendix 2 lists the
included studies, and supplementary Appendix 4 summarises their
characteristics.
Fig. 1 PRISMA 2020 flow diagram for the entire review.
Descriptive characteristics of studies, population, intervention and
outcomes
Combined, 43% (n = 46) of the studies were
conducted in North and South America, 31.8% (n = 34) in Europe
and the UK, 11.2% (n = 12) in Oceania, 13% (n = 14)
across Asia, including such countries as China, Japan, Pakistan, Iran and Thailand,
and 1% (n = 1) in Nigeria, Africa.
The age of the participants ranged between 22.6 and 74.7
years (mean 41.5; s.d. = 13.5) with 5627 (36.3%) of the overall
sample identifying as male. Fifty-eight (54.2%) studies diagnosed depression
using formal diagnoses, whereas the remaining 49 (45.8%) used self-report
measures. Of those studies in which participants reported additional physical (n = 26;
70.3%) or mental (n = 11; 29.7%) health problems, the most
common were HIV, cancer, multiple sclerosis, cardiovascular problems and
diabetes, whereas anxiety disorders (n = 8; 72.7%) were among
the most common comorbid mental health problem.
Of the total 127 comparisons included in this review, 67
(52.8%) were based on ultra-complex CBT, 36 (28.3%) on complex CBT and 24
(18.9%) on core CBT protocols. All CBT protocols were compared with multiple
comparators including TAU (57; 44.8%), no intervention and/or waiting list
conditions (36; 28.3%), another psychological therapy, including interpersonal
and psychodynamic psychotherapy (22; 17.3%), and medication alone (13; 10.2%).
Regarding the mode of delivery of the psychological interventions, 46 (36.2%)
studies used a face-to-face individual format, 40 (31.5%) used group sessions,
37 (29.1%) used an online format with and/or without a therapist's support, 3
(2.4%) used either a face-to-face or online teaching-based format and 1 (0.8%)
used self-help. The average number of sessions was 9.8 (s.d. = 4.7),
with a mean length of 72.7 min (s.d. = 27.5).
Assessment of risk of bias
The overall bias appraisal revealed that 79 studies (73.8%)
showed moderate risk, 16 (15%) studies demonstrated low risk and 12 (11.2%)
showed high risk of bias. An area of bias that was potentially problematic was
selective reporting bias: 91 (85%) of studies showed moderate or high risk of
bias. Results of the full risk of bias assessment are reported in supplementary
Appendix 5.
Network meta-analysis for main outcomes
Figure 2 shows the network of eligible comparisons for
all core CBT packages for the post-treatment outcomes from the 107 studies. The
network of evidence included 7 interventions, 15 248 participants, 90
two-arm studies and 17 multi-arm studies.
Fig. 2 Network graph and forest plot of network
meta-analysis for main outcomes.CBT, cognitive–behavioural therapy; TAU,
treatment as usual; a, low confidence of evidence; b, moderate confidence of
evidence.
Inconsistency analysis
We found evidence of statistical inconsistency through node
splitting analysis owing to comparisons of complex CBT (z = −4.67, P < 0.0001)
with no treatment, complex CBT with TAU (z = 2.94, P = 0.003)
and core CBT with no treatment (z = 2.19, P = 0.029)
(supplementary Appendix 6). The inconsistency for complex CBT compared with TAU
was due to one study, which showed a high overall risk of bias, a large
effect size and large standard error. The inconsistency for complex CBT
compared with no treatment was owing to one study, which revealed high
risk of bias because of missing data, concerns due to the unknown randomisation
procedure used and the extremely wide confidence interval. Finally, the
inconsistency for core CBT compared with no treatment was due to one study, which
had a high risk of bias for the randomisation process used and concerns of
measurement outcome and outcome reporting bias. Because consistency
(transitivity) is a central assumption of network meta-analysis, we removed all
three trials, leaving 104 RCTs in the network.
Main outcomes
Figure 2 shows the results of the network meta-analysis
for the main outcomes of all eligible trials after performing the inconsistency
analysis. All active interventions, including core CBT
(s.m.d. = −1.14, 95% CrI −1.72 to −0.55 [m.d. = −8.44, 95%
CrI −12.73 to −4.07], n = 6 studies), complex CBT
(s.m.d. = −1.24, 95% CrI −1.85 to −0.64 [m.d. = −9.18, 95%
CrI −13.69 to −4.74], n = 9), ultra-complex CBT
(s.m.d. = −1.45, 95% CrI −1.88 to −1.02 [m.d. = −10.73, 95%
CrI −13.91 to −7.55], n = 21), other psychological
psychotherapies (s.m.d. = −0.76, 95% CrI −1.35 to −0.16
[m.d. = −5.62, 95% CrI −9.99 to −1.18]; n = 3),
medication (s.m.d. = −0.80, 95% CrI −1.58 to −0.01
[m.d. = −5.92, 95% CrI −11.69 to −0.07]; n = 2)
and TAU (s.m.d. = −0.74, 95% CrI −1.24 to −0.23
[m.d. = −5.48, 95% CrI −9.18 to −1.70]; n = 2)
showed statistically significant benefits compared with no treatment. Large
heterogeneity was present in the network meta-analysis, with I = 91.5%
(95% CI 90.3–92.6%) (supplementary Appendix 7). These results were consistent
when analysed in a frequentist framework. The pairwise meta-analysis results
for the main outcomes were also consistent for core and multicomponent CBT when
compared with either TAU or no treatment (supplementary Appendix 8).
The SUCRA also supported the network meta-analysis showing
the best performing intervention as ultra-complex CBT (SUCRA = 93.9%)
followed by complex CBT (SUCRA = 77.7%) (supplementary Appendix 9).
The league table showing the results of the network
meta-analysis comparing the effects of all interventions (Fig. 3) showed that
both ultra-complex (s.m.d. = −0.71, 95% CrI −1.05 to −0.38
[m.d. = −5.25, 95% CrI −7.77 to −2.81]) and complex CBT protocols
(s.m.d. = −0.50, 95% CrI −0.95 to −0.06 [m.d. = −3.70, 95%
CrI −7.03 to −0.44]) were the only interventions that maintained a significant
effect when compared with TAU. Ultra-complex CBT was also significantly more
effective than the use of other psychological treatments
(s.m.d. = −0.69, 95% CrI −1.19 to −0.20 [m.d. = −5.11,
−8.81 to −1.48]). To ensure the certainty of evidence, we incorporated the
CINeMA judgements into Fig. 3. The evidence according to CINeMA varied
from low (n = 6 head-to-head comparisons), to moderate (n = 5)
to high (n = 7) confidence overall (supplementary Appendix
10). Funnel plots and Egger's test for assessing asymmetry indicated strong
evidence for publication bias (P < 0.0001) (supplementary
Appendix 11).
Fig. 3 Head-to-head comparisons of all intervention
groups for the main outcome network analysis.The interventions are described in
Table 1. CBT, cognitive–behavioural therapy; TAU, treatment as usual. Data are
shown as s.m.d. (95% CrI); –, no direct treatment comparisons. Darker blue
cells (bottom) show network meta-analysis estimates; lighter blue cells (top)
show direct pairwise meta-analysis estimates. The certainty of the evidence
(according to the confidence in network meta-analysis (CINeMA) framework) is:
a, very low confidence; b, low confidence; c, moderate confidence; d, high
confidence; e, very high confidence. Full results from CINeMA are provided in
supplementary Appendix 8.
Covariate adjusted network at 26 weeks (6 months) including follow-up data
The long-term effectiveness of each active intervention,
including various CBT protocols, other psychological treatment, medication and
TAU, was assessed in the covariate-adjusted network model for 26 weeks or more
use (see Fig. 4 for forest plot and Fig. 5 for league table
of comparisons). Ultra-complex CBT (s.m.d. = −1.09, 95% CI −1.61 to
−0.56 [m.d. = −8.07, 95% CI −8.58 to −4.14]), complex CBT
(s.m.d. = −0.73, 95% CI −1.36 to −0.11 [m.d. = −5.40, 95%
CI −10.06 to −0.81]), other psychological treatments (s.m.d. = −0.71,
95% CI −1.37 to −0.04 [m.d. = −5.25, 95% CI −10.14 to −0.30]) and TAU
(s.m.d. = −0.48, 95% CI −0.86 to −0.10 [m.d. = −3.55, 95%
CI −6.36 to −0.74]) maintained significance after 26 weeks post-treatment when
compared with no treatment.
Fig. 4 Network graph and forest plot of network
meta-analysis for time-adjusted analysis.CBT, cognitive–behavioural therapy;
TAU, treatment as usual; wks, weeks.