Skip to content

C3 · Lab · Psyche · A series · Interactive, in your browser

Does CBT work?

What trials found when CBT for depression was compared with waiting, other treatments and time, and how much of the headline holds up.

This isn’t advice about your own care. It explains a therapy and what the trials found.

The question has a number

One meta-analysis of CBT for depression pooled 409 trials and 52,702 patients [Cuijpers et al. 2023 (meta-analysis)]. All that evidence, and one question to ask of it.

Whether it works has an honest answer, and it’s a number. The number depends on what the treatment is compared with.

What a g of 0.79 means

At the end of treatment, against control conditions, the pooled effect across 271 comparisons was g = 0.79 (95% CI 0.70 to 0.89) [Cuijpers et al. 2023 (meta-analysis)]. On CBT 42% responded, against 19% on control: a risk ratio of 2.13 [Cuijpers et al. 2023 (meta-analysis)].

A g is hard to feel. There are two plainer ways to read it, and the first is the probability that a treated person ends up better off than a control person picked at random. For two normal curves of equal spread that’s Φ(g/√2), so 71% at g = 0.79 [Furukawa and Leucht 2011 (methods paper)].

The other is the number needed to treat, the people treated for one extra good outcome. For this g, Kraemer and Kupfer’s method gives 2.4, and Furukawa’s gives 3.8 at a control event rate of 17% [Furukawa and Leucht 2011 (methods paper); Kraemer and Kupfer 2006 (methods paper)]. Cuijpers and colleagues used Furukawa’s, so 3.8 is the number in their paper [Cuijpers et al. 2023 (meta-analysis)]. The two methods in the literature don’t agree.

Furukawa and Leucht tested both on trials of antipsychotics. Furukawa’s method landed inside the observed range in 97% of cases, Kraemer’s in 56%. That’s one area of medicine. I wouldn’t stretch it [Furukawa and Leucht 2011 (methods paper)].

Control outcomesTreated outcomes

0.79
17%
Probability of superiority
71.2%
The chance a treated person beats a control person.
Number needed to treat, Kraemer
2.4
Kraemer and Kupfer: depends on g alone.
Number needed to treat, Furukawa
3.8
At a control event rate of 17%.
FigureControl outcomes are centred on zero and treated ones shifted right by g: two normal curves of equal spread. Both number-needed-to-treat methods are named. Furukawa’s needs a control event rate.

The estimate shrinks

The 0.79 is the headline. It shrinks as the analysis gets stricter: 0.70 with outliers removed, 0.60 in the trials at low risk of bias, 0.47 after an adjustment for publication bias. The numbers needed to treat run 3.8, 4.4, 5.2 and 7.0 [Cuijpers et al. 2023 (meta-analysis)].

That adjustment is trim and fill. It imputes studies that look to be missing, 51 of them in one analysis of psychotherapy trials [Cuijpers et al. 2010 (meta-analysis)]. It rests on an assumption about what’s missing. So I read 0.47 as one estimate, not the true value.

Two other papers point the same way. Neither is about CBT alone. Across all psychotherapies for adult depression, the mean effect of 0.67 fell to 0.42 after trim and fill, and the authors wrote that “The effects of psychotherapy for adult depression seem to be overestimated considerably” [Cuijpers et al. 2010 (meta-analysis)].

The other looked at 55 US National Institutes of Health grants that began trials. Of those, 13 produced no publication. The six unpublished comparisons averaged 0.20, the 20 published ones 0.52, and all of them together 0.39 [Driessen et al. 2015 (meta-analysis)].

CBT for depression against control conditions (Cuijpers et al. 2023)

All psychological treatments of depression, NIH-funded trials (Driessen et al. 2015)

FigureEach bar is an estimate with its 95% confidence interval. Pick one to set the first figure’s g. Beside the papers’ numbers needed to treat, the figure shows the library’s own at the same g.

Compared with what?

The control matters. In the same analysis, using a waitlist was the only moderator that stayed significant after adjustment (p = 0.02). Response in the waitlist controls was 17% [Cuijpers et al. 2023 (meta-analysis)].

I should be careful here, because the paper gives no separate g for waitlist, care as usual or a pill placebo. How much the waitlist matters? I can’t put a number on it.

In a network meta-analysis of 49 trials and 2,730 people, those left untreated did better than those on a waiting list (odds ratio 2.9, 95% CI 1.3 to 5.7). The authors call it a possible nocebo condition. The differences didn’t survive their sensitivity analyses, they say [Furukawa et al. 2014 (meta-analysis)].

Against a placebo, the numbers come from anxiety rather than depression: a meta-analysis of 41 placebo-controlled trials gave g = 0.56 for target symptoms, falling to 0.33 after the publication bias adjustment [Carpenter et al. 2018 (meta-analysis)]. By disorder it ran from 1.13 for OCD, on four trials, to 0.48 for PTSD, on fourteen [Carpenter et al. 2018 (meta-analysis)].

Against other psychotherapies for depression the difference was g = 0.06 (0 to 0.12). The low-risk-of-bias, outlier-removed and publication-bias-adjusted analyses didn’t find it significant, and the authors’ verdict is that “The superiority of CBT over other psychotherapies for depression does not emerge clearly” [Cuijpers et al. 2023 (meta-analysis)].

Does it last?

Against controls, the effect fades with time: g was 0.74 at 6 to 9 months, from 78 comparisons, 0.49 at 10 to 12 months, from 22, and 0.22 at 13 to 24 months, from 8, which wasn’t significant. The later the figure, the fewer comparisons it rests on [Cuijpers et al. 2023 (meta-analysis)].

What happens after treatment is a different question. Hollon and colleagues followed people who’d responded to treatment for twelve months. Of those who’d had cognitive therapy and then stopped, 30.8% relapsed, against 76.2% of those whose medication was withdrawn onto placebo and 47.2% of those who stayed on medication [Hollon et al. 2005 (paper)].

The first gap was significant (P = .004). The second, against staying on medication, wasn’t (P = .20). The arms were small, too [Hollon et al. 2005 (paper)].

A meta-analysis of nine studies and 506 patients found fewer relapses after CBT than after stopping medication (odds ratio 2.61, 95% CI 1.58 to 4.31, NNT 5). Against continued medication the odds ratio was 1.62 (0.97 to 2.72), which wasn’t significant. The studies were few. So the authors ask for caution until it’s replicated [Cuijpers et al. 2013 (meta-analysis)].

  • Cognitive therapy, then stopped

    30.8%

    relapsed: 31 of 100

  • Medication withdrawn onto placebo

    76.2%

    relapsed: 76 of 100

  • Medication continued

    47.2%

    relapsed: 47 of 100

Cognitive therapy, then stopped against Medication continued: not significant (P = .20).

FigureOn each grid a percentage is drawn as people in a hundred, rounded to a whole dot. The dots are percentages, not real people.

Is it getting weaker?

One paper says it is. Johnsen and Friborg pooled 70 studies of CBT for depression from 1977 to 2014, and wrote that “the effects of CBT have declined linearly and steadily since its introduction” [Johnsen and Friborg 2015 (meta-analysis)]. But their effect sizes, about 1.6 and 1.7, are before-to-after changes within the CBT group, not CBT against a control. They can’t be set beside the 0.79 above [Johnsen and Friborg 2015 (meta-analysis)].

Two re-analyses disagree. Ljótsson and colleagues took the same 70 studies and found a decline only for 1977 to 1995, driven by 22 early studies, not in the 48 after [Ljótsson et al. 2017 (paper, a re-analysis)]. Cristea and colleagues re-ran it on randomised trials only, with 30 more trials, and in no analysis did year predict the difference between CBT and a control. Their verdict: the linear decline is likely spurious [Cristea et al. 2017 (paper, a re-analysis)].

Friborg and Johnsen replied, and they maintain their finding [Friborg and Johnsen 2017 (paper, a reply)]. So it’s contested. I’m not taking a side here.

Where the guideline puts it

For each level of severity, NICE’s 2022 guideline on depression in adults puts the treatments in order. In its table for more severe depression, individual CBT with an antidepressant comes first. Individual CBT on its own comes second [NICE 2022 (guideline)].

For less severe depression, guided self-help (with CBT among its materials) leads the table, followed by group CBT [NICE 2022 (guideline)]. The guideline says every option in a table can be first line. The choice is shared. The order reflects its committee’s reading of clinical and cost effectiveness [NICE 2022 (guideline)].

Limits and counter-evidence

Heterogeneity was high (I² = 85). The prediction interval says a new trial’s effect might land between −0.45 and 2.04 [Cuijpers et al. 2023 (meta-analysis)]. So the 0.79 hides a lot of spread.

The authors’ own list of limitations has signs of publication bias on it, and this: many of the trials were at high risk of bias, and the low-risk ones found smaller effects [Cuijpers et al. 2023 (meta-analysis)].

In the anxiety trials, the effect was 0.40 when everyone assigned was analysed (intention to treat), against 0.70 for completers. Only 16 of 41 trials did it that way [Carpenter et al. 2018 (meta-analysis)].

The number needed to treat depends on the method: 2.4 or 3.8 for the same g [Furukawa and Leucht 2011 (methods paper)]. Mostly, this chapter is about depression. There’s a little on anxiety. Other conditions, children and any one person’s course aren’t in it.

Sources

  1. Cuijpers et al. 2023 (meta-analysis) Cognitive behavior therapy vs. control conditions, other psychotherapies, pharmacotherapies and combined treatment for depression: a comprehensive meta-analysis including 409 trials with 52,702 patients. World Psychiatry 22(1). Open access on PubMed Central, full text, read 2026-11-18.

    The effect of CBT for depression against control conditions at the end of treatment, in four analyses; response rates; the waitlist moderator; follow-up; CBT against other psychotherapies; heterogeneity and the authors’ limits.

    • Read: full text
  2. Furukawa and Leucht 2011 (methods paper) How to obtain NNT from Cohen’s d: comparison of two methods. PLoS One 6(4). Open access on PubMed Central, full text, read 2026-11-18.

    The two ways of turning an effect size into a number needed to treat, their formulas and assumptions, and a test of both on trials of antipsychotics.

    • Read: full text
  3. Kraemer and Kupfer 2006 (methods paper) Size of treatment effects and their importance to clinical research and practice. Biological Psychiatry 59(11). PubMed record, abstract only, read 2026-11-18; its formulas are known here only as Furukawa and Leucht quote them.

    That the paper recommends the number needed to treat and the probability a treated patient does better than a control as plainer measures than a bare effect size.

    • Read: abstract
  4. Cuijpers et al. 2010 (meta-analysis) Efficacy of cognitive-behavioural therapy and other psychological treatments for adult depression: meta-analytic study of publication bias. British Journal of Psychiatry 196(3). PubMed record, abstract only, read 2026-11-18.

    All psychotherapies for adult depression, not CBT alone: the mean effect before and after the trim and fill adjustment, and the authors’ conclusion.

    • Read: abstract
  5. Driessen et al. 2015 (meta-analysis) Does publication bias inflate the apparent efficacy of psychological treatment for major depressive disorder? A systematic review and meta-analysis of US National Institutes of Health-funded trials. PLoS One 10(9). PubMed record, abstract only, read 2026-11-18.

    US National Institutes of Health grants for trials of psychological treatment of depression (all treatments, not CBT alone): how many produced no publication, and the effect in unpublished and published comparisons.

    • Read: abstract
  6. Furukawa et al. 2014 (meta-analysis) Waiting list may be a nocebo condition in psychotherapy trials: a contribution from network meta-analysis. Acta Psychiatrica Scandinavica 130(3). PubMed record, abstract only, read 2026-11-18.

    A network meta-analysis of 49 trials: the estimated effect of CBT differed by control used, no treatment did better than a waiting list, and the authors’ own hedges.

    • Read: abstract
  7. Carpenter et al. 2018 (meta-analysis) Cognitive behavioral therapy for anxiety and related disorders: a meta-analysis of randomized placebo-controlled trials. Depression and Anxiety 35(6). Author manuscript on PubMed Central, full text, read 2026-11-18.

    CBT for anxiety and related disorders against placebo: the pooled effect, the effect by disorder with its trial counts, intention-to-treat against completers, and the publication bias adjustment.

    • Read: full text
  8. Hollon et al. 2005 (paper) Prevention of relapse following cognitive therapy vs medications in moderate to severe depression. Archives of General Psychiatry 62(4). PubMed record, abstract only, read 2026-11-18.

    The continuation phase of a randomised trial: relapse over twelve months after cognitive therapy, after medication was withdrawn and after medication was continued.

    • Read: abstract
  9. Cuijpers et al. 2013 (meta-analysis) Does cognitive behaviour therapy have an enduring effect that is superior to keeping patients on continuation pharmacotherapy? A meta-analysis. BMJ Open 3(4). Open access on PubMed Central, abstract only, read 2026-11-18.

    Nine studies of relapse after acute CBT against stopped or continued medication: the odds ratios, the numbers needed to treat and the authors’ caution.

    • Read: abstract
  10. Johnsen and Friborg 2015 (meta-analysis) The effects of cognitive behavioral therapy as an anti-depressive treatment is falling: a meta-analysis. Psychological Bulletin 141(4). PubMed record, abstract only, read 2026-11-18.

    Seventy studies from 1977 to 2014 and their before-to-after effect sizes (not CBT against a control), and the authors’ finding of a decline with year.

    • Read: abstract
  11. Ljótsson et al. 2017 (paper, a re-analysis) The effects of cognitive-behavioral therapy for depression are not falling: a re-analysis of Johnsen and Friborg (2015). Psychological Bulletin 143(3). PubMed record, abstract only, read 2026-11-18.

    A re-analysis of the same 70 studies: a decline only before 1995, driven by 22 early studies, and none after.

    • Read: abstract
  12. Cristea et al. 2017 (paper, a re-analysis) The effects of cognitive behavioral therapy are not systematically falling: a revision of Johnsen and Friborg (2015). Psychological Bulletin 143(3). PubMed record, abstract only, read 2026-11-18.

    A re-run on randomised trials only, with 30 more trials: year did not predict the difference between CBT and a control.

    • Read: abstract
  13. Friborg and Johnsen 2017 (paper, a reply) The effect of cognitive-behavioral therapy as an antidepressive treatment is falling: reply to Ljótsson et al. (2017) and Cristea et al. (2017). Psychological Bulletin 143(3). PubMed record, abstract only, read 2026-11-18.

    That the original authors dispute both re-analyses and maintain their finding.

    • Read: abstract
  14. NICE 2022 (guideline) Depression in adults: treatment and management, NICE guideline NG222. Recommendations and rationale chapters, published 29 June 2022, full text, read 2026-11-18.

    The order of treatments in its tables for less severe and for more severe depression, and what the guideline says about choosing among them.

    • Read: full text

Each fact is paraphrased from the page named, read on the date shown. “Abstract” means I read the abstract only and say no more than it does. Where a sentence is my own reading, it is marked as my view.