Subgroup and Moderator Analyses of CBT for CHR-P

confidence: high · 1 sources

⚠
Working research — not a recommendation.

This page records what the evidence says a policy does, not what anyone should do about it. It has been synthesized from the cited sources but has not been independently reviewed.

What the evidence says

The claims on this page do not cover the stated topic. None of them studies people at clinical high risk for psychosis (CHR-P), and none studies CBT delivered by a clinician. All six come from one meta-analysis of 15 randomized trials of CBT-based AI conversational agents, with depression as the main outcome. For CHR-P, subgroup and moderator evidence is absent from these sources. Read the findings below as evidence about a different population and delivery format.

Within that meta-analysis, the moderators for depressive symptoms were design features and participant features. Agents supporting text, voice, and/or image interactions (multimodal) showed a larger effect than single-modality agents: g = 0.82 (95% CI [0.54, 1.09]). That comparison rests on only 2 multimodal studies. The authors say design and implementation factors likely confound it.

Agents without psychoeducational content showed a larger post-test effect on depression: g = 0.47 (95% CI [0.25, 0.69]). The authors call this preliminary because few studies were available. Both subgroup findings are associations between study features and effect size. They do not show that adding or removing a feature changes outcomes.

Two meta-regressions found moderators of the post-test effect on depression. Mean participant age had a coefficient of β = −0.02 (p < 0.05), so younger samples showed larger reductions. Study quality, assessed with the Cochrane Risk of Bias tool, had a coefficient of β = 0.11 (p < 0.05), so higher-quality studies reported larger effects. The authors read the quality result as a sign that true efficacy may be underestimated in the current literature.

Secondary outcomes are weaker. CBT-based agents decreased a negative-emotion outcome by a small amount, g = 0.21 (95% CI [0.02, 0.39]), and the authors say effects this small may not be meaningful in clinical practice. The generalized anxiety effect, g = 0.18 (95% CI [0.05, 0.31]), fell to g = 0.12 (95% CI [−0.03, 0.27]) after a publication-bias adjustment.

Where the evidence splits

The generalized anxiety result splits within the one source. The unadjusted estimate was g = 0.18 (95% CI [0.05, 0.31], p < 0.01). After trim-and-fill imputation of missing studies, it was g = 0.12 (95% CI [−0.03, 0.27], p > 0.05), and the interval includes zero. The difference comes from the publication-bias adjustment: the authors note that the initial estimate may have been inflated. The claims do not say which estimate is closer to the true effect.

Thresholds and dose-response

No threshold established in the current sources.

Open empirical questions

Findings

Multimodal CBT-based conversational agents show larger decreases in depressive symptoms than single-modality agents

confidence — direction: mediumconfidence — magnitude: lowconsensus: thin; one meta-analysis
mechanism
not established; the authors say design and implementation factors may confound the comparison
effect
decreases depressive symptoms more in multimodal agents, g = 0.82 (95% CI [0.54, 1.09]); the single-modality effect is smaller, with no figure given
conditions / moderators
based on only 2 multimodal studies; association across studies, not an experiment on modality

Conversational agents without psychoeducational content show larger decreases in depressive symptoms than those with it

confidence — direction: mediumconfidence — magnitude: lowconsensus: thin; one meta-analysis
mechanism
not established
effect
larger post-test decrease in depression without psychoeducation, g = 0.47 (95% CI [0.25, 0.69]); the effect with psychoeducation is smaller, with no figure given
conditions / moderators
limited number of studies; the authors call the finding preliminary

Younger mean participant age is associated with larger decreases in depressive symptoms

confidence — direction: mediumconfidence — magnitude: lowconsensus: thin; one meta-regression
mechanism
not established
effect
meta-regression coefficient β = −0.02 (p < 0.05); the negative sign means younger samples show larger effects. The coefficient's units are not stated in the claim.
conditions / moderators
study-level mean age, not individual-level age; the claims give no age range

Higher-quality studies report larger decreases in depressive symptoms

confidence — direction: mediumconfidence — magnitude: lowconsensus: thin; one meta-regression
mechanism
not established; the authors suggest lower-quality studies may understate true efficacy
effect
meta-regression coefficient β = 0.11 (p < 0.05); the positive sign means higher quality goes with larger effects
conditions / moderators
quality was rated with the Cochrane Risk of Bias tool

CBT-based conversational agents decrease a negative-emotion outcome by a small amount

confidence — direction: mediumconfidence — magnitude: lowconsensus: thin; one meta-analysis
mechanism
not established
effect
decreases the negative-emotion outcome at post-test, g = 0.21 (95% CI [0.02, 0.39]); the lower bound is near zero
conditions / moderators
the authors say effects this small may not represent meaningful improvement in clinical practice

The generalized anxiety effect falls to a no-measured-effect result after publication-bias adjustment

confidence — direction: lowconfidence — magnitude: lowconsensus: contested within one source; the adjusted estimate leans toward no measured effect
mechanism
not established; the adjustment imputes studies presumed missing
effect
unadjusted g = 0.18 (95% CI [0.05, 0.31], p < 0.01); adjusted g = 0.12 (95% CI [−0.03, 0.27], p > 0.05)
conditions / moderators
the authors say the initial estimate may have been inflated