MUPLAB
Bayesian Living Meta-Analysis · v1.0

The effect of auricular taVNS on learning

The effect of taVNS on accuracy and response time (RT). The two outcomes are pooled separately; positive g favors taVNS.

Data & Code: GitHubUpdated: 23 June 2026

Primary pool

Accuracy

+0.15
95% CrI [0.06, 0.25]
P(g>0) = 99.9% · τ = 0.11 · PI [-0.13, 0.44] · k = 21

Response time (RT)

+0.18
95% CrI [0.02, 0.37]
P(g>0) = 98.3% · τ = 0.24 · PI [-0.37, 0.76] · k = 14
Key finding. The strongest effect is on response time in associative & skill learning.
RT × Associative & Skill: g = 0.57 [0.28, 0.84] (fragile, k = 4)Overall: ACC g = 0.15 · RT g = 0.18ρ(ACC,RT) = 0.14 [−0.73, 0.86]

This is a living meta-analysis: the corpus is updated and versioned as new studies are published. Browse the registered analysis in the tabs, or run your own subset with the real-R engine.

1Registered analysis
2R engine
3Diagnostics
4Living versions
1 · Registered analysis: Results and figures locked with brms · bayesmeta. 2 · R engine: Pick studies, adjust priors — the analysis re-runs in real R in your browser. 3 · Diagnostics: Leave-one-out diagnostics; domain subgroups in their own tab. 4 · Living versions: Corpus updates are recorded in the version history.
Registered analysis (locked reference) — registered corpus · brms · bayesmeta · metafor. The figures here are the registered R output and do not change; they are independent of the include/exclude choices in the interactive engine below.
v1.0 — registered. 24 studies, 36 primary effect sizes (22 ACC + 14 RT). Headline values are the registered brms model; in the interactive section below you can include/exclude any study and run the analysis on your chosen subset with real R (bayesmeta) (in-browser, not saved).

Accuracy — registered brms

+0.15
Pooled g
[0.06, 0.25]
95% CrI
99.9%
P(g > 0)
[-0.13, 0.44]
Prediction interval
0.11
τ
22
k

Response time (RT) — registered brms

+0.18
Pooled g
[0.02, 0.37]
95% CrI
98.3%
P(g > 0)
[-0.37, 0.76]
Prediction interval
0.24
τ
14
k
Subgroup forest — registered
Registered subgroup forest — Accuracy
Subgroup forest — Accuracy
Registered subgroup forest — RT
Subgroup forest — RT
Contour-enhanced funnel — registered
Registered funnel — Accuracy
Contour-enhanced funnel — Accuracy
Registered funnel — RT
Contour-enhanced funnel — RT
Normal Q–Q — registered
Registered Q–Q — Accuracy
Normal Q–Q — Accuracy
Registered Q–Q — RT
Normal Q–Q — RT

Interactive explorer — real-R engine (bayesmeta · webR)

Include or exclude any publication — the forest, funnel, Q–Q and all statistics recompute for your chosen subset with real R (bayesmeta · webR) in your browser; ACC and RT are shown together; after changing the selection press Re-run — each pool recomputes in ~15–25 s. The first visit downloads the R runtime (~90 MB; later visits use the cache). Open full page ↗

Domain subgroups (registered, bayesmeta)

OutcomeDomainkPooled g95% CrI
AccuracyDeclarative Memory5+0.09[-0.06, 0.22]
AccuracyWorking Memory6+0.27[0.13, 0.40]
AccuracyCognitive Control3+0.19[-0.07, 0.45]
AccuracyAssociative & Skill Acquisition7+0.03[-0.12, 0.21]
Response time (RT)Working Memory5-0.06[-0.22, 0.11]
Response time (RT)Cognitive Control5+0.21[-0.01, 0.43]
Response time (RT)Associative & Skill Acquisition4+0.57[0.28, 0.84]
Registered extra diagnostics — leave-one-out (fixed)

Registered R/bayesmeta outputs; ACC and RT shown separately.

Leave-one-out (caterpillar) — ACC and RT
Leave-one-out (caterpillar) — Accuracy
Leave-one-out (caterpillar) — Accuracy
Leave-one-out (caterpillar) — RT
Leave-one-out (caterpillar) — RT
Domain-level leave-one-out — ACC and RT
Domain-level leave-one-out — Accuracy
Domain-level leave-one-out — Accuracy
Domain-level leave-one-out — RT
Domain-level leave-one-out — RT

Characteristics

Characteristics of included studies (paper level, 22 studies). Multi-experiment papers (Sun 2021, Ventura-Bort 2025) contribute two experiments/studies; the analysis counts 24 study units. Domain = 4-category; Outcome = ACC/RT.

StudyYearCountryDesignBlindingNAge (M)%FDomainOutcome
Jongkees 20182018Netherlands/Australia/GermanyBetweenSingle402280Cognitive ControlRT,ACC
Giraudier 20202020GermanyBetweenSingle60Declarative MemoryACC
Kühnel 20202020GermanyWithinSingle392659Associative & SkillACC
Mertens 20202020BelgiumWithinSingle412251Declarative MemoryACC
Thakkar 20202020USABetweenSingle372173Associative & SkillRT,ACC
D'Agostini 20212021BelgiumBetweenSingle712377Associative & SkillACC
Kaan 20212021USABetweenSingle622066Working MemoryRT,ACC
Sun 20212021ChinaWithinSingle46Working MemoryRT,ACC
Phillips 20222022USABetweenDouble452264Associative & SkillRT,ACC
Zhao 20222022ChinaWithin—632152Working MemoryRT,ACC
Konjusha 20232023GermanyWithinSingle3725Working MemoryACC
Sommer 20232023GermanyWithinSingle322659Cognitive ControlRT
Tian 20232023ChinaWithinSingle93Working MemoryRT,ACC
Bömmer 20242024GermanyWithinDouble272548Cognitive ControlRT,ACC
Chen 20242024ChinaWithinSingle222346Associative & SkillRT
Honda 20242024CanadaBetweenDouble452373Associative & SkillACC
Li 20252025ChinaWithinSingle61Cognitive ControlRT
Sönmez 20252025GermanyWithinSingle292551Cognitive ControlRT,ACC
Thakkar 20252025USAWithinSingle3520Associative & SkillACC
Ventura-Bort 2025 S12025GermanyWithinSingle302187Declarative MemoryACC
Çakır 20252025TürkiyeBetweenSingle80Associative & SkillRT,ACC
Mary 20262026BelgiumWithinSingle892452Declarative MemoryACC
Speed–accuracy trade-off — registered joint analysis — a joint (bivariate) Bayesian model over the 11 matched studies reporting both accuracy and RT. It tests whether taVNS buys its accuracy gain at a speed cost. The figures and values here are registered; they are independent of the include/exclude choices above.
No speed–accuracy trade-off. taVNS raises accuracy (g = 0.29 [0.17, 0.41]) with no cost to response time (RT g = 0.14 [−0.04, 0.36]); the two outcomes move together only weakly and uncertainly (ρ = 0.14 [−0.73, 0.86]). The probability of a meaningful trade-off is ≈ 0%; the most probable pattern is dual benefit (93%).
+0.29
Accuracy effect [0.17, 0.41]
+0.14
RT effect [−0.04, 0.36]
0.14
ρ (study-level) [−0.73, 0.86]
93%
P(dual benefit)
~0%
P(meaningful trade-off)
11
k (matched)

Note: these means come from the joint model over the 11 matched studies only, so they differ from the full-corpus pools (accuracy g = 0.15, RT g = 0.18). 11/24 studies matched; the rest reported accuracy or RT only and could not be paired.

Quadrant scatter — accuracy × RT

The 11 matched studies: horizontal axis = accuracy effect (g), vertical axis = RT effect (g); positive on both = favors taVNS. Upper-right quadrant = dual benefit (more accurate and faster), lower-right = more accurate but slower. Filled circle = dual benefit, open circle = accurate-but-slower. 8 studies fall in dual benefit, 3 in the more-accurate-but-slower quadrant.

Decision probabilities (by ROPE threshold)

Posterior probability of each pattern at three ROPE thresholds (0 / 0.1 / 0.2). The "trade-off" bars are effectively empty: accuracy at a speed cost is under 7%, and speed at an accuracy cost is ≈ 0%.

Sensitivity

Assuming a matched accuracy–RT sampling correlation of r = −0.5 … +0.5, the study-level ρ stays within 0.10–0.21 and its 95% CrI always spans zero; the probability of a meaningful trade-off remains ≈ 0%. The conclusion is robust.

Matched subset — by domain

DomainkAccuracy g [95%]RT g [95%]
Working Memory5+0.27 [0.12, 0.42]−0.06 [−0.23, 0.13]
Cognitive Control3+0.19 [−0.09, 0.45]+0.24 [−0.05, 0.54]
Associative & Skill Acquisition3+0.43 [0.10, 0.74]+0.52 [0.18, 0.86]

Working memory shifts toward "more accurate but slightly slower"; associative/skill learning shows the strongest dual benefit.