Measured Statistical Power of Small-Sample Anxiety Validation and Intervention Studies

What the evidence says

Across neuroimaging, randomized trials, quasi-experiments, and meta-analyses of anxiety-related outcomes, small sample size recurs as the dominant constraint on statistical power. A consumer-EEG subcohort of n = 25 is described as producing insufficient power to detect modest brain-behavior correlations, and the same source states that combining a limited sample with multiple hypothesis tests further reduces power and raises the likelihood of Type II error. This pattern is not confined to one study design: a sports-psychology intervention notes that its small sample and baseline variability "likely reduced statistical power across outcomes" despite observing effect sizes as large as η² = 0.53, and a cross-sectional psychometric study of psychological inflexibility states plainly that its sample size "may still limit... precision" and the detection of small effects. A meta-analysis of cognitive bias modification interventions goes further, stating that the field itself "is overly reliant on small studies," the majority of which do not follow CONSORT reporting guidelines.

When multiple-comparisons correction is applied to small-sample data, nominally significant associations disappear. In the EEG study, exploratory correlations between alpha-power change and state-anxiety reduction (ρ ≈ −0.44, p = 0.028) and between beta-power change and depression-subscale reduction (ρ ≈ 0.43, p = 0.032) both failed to survive false discovery rate correction. Two intervention trials that were explicitly powered a priori also returned null primary results: a dyadic intervention trial found no significant group effects on RCADS-25 anxiety (p = 0.46) or depression (p = 0.23), and enrolled 390 of a planned 450 dyads (roughly 87% of target); a movement-based training quasi-experiment found no significant between-group difference in trait anxiety, with post-hoc power for its overall mental-health comparison at 0.49, below the conventional 0.80 standard.

Required sample sizes for adequate power vary widely by design and assumed effect size across these sources: 32 participants for 90% power to detect a Cohen's f = 0.44 effect on somatic anxiety at α = 0.01; 438 dyads (146 per arm) for 80% power to detect Cohen's d = 0.33; 52 participants as the minimum for an ANCOVA at 80% power with a medium effect (f = 0.25); 285 participants as the minimum training-set size cited for machine-learning models predicting anxiety outcomes to reach stable performance error; and 276 to 333 participants for causal mediation models at 80–90% power assuming a standardized coefficient of 0.2.

Where meta-analytic evidence exists, small-sample origin is linked to inflated effect sizes. In the cognitive bias modification meta-analysis, three of four statistical outliers had effect sizes almost ten times the pooled mean; effect size decreased by 0.07 per additional quality criterion satisfied for general anxiety outcomes and by 0.10 for depression outcomes; and a negative linear relationship between publication year and effect size meant older studies showed significantly higher effect sizes while more recent studies obtained effect sizes near zero or negative. A separate meta-analysis of music-performance-anxiety interventions likewise notes that small sample sizes across the available studies "substantially limit the statistical power of meta-regression analyses to detect small moderator effects," even while the same meta-analysis reports a moderate-to-large pooled reduction in anxiety outcomes (g = 0.732) and a small-to-moderate improvement in positive functioning (g = 0.35).

Some small-sample designs did detect measurable effects. Movement-based training was associated with a within-group decrease in trait anxiety (Cohen's d = 0.34), and psychological inflexibility was associated with increases in GAD symptoms (b = 0.28), anxiety symptoms (b = 0.07), and stress symptoms (b = 0.20) in a cross-sectional sample — though that source explicitly states the design "precludes causal inference." Validation-focused claims in this set concern instrument reliability rather than intervention power: the GAD-7 showed Cronbach's α = 0.94 and the DASS-21 anxiety and stress subscales showed α = 0.86 and 0.92, both described as consistent with prior validation studies. No claim in this set reports a formal post-hoc power calculation specific to each null intervention finding — the sources flag the limitation qualitatively rather than quantify it outcome-by-outcome.

Where the evidence splits

One sports-psychology source computes an a priori sample of 32 as sufficient for 90% power on somatic anxiety, then in the same paper's discussion states that its actual sample size and baseline variability "reduced statistical power across all psychological outcomes." The claims resolve part of this: the power calculation was based on somatic anxiety's effect size alone and "may not fully represent the required sample size for the other psychological variables (e.g., cognitive anxiety, self-confidence, emotional competence, and attention), which may differ in expected effect magnitudes and variability." The split is explained for outcome-to-outcome variation, but the sources do not say whether any of those other outcomes were in fact underpowered.

A movement-based training study reports overall design-level power of 0.80 for its planned ANCOVA (sample of 65 exceeding a required minimum of 52), but a specific comparison (adjusted post-intervention mental-health score) carried a post-hoc power of only 0.49. The claims note that "post-hoc power estimates can be misleading and reflect observed rather than true effect," but do not fully reconcile why a study judged adequately powered overall produced an underpowered specific comparison.

Thresholds and dose-response

Multiple sources establish explicit sample-size thresholds tied to a target power level: 32 participants for 90% power (α = 0.01, Cohen's f = 0.44, somatic anxiety); 438 dyads (146 per arm) for 80% power (α = 0.05, Cohen's d = 0.33); 52 participants minimum for 80% power on an ANCOVA (medium effect f = 0.25); 285 participants (70% training subset) as the minimum for machine-learning models predicting anxiety outcomes to reach stable performance error; and 276–333 participants for causal mediation models depending on target power of 0.8 to 0.9 (effect size 0.2, two covariates).

A graded, dose-response relationship is reported between methodological quality and measured effect size in the cognitive bias modification meta-analysis: effect size decreased by 0.07 for each additional quality criterion satisfied for general anxiety outcomes, and by 0.10 per criterion for depression outcomes.

Open empirical questions

• Whether outcome-specific power analyses (beyond somatic anxiety) would show cognitive anxiety, self-confidence, emotional competence, or attention outcomes were adequately or inadequately powered in the sports-psychology training study.
• Whether a larger EEG subcohort would sustain the alpha-power/state-anxiety and beta-power/depression associations that were nominally significant before FDR correction but not after.
• Whether the 87%-of-target enrollment in the dyadic intervention trial (390 of 450 planned) changed the trial's power enough to explain its null primary outcome, since no post-hoc power figure is given for that trial.
• Whether replication studies powered to detect the large effect sizes seen in early, small cognitive-bias-modification studies have since been conducted at adequate power, and what they found.