Advanced appraisal of a psychiatry meta-analysis and non-inferiority claim (MEQ)

(i) Diamond vs prediction interval and I-squared. The pooled RR 0.80 with CI 0.72–0.89 indicates a statistically significant average benefit under a random-effects model (CI excludes 1). The prediction interval 0.55–1.16 is wider and crosses 1, meaning that in a new similar setting/study the true effect could include no benefit — the key advanced distinction between average-effect precision and between-study predictive uncertainty.[1] I-squared 68% indicates a substantial proportion of variability is due to heterogeneity rather than chance and should trigger clinical and methodological exploration (dose, severity, blinding, attrition), not automatic rejection of all synthesis nor uncritical class-effect claims.[2]

(ii) GRADE downgrades and certainty. Starting from RCTs (high): downgrade for risk of bias (unclear concealment; unblinded raters on subjective outcomes — RoB 2 measurement and randomisation-process concerns).[7][3] Downgrade for inconsistency (I-squared 68%, PI crossing null).[2][3] Downgrade for publication bias / small-study effects (asymmetric funnel, Egger signal; psychiatry has known selective publication threats).[4][8] Imprecision is less dominant for the average effect (CI excludes 1) but the wide PI supports cautious language about applicability across settings. Overall certainty commonly low (or low–moderate if RoB is only 'some concerns' and search is comprehensive) for the average response effect — state your judgment with domain linkage rather than a bare label.[3]

(iii) Numbers for NNT counselling. Need a dichotomous patient-important outcome (response/remission), event rates or a chosen baseline risk (CER) for the patient before you, the relative effect (RR or OR) with uncertainty, and treatment duration. Absolute risk difference = CER − (CER × RR) for a harmful-event framing of non-response, or use response rates directly if available; NNT = 1/ARR. Relative meta-analytic RR alone is insufficient for honest NNT talk.[1][3]

(iv) Non-inferiority claim. Invalid. Non-inferiority requires a pre-specified clinically justified margin (delta), appropriate analysis sets, and a CI that lies entirely on the non-inferior side of delta; CONSORT non-inferiority extension items should be reported.[5] A superiority trial whose CI extends to +11% cannot be rebranded as NI without that design; residual possible inferiority of 11% may exceed a typical acceptable margin. Press-release language is spin until the protocol and CI-vs-delta are shown.[5]

(v) Applicability (older multimorbid adult). External validity is limited: exclusion of multimorbidity, polypharmacy, frailty, and pharmacokinetic change means baseline risk, harm profile, and absolute benefits may differ — an indirectness concern in GRADE terms and a classic Rothwell external-validity problem.[6][3] Shared decision-making should present low-certainty average efficacy, emphasise monitoring for adverse effects (hyponatraemia, falls, interactions), and avoid protocol-forced switching based on the diamond alone.