IQWiG General Methods on Meta-Analysis: What Germany’s HTA Institute Expects from Your Evidence Synthesis

Germany’s AMNOG process assesses the added benefit of medicines with new active ingredients under Section 35a of Book V of the Social Code (SGB V). The Federal Joint Committee (G-BA) sets the appropriate comparator therapy (zweckmäßige Vergleichstherapie, zVT) and makes the final decision, usually after commissioning an IQWiG assessment. In this health technology assessment, a meta-analysis is evidence, not a verdict.

IQWiG’s General Methods, Version 8.0, took effect in December 2025, with the English version published on 1 July 2026 [1, 5]. Section 10.3.7 covers meta-analyses, including the standard procedure for very few studies, and Section 10.3.8 covers indirect comparisons [1]. The Methods set general principles, not a rulebook for every dossier: read them alongside G-BA requirements and assessment-specific guidance.

Establish comparability before pooling. The protocol should fix population, intervention, comparator, outcomes, follow-up and study designs, with reproducible search, screening, extraction and risk-of-bias methods; PRISMA 2020 supports transparent reporting but does not replace IQWiG’s requirements [3]. Examine baseline risk, dose, background therapy and outcome definitions: a non-significant heterogeneity test does not prove clinical comparability.

A fixed-effect model estimates a common effect; a random-effects model estimates the average of a distribution of study effects. Choose from the scientific question, not a heterogeneity test. With few studies, uncertainty in τ², the between-study variance, can move both the average effect and its interval, so report τ² and justify the estimator.

These rows summarise IQWiG’s approach to the very-few-studies setting, not universal rules for every meta-analysis or AMNOG dossier.

With three or four studies, heterogeneity is hard to estimate, and a τ² of zero may reflect too little information rather than identical effects. IQWiG therefore runs a Bayesian random-effects meta-analysis with heterogeneity priors drawn from earlier benefit assessments and compares it with a qualitative synthesis. Agreement supports a quantified result; disagreement calls for investigation, not selective reporting. Our explainer on Bayesian random-effects models covers the mechanics.

In January 2024, IQWiG reported that this approach cut the share of three- or four-study meta-analyses with non-quantifiable results from 23% to 6% [2]. That figure describes the method’s output in that setting; it does not guarantee a quantifiable result, fewer failed dossiers or a favourable decision.

IQWiG evaluates patient-relevant outcomes: mortality, morbidity, health-related quality of life (HRQoL) and adverse events, each interpreted on its own.

• A favourable composite endpoint does not show benefit on every component; analyse clinically important components separately where data allow.
• For HRQoL and patient-reported outcomes (PROs), assess validity, scoring, responsiveness and missing data.
• A responder analysis needs a threshold suited to the instrument and indication; statistical significance or a generic cut-off does not establish relevance, and dichotomising discards information.

A risk ratio is usually easier to communicate than an odds ratio. A hazard ratio compares event rates over follow-up, not the absolute risk difference at a time point, so give absolute risks at a clinically meaningful time where you can. A standardized mean difference around 0.2 is not automatically relevant; interpretation depends on the outcome and IQWiG’s criteria.

A direct randomised comparison against the G-BA comparator carries most weight when its population, outcomes and follow-up fit. Without one, an indirect treatment comparison may estimate relative effects: the Bucher method anchors the comparison on a common comparator [4], and network meta-analysis extends this to larger networks when its assumptions hold.

• Transitivity: trials must be comparable in important effect modifiers.
• Consistency: where direct and indirect evidence exist for the same comparison, they should agree; unexplained disagreement lowers confidence.
• Outcome comparability: different endpoint definitions, follow-up or analysis populations can make an estimate unusable.

Matching-adjusted indirect comparison (MAIC) and simulated treatment comparison (STC) adjust for observed characteristics only; unmeasured confounding remains. Unanchored comparisons rest on strong assumptions and are not equivalent to head-to-head trials. Map the comparator to your evidence network early. Our German ITC guide and population-adjusted indirect comparisons page go further.