GRADE: Judging Certainty of Evidence and Strength of Recommendations

Two separate questions

Certainty of evidence asks how confident we are that the estimated effect is close to the true effect. Recommendation strength asks whether the balance of benefits and harms, values, resources, equity and feasibility supports recommending an action. Strong recommendations can sometimes rest on low-certainty evidence, but the rationale must be explicit.

Certainty categories and domains

• GRADE commonly uses high, moderate, low and very-low certainty. Randomized trials usually start high and observational evidence usually starts low, but study design is not the only consideration.
• Downgrade when there are important concerns about risk of bias, inconsistency, indirectness, imprecision or publication bias. Upgrading observational evidence may be justified in defined circumstances, such as a very large effect, dose-response relationship or plausible residual confounding that would reduce an observed effect.
• Assess each critical outcome separately; certainty may differ for benefits, harms and quality of life.

Apply it to clinical decisions

• Check whether the population, intervention, comparator, outcome and follow-up match the question. A statistically significant result is not automatically clinically important.
• Consider absolute effects, confidence intervals, adverse events, patient preferences, cost and feasibility. Use MCID or other meaningful thresholds where justified, and avoid calling an estimate precise when its interval includes important benefit and harm.
• Explain uncertainty plainly to patients and teams; uncertainty is not the same as evidence of no effect.

Common mistakes and summary

• Do not label a study 'high quality' solely because it is randomized or a review includes many studies.
• Do not convert certainty ratings mechanically into a treatment rule.
• GRADE makes uncertainty visible and supports transparent decisions; it does not remove the need for clinical judgment and shared decision-making.