GRADE (Grading of Recommendations Assessment, Development and Evaluation)

GRADE is a systematic and transparent framework for rating the certainty (quality) of a body of evidence and for developing and presenting recommendations in health care. Developed by the GRADE Working Group, it provides explicit criteria for assessing whether the evidence justifies confidence in an estimated effect and for translating that evidence into actionable clinical, public health, or policy recommendations. Over 120 organizations worldwide, including the World Health Organization and Cochrane, have adopted GRADE.

Before GRADE, systems for evaluating evidence quality varied widely, were often implicit or inconsistent, and frequently relied on study design alone (placing randomized trials at the top and observational studies at the bottom without further nuance). GRADE addressed these limitations by creating a standardized, internationally recognized framework that considers multiple factors beyond study design. This transparency allows guideline users (clinicians, patients, policymakers) to understand exactly why a recommendation is strong or weak and how confident the guideline panel is in the supporting evidence. GRADE is widely used for evidence evaluation in health care guideline development, and its principles have also been adopted in environmental science, veterinary medicine, and other fields.

GRADE operates at two levels: rating the certainty of evidence and formulating recommendations.

For certainty of evidence, GRADE begins by categorizing the body of evidence for each outcome as high (starting with randomized trials) or low (starting with observational studies). One exception: when nonrandomized studies are assessed with ROBINS-I, a risk of bias tool that judges confounding and selection bias directly, their evidence starts at high, and those problems are reflected in the risk of bias rating instead. Because certainty is rated outcome by outcome, one review can hold high certainty evidence for one outcome and low certainty evidence for another. GRADE originally called this rating the quality of evidence; current GRADE guidance calls it the certainty of evidence. From its starting level, the certainty can be rated down one level for a serious concern, or two levels for a very serious one, in each of five factors: risk of bias (methodological limitations in the studies), inconsistency (unexplained heterogeneity across studies), indirectness (differences between the evidence and the question of interest in population, intervention, comparator, or outcome), imprecision (few participants or events, or a confidence interval wide enough to include effects that would lead to different decisions, such as both important benefit and no important effect), and publication bias (a strong suspicion that studies have been selectively published, for example small studies with unfavorable results missing; selective reporting of outcomes within studies is judged under risk of bias). Conversely, certainty can be upgraded for three factors, applied mainly to observational evidence that has not been rated down for any of the five factors: large effect (very large effect sizes that are unlikely to be explained by confounding), dose response gradient (a clear relationship between dose and effect), and plausible confounding that would reduce the observed effect (meaning the true effect may be even larger than observed).

The final certainty rating is expressed on four levels: high (very confident that the true effect lies close to the estimate), moderate (moderately confident: the true effect is likely to be close to the estimate, but may be substantially different), low (limited confidence: the true effect may be substantially different from the estimate), and very low (very little confidence: the true effect is likely to be substantially different from the estimate).

For recommendations, GRADE considers the certainty of evidence alongside the balance of benefits and harms, patient values and preferences, resource use, equity, acceptability, and feasibility. Recommendations are classified as strong (most informed patients would choose the recommended course of action, and only a small proportion would not) or conditional (the decision depends on individual circumstances). The direction (for or against) and strength of each recommendation are reported explicitly.

GRADE Summary of Findings tables and Evidence Profiles present the evidence in a structured format that displays the certainty assessment for each outcome, making the reasoning transparent and reproducible.

GRADE should be used when developing clinical practice guidelines, public health recommendations, or health policy decisions based on a body of evidence. It is appropriate after a systematic review has been conducted and the evidence for each relevant outcome has been identified and synthesized. GRADE can be applied to evidence from randomized trials, observational studies, or both. Cochrane's conduct standards (MECIR) require new Cochrane intervention reviews to use the GRADE considerations to assess the certainty of evidence for each outcome, usually presented in Summary of Findings tables, and many major guideline developers, including the World Health Organization, use GRADE.

A common error is applying GRADE mechanistically, treating it as a checklist to be completed quickly rather than a framework for careful deliberation. Each judgment (about risk of bias, inconsistency, imprecision, etc.) requires thoughtful consideration of the specific evidence. Another mistake is confusing the certainty of evidence with the strength of the recommendation; it is possible to make a strong recommendation based on low certainty evidence (when the consequences of inaction are severe) or a conditional recommendation based on high certainty evidence (when the benefits and harms are closely balanced). Inconsistent application across outcomes within the same guideline is another problem. Some panels apply GRADE only to their primary outcome and neglect to rate the evidence for harms, costs, and patient important secondary outcomes. Finally, failing to include patient values and preferences, resource use, and feasibility in the recommendation formulation stage reduces the comprehensiveness and applicability of the guideline.

A hypothetical guideline panel evaluates the evidence for using corticosteroids in hospitalized patients with severe pneumonia. The systematic review identifies 8 randomized trials (total n = 1,800). For mortality, the pooled risk ratio is 0.67 (95% CI: 0.52 to 0.86), based on 6 trials. The GRADE assessment rates the certainty as moderate (downgraded one level for inconsistency, I squared = 55%). For ICU length of stay, the mean difference is minus 2.1 days (95% CI: minus 3.2 to minus 1.0), rated as high certainty. For hyperglycemia (a harm), the risk ratio is 1.48 (95% CI: 1.18 to 1.86), rated as high certainty. Balancing the mortality benefit against the manageable harm of hyperglycemia, and considering that corticosteroids are inexpensive and widely available, the panel issues a strong recommendation for corticosteroid use in this population.