Table I.

The GRADE criteria and the author's interpretations and conclusions.

GRADE criteria	Author's interpretation of the GRADE criteria	Author's conclusion of the GRADE criteria	GRADE criteria for assigning level of evidence	Author's conclusion of the GRADE criteria for assigning level of evidence
Criteria which may decrease confidence in results
Limitations to study quality (risk of bias)	Reflects the (lack of) internal validity of the study. The foremost quality criterion in science	Agree with the criterion	Decrease with 1 or 2 levels if serious limitations (–1) or very serious limitations (–2) to study quality	Decreasing with 1 or 2 levels is appropriate. In some cases decreasing with even 3 levels (e.g. evidence from high to very low) is justified
Inconsistency	Reflects the (lack of) consistency of the results of a study	Agree with the criterion	Decrease with 1 or 2 levels if inconsistency is serious (–1) or very serious (–2)	Decreasing with 1 or 2 levels is appropriate
Indirectness of evidence	All studies synthesized in a systematic review should have similar patient populations, interventions, control interventions, and outcomes. It is not appropriate to do a meta-analysis combining direct and indirect evidence posing different hypotheses, except in network meta-analyses	Mostly disagree with the criterion	Decrease with 1 or 2 levels if serious (–1) or very serious indirectness (–2)	Decreasing level of evidence is not appropriate when based on summarizing results from incommensurable studies. Network meta-analyses may allow decisions for decreasing level of evidence based on indirectness
Imprecision	Reflects random error in outcome estimates. The wideness of confidence intervals is one result of a study or meta-analysis and should not be used as a quality criterion	Disagree with the criterion	Decrease with 1 or 2 levels if serious imprecision (–1) or very serious imprecision (–2)	Decreasing level of evidence based on degree of random error in the outcome estimates is not appropriate; the limitation shown by wide confidence intervals is a result of a systematic review
Probability of publication bias	Selective reporting of outcomes is a matter of internal validity of the study and belongs to the ‘limitations to study quality’ criterion. When individual studies are not at all published, the results of a systematic review are potentially biased	Agree with the criterion	Decrease with 1 or 2 levels if publication bias likely (–1) or very likely (–2)	Decreasing level of evidence is appropriate
Criteria which may increase confidence in results
Large magnitude of effect	This is a result of a study or meta-analysis and should not be used as a quality criterion. Large magnitude of effect may imply a high risk of biased results rather than increased confidence in results	Disagree with the criterion	Increase with 1 or 2 levels if large (+ 1) or very large (+ 2) evidence of association	Increasing level of evidence based on large magnitude of effect is not appropriate, because of a risk for biased conclusions
Dose-response gradient	Dose-response gradient often exists in studies assessing etiology of disease, but effectiveness of an intervention usually does not show a linear dose-response pattern	Mostly disagree with the criterion	Increase with 1 level if evidence of a dose response gradient (+ 1)	Increasing level of evidence based on a dose-response gradient is rarely appropriate when assessing effectiveness of an intervention
Residual confounding would further support inferences regarding treatment effect	If some plausible confounders have not been documented, there is no credible way to determine how adjusting these parameters would alter the effectiveness estimates.	Disagree with the criterion	Increase with 1 level if all plausible confounders would reduce a demonstrated effect (+ 1) or would suggest a spurious effect if no effect was observed (+ 1).	Increasing level of evidence is not appropriate because the confounders cannot be documented. Consequently there is a risk for biased conclusions