High, Moderate, Low, Very Low: How GRADE Rates the Certainty of Evidence

GRADE sorts medical evidence into four tiers, high, moderate, low, and very low certainty, based on how confident we can be that a study's effect estimate reflects the truth. Randomized trials start high and get downgraded across five domains. The label tells you how much a recommendation's confidence is actually earned. When a guideline reports "low certainty," it is signaling that the true effect may turn out to be substantially different from the number in the abstract.

What the four levels mean

GRADE is the Grading of Recommendations, Assessment, Development and Evaluation approach. More than 110 organizations use it, including Cochrane, the World Health Organization, and the UK's NICE, which is why the same four words keep showing up at the bottom of guideline tables. The framework was consolidated in a widely cited 2008 BMJ paper by Guyatt and colleagues, and Cochrane's Handbook Chapter 14 remains the working reference for how reviewers apply it.

The levels describe confidence in an effect estimate, not the size of the effect. High certainty means reviewers are very confident the true effect lies close to the estimate. Moderate means the estimate is probably close, but there is a real chance it differs. Low means confidence is limited and the true effect may be substantially different. Very low means the estimate is little more than a starting guess. A drug can have a large benefit rated at low certainty, and a tiny benefit rated at high certainty. Certainty and magnitude are separate axes.

Where the evidence starts

GRADE assigns a starting rating by study design. Randomized controlled trials begin at high certainty, because randomization, done well, balances known and unknown confounders across groups. Non-randomized and observational studies begin at low certainty, because, as Cochrane puts it, of the potential bias from the lack of randomization, meaning confounding and selection bias. From that starting point, a body of evidence moves down for problems and, in specific cases, back up.

The five reasons certainty gets downgraded

Risk of bias

This is about how the studies were run. Failure to conceal allocation, lack of blinding, high dropout, or selective outcome reporting all let bias creep into the result. If most of the trials feeding an estimate share these flaws, reviewers drop the rating one or two levels. A large trial with sloppy methods does not rescue certainty; volume cannot fix bias.

Inconsistency

When studies of the same question point in different directions, or their effect sizes scatter far more than chance would predict, that unexplained heterogeneity is a warning. It means some hidden difference, in the population, the dose, the outcome definition, is driving divergent answers, so any single pooled number is shakier than it looks. Consistent results across independent teams, by contrast, are reassuring.

Indirectness

Evidence is indirect when it does not quite match the question being asked. A trial in hospitalized 70-year-olds is indirect evidence for healthy 40-year-olds. A trial that measured a surrogate marker, such as a lab value, is indirect evidence for what patients care about, such as heart attacks or survival. Head-to-head comparisons that never happened, and were instead inferred across separate trials, are also indirect. Each gap costs certainty.

Imprecision

This is the confidence-interval problem. Few participants or few events produce wide intervals, and a wide interval can straddle both meaningful benefit and no benefit, or even harm. If the range of plausible truth includes outcomes that would lead to opposite decisions, the estimate is imprecise, and the rating falls. This is the domain that most often drags an otherwise clean body of small trials down to low certainty.

Publication bias

Studies with disappointing results are less likely to be published, so the literature that survives can overstate a benefit. When reviewers see signs of this, an asymmetric funnel plot, a field dominated by small industry-funded trials, a suspicious absence of null findings, they downgrade for the studies that probably exist but were never printed.

When observational evidence moves up

Observational studies can be upgraded in three situations. A large effect, on the order of a risk roughly doubled or halved with no plausible confounder to explain it, raises confidence, which is how the link between smoking and lung cancer earned strong standing without a randomized trial. A dose-response gradient, where more exposure tracks with more effect, adds credibility. And when every plausible bias would have worked to shrink the observed effect, yet an effect still appears, the real effect is likely at least as large as reported. These moves rarely lift observational evidence to high, but they can carry it to moderate.

Reading a strong versus conditional recommendation

Certainty is only half the story. GRADE separates the certainty of evidence from the strength of the recommendation. A strong recommendation means the desirable effects clearly outweigh the undesirable ones for almost everyone, so most people would want the option. A conditional (or weak) recommendation means the balance is close, or the evidence is uncertain, and the right choice depends on individual values and circumstances.

Higher certainty tends to support stronger recommendations, and GRADE guidance warns against issuing strong recommendations on low or very low certainty evidence. Yet it happens. A 2023 cross-sectional analysis in BMC Medical Research Methodology reviewed a suite of national clinical guidelines and found strong recommendations frequently attached to low or very low certainty evidence, sometimes justified by GRADE's recognized exceptions such as life-threatening situations, and sometimes not. That mismatch is the single most useful thing to check: a strong recommendation built on low certainty is a claim to read with care, because the underlying effect could still shift.

This article is educational and is not medical advice; decisions about your own care belong with a clinician who knows your history. What GRADE gives the rest of us is a shared vocabulary for a question worth asking of any health claim. Beyond what a study reported, it lets us ask how much we should trust that the finding is true.