Can I downgrade twice for imprecision when there are very few events?

Yes. A small number of events is one of the clearest cases for a two-level downgrade, particularly where the confidence interval spans both appreciable benefit and appreciable harm. State the event count in the rationale, since it is the fact doing the work.

Should imprecision be judged on the relative or absolute effect?

Assess it on the measure that supports the decision, which is usually the absolute effect. A relative risk of 0.80 means something very different at a 40% baseline risk than at a 0.4% baseline risk, and the Summary of Findings table should present absolute effects for exactly this reason.

Do I need to calculate an OIS for every outcome?

Not for every outcome, but for the outcomes in the Summary of Findings table it is the more informative of the two checks and it is quick. Where you do not calculate one, base the judgement on the decision threshold and say that is what you did rather than leaving the basis unstated.

Downgrading for Imprecision Without Guessing

8 September 20264 min

Of the five GRADE domains, imprecision is the one most often decided by feel. The confidence interval looks wide, the reviewer downgrades one level, and the rationale field says "imprecision".

That is not a rationale. It restates the domain name, and it gives a reader no way to check the judgement or to disagree with it.

The judgement has two components, and both can be stated explicitly.

Component one: does the interval cross a decision threshold?

The question is not whether the confidence interval is wide in the abstract. It is whether the interval spans values that would lead to different clinical decisions.

A pooled risk ratio of 0.85 with a 95% CI of 0.72 to 0.99 is statistically significant but the upper bound is close to no effect. Whether that warrants downgrading depends on where the threshold for a worthwhile effect sits. If a 5% relative reduction would change practice, the interval is entirely on the useful side and imprecision is not a problem. If clinicians would only act on a 20% reduction, the interval spans "act" and "do not act", and downgrading is justified.

Which means you need the threshold, and the threshold should be stated at protocol stage. It can come from a minimal clinically important difference in the literature, from guideline recommendations, or from a reasoned judgement by the review team. What it should not do is appear for the first time in the certainty assessment, chosen with the confidence interval already in view.

For continuous outcomes the same logic applies using the minimal important difference on the scale in question, or around 0.2 to 0.5 standardised mean difference units where no scale-specific MID exists, stated as an assumption.

Component two: is there enough information?

The second check is the optimal information size: whether the total sample across the pooled studies reaches the number a single adequately powered trial would require to detect the effect of interest.

Calculate the sample size a trial would need given a plausible control-group event rate, the smallest effect you would consider important, alpha of 0.05 and 80% power. Compare it to the total number of participants contributing to your meta-analysis. If the total falls short, downgrade for imprecision even where the confidence interval looks reassuring.

This check catches a specific and common situation: a meta-analysis of six small trials producing a narrow interval by chance. The interval says precision; the accumulated evidence does not support it.

Guyatt and colleagues (2011) set out both components in GRADE guidelines 6, and it remains the reference to cite when a reviewer challenges the judgement.

One level or two

Downgrade one level when the interval crosses a decision threshold or the OIS is not met. Downgrade two when the imprecision is severe: very few events, an interval spanning appreciable benefit and appreciable harm, or a total sample far below the OIS.

A working guide for the two-level case is an interval consistent with both a meaningful benefit and a meaningful harm, on evidence from a handful of small studies. Whichever you choose, write down which condition triggered it.

What a rationale looks like

Weak: "Downgraded one level for imprecision."

Defensible: "Downgraded one level for imprecision. The 95% CI (RR 0.72 to 0.99) includes values close to no effect, and the prespecified threshold for a worthwhile reduction was 15%. The pooled sample of 1,240 participants also falls short of the optimal information size of 2,100 calculated for an 8% control event rate, 15% relative reduction, alpha 0.05, power 80%."

Three sentences, and every element is checkable.

The software now asks

RevMan 11.8.0, released on 25 August 2026, raises a validation warning when a certainty-of-evidence domain has no documented rationale. Anyone with a Cochrane review open will meet it.

This is a small change with a useful effect. It moves the discovery of an undocumented downgrade from peer review, where reconstructing the reasoning months later is painful, to the point where the judgement is being made and the reasoning is still in your head.

References

Guyatt, G. H., Oxman, A. D., Kunz, R., Brozek, J., Alonso-Coello, P., Rind, D., Devereaux, P. J., Montori, V. M., Freyschuss, B., Vist, G., Jaeschke, R., Williams, J. W., Murad, M. H., Sinclair, D., Falck-Ytter, Y., Meerpohl, J., Whittington, C., Thorlund, K., Andrews, J., & Schünemann, H. J. (2011). GRADE guidelines 6. Rating the quality of evidence—imprecision. Journal of Clinical Epidemiology, 64(12), 1283–1293. https://doi.org/10.1016/j.jclinepi.2011.01.012

Common questions

Can I downgrade twice for imprecision when there are very few events?: Yes. A small number of events is one of the clearest cases for a two-level downgrade, particularly where the confidence interval spans both appreciable benefit and appreciable harm. State the event count in the rationale, since it is the fact doing the work.
Should imprecision be judged on the relative or absolute effect?: Assess it on the measure that supports the decision, which is usually the absolute effect. A relative risk of 0.80 means something very different at a 40% baseline risk than at a 0.4% baseline risk, and the Summary of Findings table should present absolute effects for exactly this reason.
Do I need to calculate an OIS for every outcome?: Not for every outcome, but for the outcomes in the Summary of Findings table it is the more informative of the two checks and it is quick. Where you do not calculate one, base the judgement on the decision threshold and say that is what you did rather than leaving the basis unstated.