Thursday, September 1, 2022

When is Imprecision "Very Serious" for a Systematic Review? New GRADE Guidance for Rating Down Two Levels Using a Minimally Contextualized Approach

Within the GRADE framework, the assessment of imprecision of an effect estimate relies on two major aspects: Optimal Information Size (OIS), which considers whether the sample size of the pooled estimate has adequate power to detect a minimally important difference, and the confidence interval (CI) approach, which requires inspecting the width of the 95% confidence interval and determining whether it crosses a meaningful threshold of effect (for instance, ranges from a null or trivial effect on one end to a potentially meaningful effect on the other).

GRADE users who are applying the framework to the assessment of the certainty of evidence in the context of a systematic review but not a practice guideline have, in previous guidance, been instructed to prioritize the OIS rather than the CI approach, which typically requires some judgment about a meaningful effect threshold, or "contextualization." However, given the fact that systematic review authors and readers can also benefit from the application of effect thresholds when assessing imprecision, new guidance has been published to help systematic reviewers apply a "minimally contextualized approach" and consider both reasons for rating down by one, two, or three levels if necessary. 

In addition to providing multiple examples in which rating down by two levels based on the span of the CI may be warranted, the paper also suggests circumstances in which systematic reviewers may rate down by two levels in the absence of concerns about the CI but in the presence of a particularly small information size. Specifically, when the CI does not overlap with threshold(s) of interest and the effect is sufficiently large, authors may still consider rating down for imprecision if the OIS (the sample size needed to detect an effect within one adequately powered trial as determined by a conventional power analysis) is not met. 

For dichotomous outcomes, in this case, one may consider rating down by two levels immediately if the ratio of the upper- to the lower-boundary of the CI is higher than 2.5 for odds ratios or 3.0 for relative risk ratios. If these criteria are not fulfilled, authors should still compare the pooled sample size to the determined OIS. For continuous outcomes, a total sample size that does not exceed 30-50% of the OIS is considered reason for rating down by two levels to "very serious" imprecision.

Finally, if a CI is so wide that it causes authors to be very uncertain about any estimate of effect, it may be acceptable to rate down by three levels for imprecision.

Zeng, L., Brignardello-Petersen, R., Hultcrantz, M., et al. (2022). GRADE Guidance 34: update on rating imprecision using a minimally contextualized approach. J Clin Epidemiol, online ahead of print. Manuscript available at the publisher's web site here.