Absolute effects across a network share no baseline risk or time horizon
Statement
An absolute effect in a network meta-analysis, a number needed to treat, a risk difference or an absolute event rate, is manufactured by combining a relative effect with a baseline risk the analyst supplies. That baseline may be the mean control event rate across studies, a pooled estimate from a baseline model, or an elicited value, and it may be common to all comparisons or specific to each; the same network yields different absolute numbers under each choice. The trials informing different comparisons also have different follow-up durations, so two absolute effects in the same display can refer to different reference populations and to different time horizons at once. The displays in routine use place these numbers side by side as if they were commensurable: bar, Cates, forest, bubble, scatter and rank-heat formats, the last recommended precisely for showing many treatments across many outcomes in one figure. GRADE summary-of-findings guidance requires an assumed comparator risk and a time frame to be stated in the table, and the network adaptation of that table carries relative and absolute effects together, so part of the problem is addressed in tabular reporting. What no convention supplies is a requirement that every cell of a displayed set share one baseline and one horizon, or failing that be annotated with its own, and no checklist for network reports asks for either.
Why it is open
Standardizing to one baseline risk reintroduces the question of whose baseline: different comparisons are informed by different trial sets, so a single control event rate is an extrapolation for every comparison whose trials were run in populations at different underlying risk, and choosing it is a judgment the display then hides. Standardizing to one horizon requires either that every comparison have evidence at that horizon, which is generally false, or an extrapolation model, which imports exactly the assumptions an absolute-effect display was meant to make transparent. Annotation is the cheaper fix and fails on the graphical formats, since a rank-heat cell or a bubble is a single mark with no room for a baseline and a follow-up, and moving them to a caption defeats the purpose of a display built for rapid comparison. The problem also has no owner: it falls between the synthesis methods literature, which treats absolute effects as a presentational step after estimation, and the reporting guidelines, which specify what a table must contain rather than what a set of numbers must have in common.
What has been tried
| Work | What it contributes |
|---|---|
| Veroniki, Bender, Glasziou, Straus & Tricco 2019, Journal of Clinical Epidemiology 111:11-22 | Sets out six graphical methods for presenting the number needed to treat from a network meta-analysis and states the problem: the calculation requires a control event rate, different rates including the mean across studies, a pooled meta-analytic estimate and an expert-opinion value give different numbers, and because the number needed to treat depends on control event rate and study duration, comparing multiple treatments on that scale may be difficult. Advises caution rather than fixing a convention. |
| Veroniki, Straus, Fyraridis & Tricco 2016, Journal of Clinical Epidemiology 76:193-199 | Introduces the rank-heat plot for presenting many treatments across many outcomes in a single figure, the display format in which absolute effects computed from different baselines and horizons are most likely to be read as commensurable. |
| Yepes-Nunez, Li, Guyatt, Jack et al. 2019, Journal of Clinical Epidemiology 115:1-13 | Develops the GRADE summary-of-findings table for network meta-analysis through structured interviews with 32 stakeholders, with relative and absolute effect estimates, certainty and treatment ranking as required components. Establishes the tabular convention that partially covers this problem; it does not require that the absolute effects for different comparisons share a baseline risk and a follow-up. |
| Guyatt, Oxman, Santesso, Helfand et al. 2013, Journal of Clinical Epidemiology 66(2):158-172 | GRADE guidance on preparing summary-of-findings tables for binary outcomes, including how to choose and state assumed comparator risks and the follow-up to which they refer. This is the closest existing standard, written for pairwise comparisons rather than for a set of comparisons shown together. |
| Hutton, Salanti, Caldwell, Chaimani et al. 2015, Annals of Internal Medicine 162(11):777-784 | The PRISMA extension for network meta-analysis, the reporting checklist that would be the natural home for a baseline-and-horizon requirement and does not contain one. |
| Dias, Sutton, Ades & Welton 2013, Medical Decision Making 33(5):607-617 | The generalized linear modeling framework for evidence synthesis in which absolute effects are produced by combining a separately estimated baseline model with the pooled relative effects, which is where the baseline enters and where a common referent could be imposed. |
Probable solution or research direction
Adopt a default and make departures declare themselves: absolute effects displayed together should be computed from one named baseline risk and one named time horizon, with the source of that baseline stated as a pooled baseline model, an external registry estimate or an elicited value. Where a comparison has no evidence at the named horizon, mark the cell rather than filling it, and report the follow-up range of the contributing trials so a reader can see how far the number was carried. Extend network reporting checklists and the summary-of-findings format to require the baseline, its provenance and the horizon for every absolute quantity, including in figures, and specify a minimal annotation for dense displays such as a companion table keyed to the cells. Then test whether it matters: recompute published network absolute-effect displays under a common baseline and horizon and under the original comparison-specific choices, and report how often the implied treatment ordering on the absolute scale changes.
Verification
Verdict. Partially addressed. One reader proposed this and the reviewer accepted it as a specific commensurability problem rather than a restatement of the need to declare an estimand, since a declared absolute target outcome does not ensure that every displayed cell uses the same baseline-risk population and time horizon. It is partially-addressed because GRADE guidance and the network summary-of-findings format already require an assumed comparator risk and a time frame in tables; what nothing covers is commensurability across a displayed set or annotation in graphical formats.
Source
Derived from reading:nma-30 of the reviewed corpus.