Draft methodological review: transitivity in the incretin… — submissions
The 9 submissions received, published in full with declared interests and secretariat responses.
§2Submissions and responses
9 submissions were received. Each is published in full below with its declared interest, the secretariat response and the disposition. The Institute publishes submissions it did not accept in the same form as those it did.
Results at materially different follow-up durations are pooled
The respondent submits on the draft synthesis addressing Are the trials in the incretin evidence network sufficiently similar to support indirect comparison?. This submission concerns the methods rather than the conclusion, on the view that a conclusion is only as good as the search that produced it.
Contributing trials report the outcome at durations ranging across more than a year. The draft pools them into a single estimate. The respondent states that a mean at one duration and a mean at another are not estimates of the same quantity when the effect is still changing.
The respondent proposes that estimates be reported by duration band.
The secretariat accepts this submission. Pooling across durations assumes a plateau the contributing trials do not demonstrate.
Estimates are now reported by duration band, with the number of contributing trials and participants in each band stated, and no estimate is pooled across bands.
Regulatory assessment documents were not searched
Having read the draft synthesis on Are the trials in the incretin evidence network sufficiently similar to support indirect comparison?, the respondent puts one point to the committee.
The respondent states that regulatory assessment reports frequently contain analyses that never appear in journals, and that a search limited to bibliographic databases will miss them.
The respondent proposes that regulatory documents be searched for every review in the series.
The respondent has read submission 001 and asks that this submission be considered with it.
The secretariat accepts this submission in part. Regulatory assessment documents are searched. The proposal that they be treated as equivalent to a full study report is declined, because the level of detail varies and is often insufficient for risk-of-bias assessment.
Published regulatory assessment documents are now searched as a named source, are reported as a distinct evidence class in the included-studies table, and contribute to the assessment while being excluded from pooled estimates where risk of bias could not be assessed from the document.
Statistical heterogeneity is treated as though it measured clinical heterogeneity
The respondent has read the draft synthesis on Are the trials in the incretin evidence network sufficiently similar to support indirect comparison? and makes one submission.
The draft reports a heterogeneity statistic and proceeds to pool where it is low. The respondent states that a low statistic in a small set of trials is uninformative, and that clinical and methodological similarity should be assessed before any statistic is consulted.
The respondent proposes that the decision to pool be justified on clinical grounds first and that the statistic be reported as a description rather than used as a threshold.
The secretariat accepts this submission in part. The decision to pool is now made on clinical and methodological grounds and stated as such. The statistic continues to be reported, because readers expect it and its absence would be read as concealment.
The decision to pool is now justified in prose against the population, intervention, comparator and outcome before any statistic is presented, and the heterogeneity statistic is reported with its confidence interval and a note of the number of contributing studies.
The review protocol is described but its registration record is not linked
The respondent’s comment on the draft review of Are the trials in the incretin evidence network sufficiently similar to support indirect comparison? arises from comparing the included set against the respondent’s own knowledge of the field.
The respondent asks for the prospective registration record of the review protocol so that the registered outcomes can be compared with the reported ones.
The respondent states that without it the claim of prospective registration cannot be checked.
The secretariat accepts this submission in part. The protocol is published in full on the Institute site with its version history, which permits the comparison the respondent asks for. Where an external registration identifier is not held by the Institute, the review says so rather than supplying one.
Each synthesis now links its own protocol with the version in force at screening, and states explicitly where an external registration identifier is not held, so that the absence is a recorded fact.
Intention-to-treat and efficacy-estimand results are combined without distinction
The respondent read the draft review of Are the trials in the incretin evidence network sufficiently similar to support indirect comparison? and has confined this submission to a single matter.
Several contributing trials report both a treatment-policy result and an efficacy result that censors at discontinuation. The draft draws from whichever is reported first in each paper.
The respondent, a trial statistician, states that the two answer different questions, that the difference is substantial where discontinuation is common, and that a synthesis should choose one and say which.
The respondent’s submission overlaps with submission 001 and was prepared without sight of it.
The secretariat accepts this submission. Mixing estimands within a single pooled estimate is a defect of the synthesis rather than of the trials.
The treatment-policy estimand is used throughout as the primary analysis, the efficacy estimand is reported as a secondary analysis where available, and the estimand used is stated in every row of the summary of findings.
The timepoint at which each outcome was extracted was chosen after the data were seen
This submission addresses the draft synthesis on Are the trials in the incretin evidence network sufficiently similar to support indirect comparison? from the standpoint of a reader who will use the summary and not the appendices.
The methods state that outcomes were extracted at the latest available timepoint. Several contributing trials report the same outcome at three timepoints, and taking the latest is a choice that can be made in the knowledge of what each shows.
The respondent proposes that the extraction timepoint be prespecified in the protocol, and that where it was not, the review report the estimate at every reported timepoint so that the choice is visible.
The secretariat accepts this submission. An extraction rule applied after the data are visible is a degree of freedom and should be recorded as one.
The extraction timepoint is now prespecified in the protocol for new reviews. Where a review predates the requirement, estimates are reported at every timepoint the contributing trials report.
Absence of evidence is presented in a form a reader will take as negative evidence
This submission concerns the draft review of Are the trials in the incretin evidence network sufficiently similar to support indirect comparison?. The respondent has conducted reviews on adjacent questions and the observation arises from that experience.
Where the Institute has identified no study, the draft states that no evidence was found. In several places that sentence sits immediately after a paragraph describing an adverse outcome, and the juxtaposition invites the reading that the compound was studied and found wanting.
The respondent proposes a standing formulation, used identically wherever the situation arises, distinguishing an outcome that was studied and not demonstrated from an outcome that has not been studied at all.
The secretariat accepts this submission. The two states are different, they support different decisions, and the draft rendered them in language a reader could not reliably separate.
A standing formulation has been adopted and is applied wherever an evidence gap is recorded, distinguishing an outcome assessed and not demonstrated from an outcome not assessed. The formulation is identical at every occurrence so that it can be recognised at a glance.
Doses differing several-fold are pooled without examining dose-response
The respondent notes that the review question — Are the trials in the incretin evidence network sufficiently similar to support indirect comparison? — is answerable only if the contributing trials measured the same thing, and submits with that in view.
Contributing trials administer doses that differ by a factor of several. The draft pools them and reports one estimate. The respondent states that where a dose-response relationship exists the pooled figure corresponds to no dose that anyone receives.
The respondent proposes that estimates be reported by dose and that dose-response be examined where the data allow.
The secretariat accepts this submission. A pooled estimate across doses is an estimate for an average dose that no protocol specifies.
Estimates are now reported by dose group, a dose-response examination is reported where three or more dose levels contribute, and the pooled across-dose estimate is removed rather than retained alongside.
Overlapping primary studies across included reviews are counted more than once
The draft review of Are the trials in the incretin evidence network sufficiently similar to support indirect comparison? was read against its registered protocol.
The draft is an overview of reviews and several included reviews share primary studies. The respondent states that the overview reports the total number of participants across reviews as though the sets were disjoint.
The respondent proposes that overlap be assessed and reported, and that no total be given without correcting for it.
The secretariat accepts this submission. An inflated participant total overstates the evidence base by an amount the reader cannot see.
Overlap between included reviews is now assessed at primary-study level and reported in a citation matrix, and participant totals are reported for the union of primary studies rather than as a sum across reviews.
References cited on this page
References are numbered in order of first citation in this document. Each superscript in the text links to its entry below.
- International Organization for Standardization. ISO/IEC 17025:2017 General Requirements for the Competence of Testing and Calibration Laboratories. ISO/IEC Standard 2017;3rd edition. identifier not held by the Institute
Identifiers are reproduced only where the Institute holds them. Where a digital object identifier or PubMed identifier is not shown, the Institute has recorded the journal and year and has not constructed an identifier.