Death's neighbours: dichotomy placement and the death-to-dependency exchange in randomised trials of intracranial-pressure-lowering strategies — analysis specification, count tables and confirmatory code

Pre-results analysis record for a re-expression of published Glasgow Outcome Scale-Extended (GOS-E) distributions from randomised trials of strategies that lower intracranial pressure or the response to it: Eurotherm3235 (Andrews 2015), DECRA (Cooper 2011), RESCUEicp (Hutchinson 2016; 6 and 12 months) and BEST:TRIP (Chesnut 2012), with POLAR (Cooper 2018) as a pre-specified secondary stratum in collapsed 3-category form. NABIS:H II and Clifton 2001 are excluded with the reason stated (5-point GOS only). This record was deposited before any manuscript text was written. It contains the frozen analysis specification (v0.2, 17 September 2026), the count tables with the source table and text-line anchor for every cell, the confirmatory script (v1.0, frozen), its output unchanged, the software versions, licences and a SHA-256 manifest. A rerun of the script with the fixed seed reproduces the deposited output exactly; the proportional-odds odds ratios it computes reproduce the trials' published ordinal odds ratios, as a transcription check. The analysis quantifies how the sign and significance of each trial's functional-outcome result depend on where the unfavourable/favourable dichotomy is placed relative to death (four cuts used in the literature: GOS 1–3, GOS-E 1–4, GOS-E 1–5, GOS-E ≥ 4), decomposes between-arm mortality differences into their destination categories (vegetative state or lower severe disability; upper severe disability; moderate disability or better), reports win probabilities and an exchange rate (extra dependent survivors per death averted) on the randomised denominators, and states with proof the proposition that a transfer of deaths into category j is registered as benefit only by cuts strictly below j. Published aggregate data only; no individual participant data; no new participants; no trial author was contacted. The deposit contains no publisher PDFs, no extracted article text and no reproduced abstracts; primary reports must be obtained from their publishers. Counts were checked against the page images of the source tables by two authors before deposit. Licence: text and data CC BY 4.0; code (the code/ directory) MIT. AI tools (Claude, Anthropic) assisted with source extraction, statistical implementation, cross-checking and drafting under the authors' direction; this is disclosed in the manuscript. The named authors take responsibility for the content.