Leave One Out

Briefing

Kastner reads Corvane's forest plot aloud for the third time. Eight trials, a pooled odds ratio for hospitalization below one, an interval that just clears the null. "Robust," the slide says. Brook is not so sure: one trial, he says, did most of the work. Vale slides the trial file across the table and asks the question the Board will ask tomorrow: how robust is robust?

Specification

• Data: quayside_trials.rds, the same eight trials as the rest of Case E2.
• Outcome: heart failure hospitalization within 12 months; odds ratio, Velquor versus control; below 1 favors Velquor.
• Model: random effects meta-analysis of log odds ratios with Woolf variances and Wald type 95 percent confidence intervals.
• The trial: with all eight trials the 95 percent CI excludes 1. Exactly one trial, when omitted, gives a CI that includes 1, and this holds under the DerSimonian-Laird, REML, and Paule-Mandel estimators alike.
• OR: the pooled random effects OR without that trial, with the REML estimator.
• W: that trial's weight in the inverse variance fixed effect analysis of all eight trials, in percent.

Flag

epi{STUDY_OR_W}

Format example: epi{quayside-x_0.91_12.4}. Unrounded numbers are fine: the preview shows what will be checked.

Exhibits

R files can run code when loaded in R older than 4.4.0. Use R 4.4.0 or later and check the checksum before readRDS() or load().