227 OPERAM: cluster randomised controlled trial cluster correlation coefficient (ICC) of 0.02 [46], and variable cluster sizes from 12 to 38 (mean 25) patients [28,47]. The 30% relative risk reduction was based on assessment of the effect that we did not want to miss [48]. The primary analysis was performed according to intention-to-treat, including all clusters and patients in the allocated groups. The between-group difference for the primary outcome was analysed using a mixed-effects Cox proportional hazards model with a fixed effect for the intervention group and random effects for site and attending hospital physician [49,50]. Patients were censored at death to calculate cause-specific HRs. An additional analysis used extensions of the FineGray proportional hazards model that accounts for clustering in competing risk settings, treating death as the competing event to calculate subdistribution HRs [49]. Statistics were reported with their respective 95% confidence intervals (CIs) and p-values. All-cause deaths, cancer deaths, all-cause hospitalizations, falls and preventable drug-related hospital admissions were analysed similarly. Between-group differences for in-hospital death, drug-drug interaction, and drug overuse/underuse/misuse were analysed using mixed-effects logistic regression with fixed and random effects as above. Between-group differences for continuous outcomes were analysed using mixed-effects linear regression models with fixed and random effects as above, and adjustment for baseline values. Pre-specified subgroup analyses considered sex, age (<80 years vs. ≥80 years), home accommodation (independently living versus non-independently living), presence of dementia, number of drugs (<10 per day versus ≥10 per day), number of comorbidities (<median versus ≥median), cluster specialty (medical versus surgical) and trial site. Pre-specified sensitivity analyses adjusted for baseline characteristics and investigated time variation of the intervention effect [51]. A post-hoc added sensitivity analysis only considered data collected in interviews conducted within protocol-specified time windows. Per-protocol analyses were performed for time-to-first-event outcomes, omitting attending hospital physicians and patients with pre-defined protocol deviations (allocated intervention not received, cluster size <5 patients, violated inclusion or exclusion criteria) and intervention group patients for whom none of the STOPP/ START recommendations were implemented at month 2 [52]. The detailed statistical analysis plan is described in a supplement. All analyses were performed using R version 3.6.0. software [53]. 3
RkJQdWJsaXNoZXIy MTk4NDMw