Randomizing Facilities, Analyzing Patients

Residents within a nursing home are not independent: they share staff, prescribing culture, documentation practices, and infection-control conditions. If the protocol analyzes 480 residents as independent observations, the standard error will be too small and the type I error rate will be inflated. A simple quantification uses the design effect: DE = 1 + (m - 1)ICC, where m is cluster size. With m = 40 and an ICC of 0.05, DE = 1 + 39(0.05) = 2.95; the effective sample size is about 480 / 2.95 = 163 residents, not 480. Even with a modest ICC of 0.02, DE = 1.78 and the effective n is about 270. The analysis should use a mixed-effects model with a random intercept for nursing home, or a cluster-level summary analysis, and should adjust for baseline prescribing rates. With only 12 clusters, the cluster-level degrees of freedom are limited, so the trial is likely underpowered for plausible effects. The team should either increase the number of homes, reduce cluster size while increasing clusters, or pre-specify a cluster-level analysis with an honest power calculation that incorporates the ICC and baseline imbalance.

With 12 clusters, chance imbalance is a serious threat. The home with 90% antipsychotic use could dominate the intervention arm and bias the comparison. Strengthening options include stratified randomization by baseline potentially inappropriate medication rate or antipsychotic rate, or constrained randomization to ensure balance on key facility characteristics. The primary analysis should adjust for baseline facility-level prescribing rates and resident-level confounders such as age, dementia diagnosis, and length of stay. The team should pre-specify the primary estimand and avoid interpreting unadjusted resident-level differences. It should also report cluster-level outcomes and consider a run-in period to identify homes with stable prescribing. Because the intervention is delivered at the facility level, contamination is less likely between arms, but staff turnover and shared prescribers could still dilute the effect. A process evaluation should measure fidelity. Finally, the power calculation should reflect the number of clusters, not just the number of residents; if 12 clusters cannot detect a realistic effect, the trial should be redesigned before launch rather than analyzed with an inappropriate t-test.

A waiver of individual consent is not automatic just because the intervention is a facility policy. The ethics board should require that the research poses minimal risk, that the waiver will not adversely affect residents' rights or welfare, that individual consent is impracticable, and that residents will be provided with relevant information afterward when feasible. Because nursing-home residents are a vulnerable population, safeguards should include facility-level agreement from resident councils, an independent safety monitor, a clear opt-out mechanism for residents or substitute decision-makers, and protection of privacy during data extraction. The intervention changes prescribing workflows, so residents may receive fewer potentially inappropriate medications, but they may also lose medications they need. The protocol should specify how adverse drug withdrawal events will be detected and reported. Consent materials should be written in plain language and offered in the residents' preferred languages. If any resident-level data are collected beyond routine care, additional consent or a separate authorization may be required. The waiver should be time-limited and reviewed if the intervention is adopted beyond the study.