O'Shea 2014.

Methods	Cluster RCT.
Participants	304 long‐stay care home residents in Ireland living with dementia according to DSM‐IV or any other diagnosis by a clinician, nurses judgement, nurses records (or a combination) or prescribed any medication for AD. Mean age: 85.4 years.
Interventions	Intervention: group RT. Control: treatment as usual.
Outcomes	Quality of life: QoL‐AD. Behavioural: Cohen Mansfield Agitation Inventory. Mood‐related outcomes: CSDD. Carer: Modified ZBI.
Length and frequency of intervention	3‐4 sessions per week for a mean of 14 weeks (range 12‐17 weeks). Session duration unspecified.
Time points measured	Baseline and 18‐22 weeks' postrandomisation.
Number of participants who did not complete study	76/304 (25%).
Notes	Approximately 75% of care homes recorded close to the mean target of 3 or 4 sessions per week.
Risk of bias
Bias	Authors' judgement	Support for judgement
Random sequence generation (selection bias)	Low risk	Randomisation at the level of the long‐stay residential unit. Randomisation was on a ratio of 1:1 and was stratified by public and private residential units (one‐third public to two‐thirds private, reflecting the overall distribution of beds in the region).
Allocation concealment (selection bias)	Low risk	Concealment of group allocation achieved by giving the responsibility for sequence generation and group allocation to a researcher who was independent of the study and its investigators.
Blinding of outcome assessment (detection bias) All outcomes	Low risk	Research nurses involved in data generation and collection were blinded to group allocation of participating units. Data analysis undertaken by researchers and statisticians blinded to group allocation.
Incomplete outcome data (attrition bias) All outcomes	Low risk	25 residents lost to follow‐up in intervention group (18 died, 1 was transferred, 2 in hospital, 1 withdrew and 3 too ill to participate) and 27 in control group (18 died, 3 too ill, 2 in hospital and 1 was transferred). Paper reported that all results were insensitive to the inclusion of missing data using multivariate imputation by chained equations.
Selective reporting (reporting bias)	Low risk	All outcomes listed in the methods section and study protocol were reported and there was no evidence of selective outcome reporting.
Other bias	Low risk	Cluster RCT. 18 clusters (9 intervention and 9 control). Although study authors analysed data appropriately, review authors needed to extract data (from a table) that did not account for clustering. The study authors reported the ICC for each measure, which the review authors extracted and used to calculate the effective sample size to enter into Review Manager 5 (RevMan 2014). No evidence of recruitment bias or missing clusters after randomisation. Cluster‐specific baseline adjustment was implemented to prevent baseline imbalance creating bias.
Availability of training and supervision Objective outcome measures	Low risk	Staff training involved a structured education programme, facilitated by experienced nurse educators, delivered over 3 days and augmented by telephone support and onsite visits.
Availability of manual or protocol for intervention All outcomes	Low risk	A structured education programme for staff was delivered and staff were trained in intervention design. Staff target was 1 planned formal session and 3 spontaneous sessions per week.