Krogh 2009.

Methods	RCT (parallel group)
Participants	Referred from general practitioners, private psychiatrists, psychologists and psychiatric wards institutions. Included if met criteria for major depression Mean age 38.9 73.9% women N = 165
Interventions	1. Strength circuit training (n = 55) 2. Aerobic (machine‐based) training (n = 55) 3. Relaxation control (n = 55 Twice‐weekly intervention for 32 sessions delivered over a 4‐month period
Outcomes	Hamilton Rating Scale for Depression
Notes	Intention‐to‐treat analysis Significant drop‐outs in each group Changed sample size calculation after first 50 participants on basis of observed standard deviation
Risk of bias
Bias	Authors' judgement	Support for judgement
Random sequence generation (selection bias)	Low risk	Computerised restricted randomisation with a block size of 8
Allocation concealment (selection bias)	Low risk	The block size and allocation sequence were unknown to the DEMO trial staff
Blinding (performance bias and detection bias) participants	Unclear risk	Participants not blind, but unclear what influence this had on bias
Blinding (performance bias and detection bias) those delivering intervention	Unclear risk	Physiotherapists delivering the intervention were not blind. Unclear how this influenced risk of bias
Blinding (performance bias and detection bias) outcome assessors	Low risk	The assessor was blind to intervention group. The investigators asked the outcome assessors to guess intervention group. The kappa values for agreement between the right allocation and the guessed allocation were 0.15 and 0.05 for the assessments at 4 and 12 months respectively
Incomplete outcome data (attrition bias) All outcomes	Low risk	137/165 were available for follow‐up at the end of the intervention. Eighteen were lost to follow‐up and 10 refused to participate (8/55 in strength group, 7/55 in aerobic group and 13/55 in the relaxation group). The authors used a likelihood‐based mixed‐effect model with an unstructured variance matrix available in SPSS, which is able to handle missing data with higher precision and power than last observation carried forward. The authors reported no significant difference between missing participants and participants included in the analyses at either 4 or 12 months, and concluded that it was reasonable to assume that the missing data were 'missing at random'
Selective reporting (reporting bias)	Low risk	All prespecified outcomes seem to have been reported. Protocol was published in advance of the trial
Other bias	Unclear risk	The authors repeated power calculations part‐way through the trial, and reduced the sample size as the standard deviation was lower than anticipated