Dennis 2005.

Methods	Multicentric parallel randomised controlled trial Setting: multicentric study involving many countries, mainly UK Sample size: 1000 patients based on 85% power to detected and absolute risk difference for death or poor outcome of 9%. Type one error: 0.05
Participants	321 patients: 144 male, 177 female; mean age 76 (10); dysphagic stroke patients Inclusion criteria: recent stroke (within 7 days before admission), first‐ever or recurrent, if the responsible clinician was uncertain of the best feeding (PEG or NGT) Exclusion criteria: patients with subarachnoid haemorrhage
Interventions	PEG (n = 162) NGT (n = 159)
Outcomes	Mortality or poor outcome Overall survival Utility score (EUROQoL) Quality of life (EUROQoL) Length of hospital stay Adverse events in hospital stay Pneumonia Causes of death Treatment effect Number of tubes inserted Reasons for stopping feeding Vital status Functional ability (Modified Rankin scale) Clinicians' satisfaction about enteral feeding Time in enteral nutrition
Notes	Follow‐up: six months Outcomes 3, 10 and 13 were not suitable for analysis
Risk of bias
Bias	Authors' judgement	Support for judgement
Random sequence generation (selection bias)	Low risk	Computer‐generated, stratified by country, age, gender, and predicted probability of poor outcome (by minimisation)
Allocation concealment (selection bias)	Low risk	The randomisation systems were housed on a secure server with access permitted, via a password. Participating centres were issued with codes in order for them to access the randomisation services (three separate numerical codes).
Blinding of participants and personnel (performance bias) All outcomes	High risk	According to the authors, "the randomising clinician, the clinical team, and the patients were not unaware to treatment allocation— doing so would have been impossible".
Blinding of outcome assessment (detection bias) All outcomes	High risk	According to the authors, "the randomising clinician, the clinical team, and the patients were not unaware to treatment allocation— doing so would have been impossible". However, 6 month of follow‐up was carried out for the following variables: patients’ vital status, functionalability with themodified Rankin score (MRS), 19 place of residence, method of feeding, and quality of life with the EUROQoL. For these variables, the authors referred that "follow‐up was masked to treatment allocation (except where patients or carers inadvertently unmasked an interviewer at follow‐up; such occurrences were unusual but their frequency was not systematically recorded)". Because of these divergences the study was considered as of high risk of bias.
Incomplete outcome data (attrition bias) All outcomes	Low risk	Flow of patients was clearly reported
Selective reporting (reporting bias)	Low risk	Relevant outcomes were analysed
Other bias	Low risk	None suspected