SUMMARYThe design of randomized controlled trials to assess the efficacy of pharmacological measures for the prevention of the gastrointestinal side‐effects of anti‐inflammatory drugs requires an accurate estimate of excess risk under controlled conditions. Photocopies of 952 randomized controlled trial publications were obtained after scanning titles and abstracts of a MEDLINE computer search, 427 were excluded for obvious reasons, and 525 were again photocopied after obliterating source and results. Selection criteria were: the presence of a non‐anti‐inflammatory drug control group; at least 4 days of therapy; at least 3 days without anti‐inflammatory drugs before randomization; no complicating background drugs; mention of side‐effects; and a clear differentiation of gastrointestinal complications. Observer error, with two independent readings, for inclusion suitability in the study was 19% for Methods and 9% for Results. For the 44 aspirin trials, the mean therapy duration was 357 days; the unweighted rate difference between therapy and control groups (± 1 S.E.M.) for ulcer was 0.006 ± 0.003, for gross haemorrhage 0.006 ± 0.002 and for unspecified gastric symptoms 0.03 ± 0.01. In 123 non‐aspirin non‐steroidal anti‐inflammatory drug (NA—NSAID) trials, the mean duration was 67 days; the unweighted rate difference for ulcer was 0.0005 ± 0.0003, for gross haemorrhage 0.007 ± 0.004 and for unspecified gastric symptoms 0.02 ± 0.005. Risk differences were also pooled using the DerSimonian and Laird method, which weights studies inversely according to variance. Using this method, only the unspecified gastric symptoms for non‐aspirin non‐steroidal anti‐inflammatory drugs (NA—NSAIDs) and the haemorrhage for aspirin were found to be statistically significant. Longer studies have higher risk differences. Randomized control trials to determine prophylactic efficacy against haemorrhage (that is, to demonstrate a reduction of ulcer rate in the therapy group to the rate of controls) would require 190 patients in each group for NA—NSAIDs in studies of 2–6 months; 950 subjects would be needed to detect a 50% reduction. Randomized control trials to determine a reduction in ulcer rate to that of controls in patients on aspirin for more than 6 months would require 700 subjects in each group; 3346 subjects would be needed to detect a 50