To determine whether there is a difference in the relative efficacy of individual non-steroidal anti-inflammatory drugs (NSAIDs) when used in the management of osteoarthritis (OA) of the knee. We searched Medline (1966-1995) and Bids Embase (Jan-Dec, 1980-1995). The searches were limited to publications in the English language, and were last perfomed in November 1996. We used modified Cochrane Collaboration search strategy to identify all randomised controlled trials. The MeSH heading osteoarthritis was combined with the generic names of the 17 non-aspirin NSAIDs licensed in the UK for the management of OA in general practice. The search of Embase used the term "osteoarthritis" if present in the abstract, title or keywords, and was combined with the generic names of the 17 non-aspirin NSAIDs, only if they were mentioned in the title, abstract or keywords. All double blind, randomised controlled trials, in the English language, comparing the efficacy of two non-aspirin NSAIDs in the management of osteoarthritis of the knee, were selected. Only trials with subjects aged 16 years and over, with clinical and/or radiological confirmation of the diagnosis of OA knee were included. Studies which compared one "trial" NSAID with one "reference" NSAID were included provided they were non-aspirin NSAIDs available in the UK and were licensed for the treatment of OA by general practitioners. Trials which were placebo-controlled and which also involved the comparison of two NSAIDs were also included. The methodological design of each study was scored according to a pre-determined system. The three main outcome measures of pain, physical function and patient global assessment were chosen based on the core set agreed upon by OMERACT (Outcome Measures in Rheumatology Clinical Trials). These were used to determine the power of each trial. The equivalency of NSAID doses was calculated using the percentage of the recommended maximum daily dose. Sample size estimates for the detection of clinically relevant changes in outcome measures used in the assessment of OA knee were used for power calculations. These calculations were performed to determine whether the trials were of a sufficient size to detect clinically relevant differences which were statistically significant. The calculations incorporate estimates of standard deviation, and minimum, median and maximum differences (delta) between drugs which are deemed to be clinically important. The number of "withdrawals due to lack of efficacy" was also selected as an outcome measure for this review. The Peto odds ratio and 95% confidence intervals were calculated where possible. The results of studies which compared the same trial and reference NSAIDs were combined where possible. Of the 1151 trials identified by the search strategy, 22 involved knee osteoarthritis only. Sixteen of these trials fulfilled the inclusion criteria and were entered in the review. Eight NSAIDs were represented in these trials. Etodolac was represented in 11 trials. The reference NSAID in these trials was piroxicam (n=3), naproxen(n=3), diclofenac (n=3), indomethacin (n=1), and, nabumetone (n=1). The reported methodological design of the trials was poor, with a median score of 3 (out of a maximum of 8). The results of the trials comparing the same trial and reference NSAIDs were pooled for the outcome "withdrawal due to lack of efficacy". For the comparison, etodolac vesus piroxicam, the odds ratio favoured etodolac i.e. patients receiving etodolac were less likely to withdraw due to lack of efficacy. The dose of etodolac used in each of these three studies, however, was greater than the corresponding dose of piroxicam (based on percentage maximum daily dose). The significance of these results is therefore questionable. For the comparisons etodolac versus diclofenac, and etodolac versus naproxen, there were no clear differences betw