Exercise is an effective treatment for osteoarthritis. However, the effect may vary from one patient (or study) to another.

To evaluate the efficacy of exercise and its potential determinants for pain, function, performance, and quality of life (QoL) in knee and hip osteoarthritis (OA).

We searched 9 electronic databases (AMED, CENTRAL, CINAHL, EMBASE, MEDLINE Ovid, PEDro, PubMed, SPORTDiscus and Google Scholar) for reports of randomised controlled trials (RCTs) comparing exercise-only interventions with usual care. The search was performed from inception up to December 2017 with no language restriction. The effect size (ES), with its 95% confidence interval (CI), was calculated on the basis of between-group standardised mean differences. The primary endpoint was at or nearest to 8 weeks. Other outcome time points were grouped into intervals, from<1 month to≥18 months, for time-dependent effects analysis. Potential determinants were explored by subgroup analyses. Level of significance was set at P≤0.10.

Data from 77 RCTs (6472 participants) confirmed statistically significant exercise benefits for pain (ES 0.56, 95% CI 0.44–0.68), function (0.50, 0.38–0.63), performance (0.46, 0.35–0.57), and QoL (0.21, 0.11–0.31) at or nearest to 8 weeks. Across all outcomes, the effects appeared to peak around 2 months and then gradually decreased and became no better than usual care after 9 months. Better pain relief was reported by trials investigating participants who were younger (mean age<60 years), had knee OA, and were not awaiting joint replacement surgery.

Exercise significantly reduces pain and improves function, performance and QoL in people with knee and hip OA as compared with usual care at 8 weeks. The effects are maximal around 2 months and thereafter slowly diminish, being no better than usual care at 9 to 18 months. Participants with younger age, knee OA and not awaiting joint replacement may benefit more from exercise therapy. These potential determinants, identified by study-level analyses, may have implied ecological bias and need to be confirmed with individual patient data.