We included 42 studies involving 1453 participants. The trials included participants who had some residual motor power of the paretic arm, the potential for further motor recovery and with limited pain or spasticity, but tended to use the limb little, if at all. The majority of studies were underpowered (median number of included participants was 29) and we cannot rule out small-trial bias. Eleven trials (344 participants) assessed disability immediately after the intervention, indicating a non-significant standard mean difference (SMD) 0.24 (95% confidence interval (CI) -0.05 to 0.52) favouring CIMT compared with conventional treatment. For the most frequently reported outcome, arm motor function (28 studies involving 858 participants), the SMD was 0.34 (95% CI 0.12 to 0.55) showing a significant effect (P value 0.004) in favour of CIMT. Three studies involving 125 participants explored disability after a few months of follow-up and found no significant difference, SMD -0.20 (95% CI -0.57 to 0.16) in favour of conventional treatment.