STUDY DESIGN: Methodological study nested within a multicentre randomised controlled trial (RCT) of yoga plus usual general practitioner (GP) care vs usual GP care for chronic low back pain. OBJECTIVE: To explore the treatment effects of non-compliance using three approaches in an RCT evaluating yoga for low back pain. SUMMARY OF BACKGROUND DATA: A large multicentre RCT using intention-to-treat (ITT) analysis found that participants with chronic low back pain who were offered a 12-week progressive programme of yoga plus usual GP care had better back function than those offered usual GP care alone. However, ITT analysis can underestimate the effect of treatment in those who comply with treatment. As such, the data were analysed using other approaches to assess the problem of non-compliance. The main outcome measure was the self-reported Roland Morris Disability Questionnaire (RMDQ). METHODS: Complier average causal effect (CACE) analysis, per-protocol analysis and on-treatment analysis were conducted on the data of participants who were fully compliant, predefined as attendance of at least three of the first six sessions and at least three other sessions. The analysis was repeated for participants who had attended at least one yoga session (i.e. any compliance), which included participants who were fully compliant. Each approach was described, including strengths and weaknesses, and the results of the different approaches were compared with those of the ITT analysis. RESULTS: For the participants who were fully compliant (n=93, 60%), a larger beneficial treatment effect was seen using CACE analysis compared with per-protocol, on-treatment and ITT analyses at 3 and 12 months. The difference in mean change in RMDQ score between randomised groups was -3.30 [95% confidence interval (CI) -4.90 to -1.70, P<0.001] at 3 months and -2.23 (95% CI -3.93 to -0.53, P=0.010) at 12 months for CACE analysis, -3.12 (95% CI -4.26 to -1.98, P<0.001) at 3 months and -2.11 (95% CI -3.33 to -0.89, P=0.001) at 12 months for per-protocol analysis, and -2.91 (95% CI -4.06 to -1.76, P<0.001) at 3 months and -2.10 (95% CI -3.31 to -0.89, P=0.001) at 12 months for on-treatment analysis. For the participants who demonstrated any compliance (n=133, 85%), the results were generally consistent with the fully compliant group at 3 months, but the treatment effect was smaller. The difference in mean change in RMDQ score between randomised groups was -2.45 (95% CI -3.67 to -1.24) for CACE analysis, -2.30 (95% CI -3.43 to 1.17) for per-protocol analysis and -2.15 (95% CI -3.25 to -1.06) for on-treatment analysis, which was slightly less than that for ITT analysis. In contrast, at 12 months, per-protocol and on-treatment analyses showed a larger treatment effect compared with CACE and ITT analyses: per protocol analysis -1.86 (95% CI -3.02 to -0.71), on-treatment analysis -1.99 (95% CI -3.13 to -0.86) and CACE analysis -1.67 (95% CI -2.95 to -0.40). CONCLUSION: ITT analysis estimated a slightly smaller treatment effect in participants who complied with treatment. When examining compliance, CACE analysis is more rigorous than per-protocol and on-treatment analyses. Using CACE analysis, the treatment effect was larger in participants who complied with treatment compared with participants who were allocated to treatment, and the difference between ITT and CACE analyses for the fully compliant group at 3 months was small but clinically important. Per-protocol and on-treatment analyses may produce unreliable estimates when the effect of treatment is small. INTERNATIONAL STANDARD RANDOMISED TRIAL NUMBER REGISTER: ISRCTN 81079604.