Abstract
Background:
Non-adherence to interventions is common in randomized controlled trials (RCTs), complicating the interpretation of treatment effects. The intention-to-treat (ITT) principle estimates the treatment effect of assignment to intervention but does not reflect efficacy among those who adhere. Per-protocol (PP) analyses attempt to address this but introduce selection bias by violating randomisation. The complier average causal effect (CACE) provides an efficacy estimand among compliers while preserving randomisation. This study aimed to provide an empirical comparison of ITT, PP, and CACE approaches using individual participant data (IPD) from trials of depression interventions in primary care.
Methods:
We analysed IPD from the Depression in General Practice (Dep-GP) collaboration, comprising seven eligible RCTs with 3,467 participants. Trials reported continuous (depression symptom scores) or binary (treatment response) outcomes. Adherence was defined within the intervention group. We conducted a two-stage IPD meta-analysis to estimate treatment effects under ITT, PP, and CACE. Results were expressed as differences in standardised mean difference (ΔSMD) for continuous outcomes and as ratios of odds ratios (ROR) for binary outcomes. One-stage mixed-effects models were performed as secondary analyses.
Results:
For binary outcomes, both PP and CACE analyses produced larger effects than ITT (ROR for PP vs ITT: 1.09; 95% CI, 1.05–1.14; P < .001; CACE vs ITT: 1.19; 95% CI, 1.00–1.42; P < .05). For continuous outcomes, CACE yielded a larger effect than ITT (ΔSMD = 0.10; 95% CI, 0.01–0.20; P < .05), while PP did not differ from ITT (ΔSMD = 0.03; 95% CI, –0.01 to 0.08; P = .167). Sensitivity analysis, excluding the TREAD trial, yielded larger effect by the PP method (ΔSMD = 0.05; 95% CI, 0.01–0.09; P < .05).
Discussion:
Our findings demonstrate that CACE provides a causal efficacy estimand that diverges from the treatment policy effect estimated by ITT, while PP yields similar but potentially biased results. This highlights the importance of distinguishing between estimands in the presence of non-adherence and illustrates empirically how they differ in practice.
Conclusions:
Current RCT reporting recommendations should be updated to require routine reporting of CACE alongside ITT, together with adherence information, to provide a more complete and transparent account of treatment effects.
Non-adherence to interventions is common in randomized controlled trials (RCTs), complicating the interpretation of treatment effects. The intention-to-treat (ITT) principle estimates the treatment effect of assignment to intervention but does not reflect efficacy among those who adhere. Per-protocol (PP) analyses attempt to address this but introduce selection bias by violating randomisation. The complier average causal effect (CACE) provides an efficacy estimand among compliers while preserving randomisation. This study aimed to provide an empirical comparison of ITT, PP, and CACE approaches using individual participant data (IPD) from trials of depression interventions in primary care.
Methods:
We analysed IPD from the Depression in General Practice (Dep-GP) collaboration, comprising seven eligible RCTs with 3,467 participants. Trials reported continuous (depression symptom scores) or binary (treatment response) outcomes. Adherence was defined within the intervention group. We conducted a two-stage IPD meta-analysis to estimate treatment effects under ITT, PP, and CACE. Results were expressed as differences in standardised mean difference (ΔSMD) for continuous outcomes and as ratios of odds ratios (ROR) for binary outcomes. One-stage mixed-effects models were performed as secondary analyses.
Results:
For binary outcomes, both PP and CACE analyses produced larger effects than ITT (ROR for PP vs ITT: 1.09; 95% CI, 1.05–1.14; P < .001; CACE vs ITT: 1.19; 95% CI, 1.00–1.42; P < .05). For continuous outcomes, CACE yielded a larger effect than ITT (ΔSMD = 0.10; 95% CI, 0.01–0.20; P < .05), while PP did not differ from ITT (ΔSMD = 0.03; 95% CI, –0.01 to 0.08; P = .167). Sensitivity analysis, excluding the TREAD trial, yielded larger effect by the PP method (ΔSMD = 0.05; 95% CI, 0.01–0.09; P < .05).
Discussion:
Our findings demonstrate that CACE provides a causal efficacy estimand that diverges from the treatment policy effect estimated by ITT, while PP yields similar but potentially biased results. This highlights the importance of distinguishing between estimands in the presence of non-adherence and illustrates empirically how they differ in practice.
Conclusions:
Current RCT reporting recommendations should be updated to require routine reporting of CACE alongside ITT, together with adherence information, to provide a more complete and transparent account of treatment effects.
| Original language | English |
|---|---|
| Journal | BMC Medical Research Methodology |
| Publication status | Accepted/In press - 29 Dec 2025 |