Intention-to-Treat (ITT) vs Modified ITT (mITT) vs Per Protocol (PP) vs As Treated (AT) vs Complier Average Causal Effect (CACE): Choosing the Right Analysis in Clinical Trials
- Mayta
- Aug 28
- 2 min read
Strategy | Core Logic | Maintains Randomization? | Bias Profile | Best Use Case |
Intention-to-Treat (ITT) | Analyze everyone as randomized, regardless of what happened next. | ✅ Yes | 🔽 Lowest bias (but dilution risk) | Primary analysis for effectiveness (policy relevance, real-world impact). |
Modified ITT (mITT) | ITT with defined exclusions (e.g., no post-randomization data at all). | ⚠️ Partially | Moderate (selection creep) | When minimal follow-up is missing—pragmatic compromise. |
Per Protocol (PP) | Keep only strict adherers, exclude violators. | ❌ No | 🔼 High (selection bias) | Secondary analysis to explore efficacy under perfect use. |
As Treated (AT) | Re-sort patients by what they actually received. | ❌ No | 🔼 High (confounding) | Exploratory; mimics observational analysis. Not valid for primary inference. |
Complier Average Causal Effect (CACE) | Estimates effect among true compliers only. | ✅ Yes (under IV assumptions) | Medium | When non-compliance is substantial but symmetric. More causal than AT/PP. |
🧠 Interpretive Flowchart
When deciding which analysis to apply, follow this sequence:
Primary aim = effectiveness?→ Use ITT as default .
Large dropout with zero data?→ Consider mITT (must predefine rules, symmetric).
Want efficacy under ideal adherence?→ Add PP as sensitivity.
Major crossover or non-adherence?→ Report ITT but add CACE to salvage causal interpretation.
Tempted by AT?→ Remember: this is essentially an observational re-analysis — use only as exploratory.
IV assumptions behind CACE
When you use randomization as an instrumental variable (IV), four key assumptions must hold:
Relevance: Randomization actually influences treatment received (i.e., assignment matters).
Exclusion restriction: Randomization affects the outcome only through the treatment, not directly.
Independence: Randomization is independent of potential outcomes (held by design in an RCT).
Monotonicity: No “defiers”: no one systematically does the opposite of their assignment.
How it’s calculated
It uses an instrumental variable (IV) framework, with randomization as the instrument:
CACE ≈ ITT effect / Proportion of compliers
Example:
If ITT risk difference = 5% improvement,
and 70% of patients were true compliers,
Then CACE ≈ 5% ÷ 0.7 = 7.1% (the estimated effect in compliers).
🔍 Secret Insights
Regulators and CONSORT guidance: ITT must always be presented, even if secondary analyses are added.
Non-Inferiority trials: ITT alone is dangerous (dilution may favor NI); both ITT + PP must agree.
Pragmatic vs Explanatory framing: Pragmatic = ITT; Explanatory = PP. Choice of analysis reflects the clinical stake.
Policy makers vs clinicians: Policy → ITT; Clinicians → may look at PP/AT to understand “what happens if patients really take it”.
✅ Key Takeaways
ITT = safest, lowest bias, preserves randomization.
mITT = compromise, but risk of creeping bias.
PP = efficacy under ideal conditions, but selection bias threat.
AT = exploratory, mimics observational study.
CACE = advanced tool when compliance is a real problem.
Comments