mITT vs CACE in Clinical Trials: Cleaning Convenience or Causal Clarity?

🧪 Modified Intention-to-Treat (mITT): Cleaning for Convenience, at a Cost

🎯 What It Really Estimates

mITT estimates the effect of being assigned and partially engaging with the treatment. It excludes post-randomization non-starters—those who never began treatment or missed key data.

🔍 It sounds tidy, but breaks the causal chain. Why? Because exclusions after randomization violate the principle of exchangeability.

🚨 Why It’s Statistically Dangerous

Breaks randomization: Removing patients post-randomization reshuffles the group characteristics in an uncontrolled way.
Selection bias: Patients who start treatment may differ systematically (healthier, more motivated).
Unclear interpretation: Is it treatment effect or a byproduct of selective inclusion?

🔍 Secret Insight: mITT may look like "clean" data, but it's like cleaning a randomized poker game by discarding hands you don’t like. You’re now playing a different game.

🧬 Complier Average Causal Effect (CACE): Modeling Reality, Respecting Randomization

🎯 What It Really Estimates

CACE aims to estimate what the treatment does to people who would comply no matter what—those who would take treatment if assigned and not if not assigned.

🔬 It isolates the causal effect among “baseline compliers” using instrumental variable (IV) techniques or principal stratification.

🛠 How It Works

Everyone stays in the dataset—no one is removed.
Mathematical models estimate a latent subgroup (the compliers).
CACE = ITT effect ÷ Proportion of compliers.

🧠 This protects the exchangeability granted by randomization while honing in on realistic adherence.

✅ Summary: Core Differences

Aspect	ITT	mITT	CACE
Keeps all randomized?	✅ Yes	❌ No	✅ Yes (but re-focuses)
Preserves randomization?	✅ Yes	❌ No	✅ Yes
Excludes anyone?	No	Yes—by design	No (but focuses analysis on compliers)
Causal interpretation?	Yes (policy level)	No (biased)	Yes (compliance-adjusted)

📊 Head-to-Head: mITT vs CACE

Feature	mITT	CACE
Keeps everyone?	❌ No	✅ Yes
Preserves randomization?	❌ No	✅ Yes
Causal interpretation?	❌ Biased	✅ Yes, for compliers
Method basis	Exclusions, convenience	IV/Principal stratification
Real-world insight	🚫 Limited	✅ High clinical fidelity

🧠 Visual Analogy: Trial as a Gym

Population	Analytic View	Method
Everyone invited	ITT	Full RCT
Only who entered	mITT	Post-hoc cut
Those who’d always comply	CACE	IV modeling

🎯 Sentence Practice

Try completing this:

“I want to estimate the effect of [intervention] in patients who…”

“...are offered treatment” → Use ITT
“...start the treatment” → Use mITT (beware!)
“...would comply if assigned” → Use CACE

🔍 Secret Insight Sidebar

mITT may be accepted by journals, but not defensible as causal inference.
CACE is rigorously causal, but needs assumptions (no defiers, exclusion restriction).
Don’t confuse mITT cleanliness for clarity. It filters out more than data—it filters out meaning.

✅ Key Takeaways

mITT introduces bias—convenient but not causal.
CACE keeps the rigor of ITT while adjusting for real-world compliance.
Use CACE when you care about what works for patients who follow the plan.
Preserve randomization at all costs—don’t toss out the experiment with the dropout.