Intention-to-Treat & Sensitivity Analysis
A Step 1 biostatistics lesson on intention-to-treat analysis (analyze as randomized, preserves randomization, conservative/toward-null in superiority trials, anti-conservative in non-inferiority) contrasted with per-protocol and as-treated, paired with sensitivity analysis as the tool that stress-tests ITT results against missing-data and other assumptions.
Intention-to-treat (ITT) analysis keeps every randomized participant in their originally assigned group — regardless of adherence, crossover, or dropout ("once randomized, always analyzed"). This preserves the balance of known and unknown confounders created by randomization, protecting against selection bias and confounding; by retaining non-adherers and dropouts it also limits attrition bias, though their missing outcomes still require handling. Because ITT keeps non-adherers and crossovers in their assigned arm, it dilutes the apparent treatment effect and biases toward the null in superiority trials — a conservative, real-world estimate of effectiveness.
Sensitivity analysis is the companion tool: it re-runs the primary analysis under different assumptions (alternative missing-data imputation, per-protocol population, excluding outliers, different models) to test whether the conclusion is robust or fragile. Boards pair these concepts because the biggest threat to an ITT result — missing outcome data from dropouts — is exactly what a sensitivity analysis stress-tests.
- ITT = analyze by the group randomized to — "once randomized, always analyzed"
- Preserves randomization → maintains confounder balance → prevents selection bias / confounding
- Conservative in superiority trials: including non-adherers/dropouts dilutes effect → biases toward the null
- Per-protocol (PP): only adherent completers analyzed → breaks randomization → tends to overestimate efficacy (selection/attrition bias)
- As-treated: analyzed by treatment actually received → also breaks randomization → confounding
- ITT ≈ effectiveness (real world); PP ≈ efficacy (ideal conditions)
- Non-inferiority trials: ITT is NOT conservative — bias toward null makes arms look similar → favors declaring non-inferiority → report both ITT and PP
- Sensitivity analysis = vary assumptions/methods; if the conclusion holds, it is robust
| Feature | Intention-to-Treat | Per-Protocol | As-Treated |
|---|---|---|---|
| Analysis group | As randomized | Only adherent completers | By treatment received |
| Randomization preserved | Yes | No | No |
| Main bias risk | Dilution toward null | Selection / attrition | Confounding |
| Effect estimate | Conservative (effectiveness) | Often overestimates (efficacy) | Variable |
| Preferred as primary | Superiority RCT | Non-inferiority (report both) | Rarely |
Vignette: An RCT randomizes 500 patients to a new anticoagulant vs warfarin. In the drug arm, 40 patients stop the medication due to bleeding; in the warfarin arm, 15 patients switch to the new drug. Investigators want the analysis least prone to bias for the primary superiority comparison of stroke prevention.
Answer / next step: Intention-to-treat — analyze all 500 patients in their originally randomized groups, keeping the 40 non-adherers and 15 crossovers in their assigned arms. This preserves the confounder balance from randomization and gives an unbiased (conservative) estimate. Choosing per-protocol here would selectively drop the sickest/side-effect patients, break randomization, and overestimate the drug's benefit.
"Once randomized, always analyzed." — ITT keeps each patient in their original arm no matter what happens after randomization.
- ITT = Intent, not treatment received.
- Directionality (superiority trials): ITT dilutes, PP inflates.
- Sensitivity analysis: repeat the primary analysis under alternate assumptions to test robustness
- Common variations: ITT vs per-protocol, different missing-data imputation, excluding outliers, alternate models/subgroups
- Missing-data methods: complete-case, last observation carried forward (LOCF), multiple imputation, best-case/worst-case (extreme-case) analysis
- Robust = conclusion unchanged across methods; fragile = conclusion depends on one assumption
- Differential dropout (unequal loss between arms) threatens validity → sensitivity analysis is especially important
- Reported alongside ITT to show that dropouts/missing data did not manufacture the result
Vignette: A 2-year trial of a weight-loss drug loses 20% of participants to follow-up, with more dropouts in the placebo arm. The primary ITT result favors the drug. To confirm the finding is not an artifact of how missing data were handled, the authors repeat the analysis under extreme-case assumptions — counting all dropouts as failures, then as successes (worst- and best-case bounds) — and again with multiple imputation.
Concept / next step: This is a sensitivity analysis. If the drug still shows benefit across all scenarios, the conclusion is robust. If the effect disappears under plausible missing-data assumptions, the result is fragile and may reflect differential dropout rather than a true treatment effect.
Practice Research Methods now
Board-style questions, spaced-repetition flashcards, and a Socratic AI tutor — free to start.