Eduovisual
Biostatistics & Population Health
Intention-to-treat vs per-protocol analysis
— A randomized controlled trial (RCT) vignette mentions dropouts, crossover, nonadherence, or "lost to follow-up".
— The stem contrasts two effect sizes ("ITT showed RR 0.92, per-protocol showed RR 0.74") and asks which to believe.
— A surgical or device trial highlights patients who never received the assigned procedure.
— A noninferiority or equivalence trial is described (PP becomes critically important here).
— A question about pragmatic vs explanatory trials, FDA approval standards, or regulatory reporting.
Board pearl: ITT preserves randomization and reflects real-world effectiveness; per-protocol estimates biological efficacy under ideal adherence but reintroduces selection bias by conditioning on a post-randomization variable (adherence).
Key distinction: ITT answers "Does the strategy of offering this treatment work?" — PP answers "Does the treatment itself work when actually taken correctly?" Both questions are legitimate; the trial design dictates which is primary.
— "1,000 patients were randomized to drug X or placebo."
— "200 patients in the drug X group stopped therapy due to side effects."
— "150 patients in the placebo group crossed over to drug X."
— "Investigators analyzed only those who completed 12 months of therapy."
— Results: "Among completers, drug X reduced mortality by 30%."
— "Analyzed according to original group assignment"
— "Regardless of treatment received"
— "Including all randomized patients"
— "Among those who completed the protocol"
— "Excluding nonadherent participants"
— "Analyzed by treatment actually received" (this is technically as-treated, but stems blur the line)
— "Restricted to patients who took ≥80% of doses"
— High dropout rate (>20%) → ITT may bias toward the null; PP overestimates effect.
— Asymmetric dropouts between arms → both analyses are vulnerable; consider why.
— Crossover (common in oncology and surgical trials) → PP/as-treated will inflate apparent benefit.
— Open-label design → adherence bias more likely; ITT protects partially.
Board pearl: The CONSORT diagram at the start of every RCT publication exists specifically to show readers the flow from randomization → allocation → follow-up → analysis, so you can see how many were dropped before the analytic dataset was finalized.
Step 3 management: When reading a trial that will inform your patient's care, first locate the ITT result for the primary outcome; treat the per-protocol number as supportive, not definitive. If the two differ dramatically, the trial has an adherence or crossover problem and external validity should be questioned before adopting the therapy in your clinic.
— Were allocation sequence and concealment adequate?
— Baseline characteristics balanced? If imbalanced after randomization, suspect chance or small sample.
— How many assigned to intervention actually received it?
— How many in control received the intervention anyway (contamination)?
— Lost to follow-up <5% = minimal concern; 5–20% = sensitivity analyses required; >20% = serious threat to validity.
— Differential loss between arms is more concerning than equal loss.
— ITT population: all randomized.
— Modified ITT (mITT): all randomized who received ≥1 dose or had ≥1 post-baseline measurement — a common compromise but breaks pure ITT and can introduce bias.
— Per-protocol population: subset meeting adherence/protocol criteria.
— Safety population: usually as-treated, for adverse event reporting.
— Last observation carried forward (LOCF) — outdated, biased.
— Multiple imputation — preferred modern approach.
— Complete case analysis — equivalent to PP for missingness; assumes missing completely at random.
Key distinction: Modified ITT ≠ ITT. Trials labeled "ITT" that exclude even one randomized patient have technically performed mITT. Regulators (FDA, EMA) increasingly scrutinize this distinction because excluding early dropouts can selectively remove the sickest patients.
Board pearl: Adverse events are almost universally reported in the as-treated/safety population, not ITT — because attributing a side effect to a drug a patient never swallowed makes no biological sense. Efficacy = ITT; safety = as-treated.
— If the number analyzed equals the number randomized → ITT.
— If the number analyzed is smaller (excluding dropouts/nonadherent) → per-protocol or mITT.
— "Primary analysis was performed on the intention-to-treat population, defined as all randomized participants" → classic ITT.
— "Per-protocol analysis included participants who received ≥80% of assigned doses and completed the final visit" → PP.
— "As-treated analysis grouped participants by the intervention actually received" → as-treated.
— Superiority trial → ITT is primary (conservative, preserves randomization).
— Noninferiority or equivalence trial → both ITT and PP must agree; PP is often considered equally or more important because ITT biases toward the null (no difference), which falsely favors noninferiority.
— If ITT result is closer to null than PP → typical pattern; suggests nonadherence diluted the effect.
— If PP result is closer to null than ITT → unusual; suggests the protocol-completers were systematically different (selection bias).
— FDA generally requires ITT as primary for drug approval in superiority trials.
— FDA requires concordance between ITT and PP for noninferiority claims.
Board pearl: A trial reporting only per-protocol results for a superiority claim should raise immediate suspicion of selective reporting bias — a major threat in industry-sponsored studies.
Step 3 management: When a guideline cites a trial to support a therapy you're considering prescribing, look up whether the headline number is ITT or PP. ITT results translate more honestly to your real, imperfectly adherent clinic patients.
— ITT primary → PP as sensitivity (and vice versa in noninferiority trials).
— Multiple imputation of missing data vs complete case.
— Tipping-point analyses: how extreme would missing-data assumptions need to be to overturn the result?
— An estimand precisely defines what is being estimated, including how "intercurrent events" (treatment discontinuation, rescue therapy, death) are handled.
— Five strategies for intercurrent events: treatment policy (ITT-like), hypothetical, composite, while-on-treatment, principal stratum.
— Replaces the simplistic ITT-vs-PP dichotomy with a more nuanced specification.
— Instrumental variable analysis using randomization as the instrument.
— Inverse probability weighting for adherence.
— Complier average causal effect (CACE) estimation — estimates effect among those who would comply with either assignment.
— Pragmatic trials test real-world effectiveness → ITT is naturally aligned.
— Explanatory trials test biological efficacy under ideal conditions → PP may be more informative.
Key distinction: ITT measures effectiveness (does offering this work in practice?); PP measures efficacy (does this work when actually used as designed?). Both have legitimate roles; the question is which matches your clinical decision.
Board pearl: When dropouts are informative (e.g., patients drop out because they're feeling worse), neither ITT nor PP gives an unbiased estimate — but ITT is generally less biased because it preserves the comparability achieved by randomization. PP conditions on a post-randomization variable, which fundamentally breaks randomization's protection.
— Primary: ITT — conservative, regulator-preferred.
— Secondary/sensitivity: PP — confirms effect isn't entirely driven by nonadherence dilution.
— Rationale: if ITT shows benefit despite nonadherence, the strategy works in the real world.
— Both ITT and PP must support noninferiority.
— ITT biases toward the null (no difference) — which falsely favors noninferiority by making arms look similar.
— PP is therefore essential to rule out spurious noninferiority.
— ITT is the natural and almost exclusive choice.
— Captures real-world adherence patterns as part of the "treatment."
— PP may be primary because the question is biological, not behavioral.
— Always as-treated — attribute harms to actual exposure.
— Especially tricky: patients randomized to surgery may not undergo it; randomized to control may cross over.
— ITT remains the standard but as-treated analyses are typically reported alongside, with explicit caveats.
Step 3 management: When counseling a patient about a therapy, the ITT relative risk reduction is the most honest number to share, because it reflects the strategy of starting the therapy — including the realistic probability your patient will stop it for side effects. Per-protocol numbers describe an idealized patient who took every dose perfectly.
Board pearl: The phrase "once randomized, always analyzed" is the ITT mantra. Memorize it — it appears on virtually every biostatistics question that touches this concept.
— Antihypertensive trial: 25% of drug-arm patients stop due to cough; analyzed in ITT, the blood pressure reduction looks modest.
— PP analysis (only those who tolerated the drug) shows dramatic BP reduction.
— Correct interpretation: The drug works biologically, but in practice 1 in 4 patients won't tolerate it — both numbers matter for prescribing decisions.
— Patients on placebo arm are allowed to cross over to the active drug at progression.
— ITT overall survival comparison is diluted because placebo arm now contains many treated patients.
— FDA still typically requires ITT as primary; rank-preserving structural failure time (RPSFT) adjustments may be presented as secondary.
— "Among patients taking ≥80% of statin doses, LDL fell by 50%; ITT analysis showed 35% reduction."
— The PP number reflects pharmacology; the ITT number reflects what happens when you write the prescription.
— Some trials use a placebo run-in to exclude nonadherent patients before randomization.
— This is not the same as PP analysis — it occurs pre-randomization, so randomization is preserved, but external validity suffers (your real patients weren't pre-screened for adherence).
Key distinction: A pre-randomization run-in preserves internal validity of ITT but limits generalizability. A post-randomization adherence filter (PP analysis) breaks randomization and threatens internal validity. Both reduce the population the result applies to, but in different ways.
Board pearl: When ITT shows benefit, you can confidently tell your patient "starting this drug reduces your risk." When only PP shows benefit, you must add "if you can take it consistently" — a meaningful caveat for chronic disease management.
— Patients randomized to CABG who refuse surgery and get medical therapy → still analyzed in the CABG arm under ITT.
— Patients randomized to medical therapy who later get CABG for worsening symptoms → still analyzed in medical arm under ITT.
— As-treated analysis would reassign them; this breaks randomization because the decision to cross over is prognostically informative (sicker patients cross over).
— ITT is critical because patients who learn their allocation may seek the active procedure elsewhere.
— Blinding of patients, operators, and outcome assessors is essential.
— Early procedures in a trial may have worse outcomes than later ones.
— PP analysis excluding early "learning" cases is sometimes proposed but is post hoc and biased.
— "Patients randomized to surgery had 5% perioperative mortality; therefore surgery is dangerous."
— But under ITT, that 5% includes patients who never had surgery (so it's an underestimate of operative risk) — and the comparison group includes patients who eventually had surgery anyway.
— Operative mortality specifically requires an as-treated lens; strategy comparison requires ITT.
CCS pearl: When ordering a procedure for a CCS patient, you're committing to a strategy, not guaranteeing the patient will tolerate, complete, or benefit from it. The trial evidence you rely on (ITT) already accounts for some of that messiness — which is why ITT-based guidelines apply better to your unselected clinic population.
Board pearl: In any trial where >10% of patients didn't receive their assigned procedure, demand both ITT and as-treated results before drawing causal conclusions.
— Should be analyzed within the ITT framework.
— Tested for interaction (does the treatment effect differ across the subgroup?), not just within-subgroup significance.
— Multiple testing correction is essential.
— High risk of false-positive findings.
— Should be hypothesis-generating only, never practice-changing.
— Often under-enrolled; trial ITT results may not generalize well.
— Higher dropout rates due to comorbidity, mortality from competing causes.
— Competing risks can distort ITT mortality endpoints.
— Often excluded from trials → ITT effect estimates don't apply to these patients.
— When included, higher dropout for adverse events shifts ITT toward null.
— Frail patients have lower adherence → in PP analyses, they're systematically excluded → PP results overestimate efficacy in the very populations where prescribers most need guidance.
Key distinction: External validity (does this trial apply to my patient?) is separate from the internal ITT vs PP debate. A trial can have impeccable ITT analysis but still not apply to your 85-year-old with CKD stage 4 because such patients were excluded at enrollment.
Step 3 management: When applying RCT evidence to elderly or comorbid patients, mentally adjust expected benefit downward (lower adherence, competing mortality) and expected harm upward (more side effects, drug interactions). The ITT relative risk reduction is your starting point, but absolute risk in your specific patient determines NNT.
Board pearl: A statistically significant ITT subgroup interaction is suspicious if not pre-specified — the Step 3 answer is usually "hypothesis-generating, requires confirmatory trial."
— Rare due to ethical/regulatory barriers; most evidence is observational.
— When RCTs exist (e.g., antihypertensives in pregnancy, antenatal steroids), ITT remains the standard.
— Loss to follow-up is particularly problematic if pregnancy outcomes (miscarriage, stillbirth) are themselves the endpoint — dropouts may be informative.
— Adherence often depends on caregiver behavior, complicating PP definitions.
— Long-term follow-up trials (e.g., asthma controller medications) face high attrition.
— FDA Pediatric Research Equity Act requires pediatric data; ITT analysis is standard.
— ITT applies at the cluster level, not the individual level.
— Individuals who move between clusters are analyzed by original cluster.
— Common in implementation science and public health interventions.
— Designed to mirror real-world practice.
— Broad eligibility, minimal protocol enforcement, routine-care comparators.
— ITT is essentially the only sensible primary analysis.
— Example: PRECIS-style trials of care delivery models, telehealth, decision aids.
— Allow modifications (sample size, randomization ratio, arm dropping) based on interim data.
— ITT principles still apply but require careful pre-specification of the analytic population.
Board pearl: In vaccine efficacy trials, the "ITT" population is sometimes defined as all randomized; the "per-protocol" population includes only those who received all doses and had no prior infection. Headline efficacy numbers are often PP-like, but real-world effectiveness corresponds better to ITT.
Key distinction: Vaccine efficacy (PP, controlled trial) is typically higher than effectiveness (ITT-like, observational, real-world). When counseling patients, the effectiveness number is more honest.
— Excluding nonadherent patients selectively removes those who tolerated the drug poorly or had early adverse events.
— Remaining "completers" are systematically healthier and more adherent — survivorship bias.
— Guideline writers adopting PP-driven results overstate real-world benefit.
— Dilution toward the null.
— Particularly problematic in oncology where crossover at progression is ethically standard.
— May cause regulators to reject effective therapies if not handled with sensitivity analyses.
— ITT in a noninferiority trial biases toward "no difference" — which is the very thing the trial is trying to demonstrate.
— Without concordant PP results, a noninferiority claim is suspect.
— Analyzing safety by ITT (rather than as-treated) attributes side effects to patients who never received the drug — artificially lowering the apparent toxicity rate.
— Trials that quietly switch from a pre-specified ITT primary to a post hoc PP analysis because ITT was non-significant — a form of p-hacking.
— COMPare project and similar initiatives have documented this widely.
— Trials with non-significant ITT but significant PP may be published emphasizing PP results.
— Readers must locate the pre-specified primary analysis in the protocol or registration (ClinicalTrials.gov).
Board pearl: When ITT and PP diverge substantially, the trial is telling you something important about adherence, tolerability, or generalizability — don't ignore the discrepancy.
Step 3 management: Before adopting a new therapy into your practice based on a single trial, verify that the pre-registered primary analysis was ITT and that the published headline matches it. Discordance is a red flag for biased reporting.
— >20% lost to follow-up → primary analysis is fragile regardless of ITT vs PP.
— Differential attrition (>5% difference between arms) → especially concerning.
— >10% crossover → ITT effect substantially diluted.
— Demand sensitivity analyses (IV, RPSFT, CACE).
— Effect sizes differ by >50% → investigate why.
— Often reveals adherence-related effect modification.
— Inadequate; PP must also support noninferiority.
— Excluding randomized patients (even for "never receiving treatment") breaks pure ITT.
— Acceptable if pre-specified and small, suspicious if post hoc.
— LOCF is outdated; multiple imputation or mixed models are modern standards.
— Complete case analysis = silent PP.
— Not automatically invalidating, but warrants extra scrutiny.
— Local journal club or EBM consultation for high-stakes practice-changing trials.
— Pharmacy & Therapeutics committee review before formulary adoption.
— Cochrane review for synthesized, methodologically rigorous summaries.
CCS pearl: Translating trial evidence to bedside orders requires recognizing that CCS scenarios reflect real-world (ITT-like) populations — not the idealized completers of PP analyses. Order therapies whose ITT evidence supports use, and anticipate that your patient may be among the nonadherent.
Board pearl: The single most important question to ask of any RCT: "What was the pre-specified primary analysis, and was it ITT?" If yes and significant, the evidence is robust. If no, dig deeper.
— Excludes randomized patients who never received treatment or had no post-baseline measurement.
— Compromises ITT purity but is widely used.
— Acceptable when pre-specified and exclusions are small/symmetric.
— Groups patients by actual treatment received.
— Used routinely for safety endpoints.
— Breaks randomization for efficacy — biased.
— Restricts to protocol-adherent patients.
— Sensitivity analysis in superiority trials; co-primary in noninferiority.
— Estimates effect among "compliers" (those who would adhere regardless of assignment).
— Uses instrumental variable methods.
— Preserves causal interpretation under specific assumptions.
— Includes only follow-up while patient is on assigned therapy.
— Censors at discontinuation.
— Common in cardiovascular safety trials; introduces informative censoring.
— Assigns worst (or best) outcome to all missing patients.
— Tests robustness; if conclusion holds under worst-case, very robust.
— Identifies the threshold of missing-data assumptions at which the conclusion flips.
Key distinction: ITT preserves randomization; as-treated discards it; per-protocol conditions on a post-randomization adherence variable, partially breaking randomization. The hierarchy of internal validity for causal inference: ITT > mITT > PP > as-treated.
Board pearl: The safety population in nearly every drug trial is as-treated, not ITT — because attributing a drug side effect to someone who never took the drug is biologically nonsensical. Efficacy and safety analyses legitimately use different populations.
— Random sampling affects external validity (generalizability).
— Randomization affects internal validity (causal inference).
— ITT/PP debate is about preserving randomization, not sampling.
— Independent of ITT/PP.
— Open-label trials can still use ITT; double-blind trials can violate ITT via differential dropout.
— Selection bias occurs at enrollment.
— PP analysis introduces a post-randomization selection bias by conditioning on adherence.
— Randomization controls baseline confounding.
— PP analysis can reintroduce confounding because adherence correlates with prognosis.
— Screening trial concepts; distinct from ITT but compounded if screening trial uses PP analysis.
— Adherent patients have better outcomes even on placebo — a well-documented phenomenon (e.g., Coronary Drug Project, 1980).
— This is precisely why PP analyses are biased: adherent patients differ systematically from nonadherent ones.
— Calculated from absolute risk reduction.
— ITT-derived NNT reflects real-world strategy.
— PP-derived NNT reflects idealized scenario — usually smaller (more optimistic).
Board pearl: The Coronary Drug Project (1980) showed that patients adherent to placebo had ~50% lower mortality than nonadherent placebo patients — proving that adherence itself is a marker of overall health behavior, independent of drug effect. This is the empirical foundation for distrusting PP analyses.
Key distinction: Don't confuse ITT (analytic principle) with CONSORT (reporting standard) or GCP (conduct standard). All three improve trial trustworthiness but operate at different stages of the research pipeline.
— Look at the abstract: which population was analyzed?
— Confirm against the registered protocol on ClinicalTrials.gov.
— Use absolute risk reduction from the ITT analysis.
— NNT = 1/ARR.
— This is the most honest number for shared decision-making.
— If PP shows benefit and ITT does not, the drug may work biologically but adherence is a clinical barrier.
— Identifying and addressing adherence barriers in your patient may partially close the gap between ITT effectiveness and PP efficacy.
— Adjust for baseline risk (higher baseline risk → larger absolute benefit, smaller NNT).
— Adjust for likely adherence (poor adherence → expect closer to ITT result; excellent adherence → may approach PP result).
— Chronic disease therapies (statins, antihypertensives, DOACs) show ITT-PP gaps largely driven by discontinuation.
— Secondary prevention success requires adherence interventions: pill counts, refill monitoring, simplified regimens, combination pills, patient education.
— Present ITT numbers as the expected outcome of starting therapy.
— Discuss tolerability and adherence honestly.
— Use decision aids that incorporate real-world (ITT-like) effect sizes.
Step 3 management: When initiating long-term therapy for secondary prevention (post-MI beta-blocker, post-stroke antiplatelet, statin after ACS), counsel patients using ITT-derived NNTs — they represent what realistically happens to patients started on therapy, accounting for the ~30–50% who will discontinue within a year.
Board pearl: Polypill strategies (combination therapy in one pill) consistently improve adherence and narrow the ITT-PP gap — a real-world application of analytic insight to clinical care.
— Self-report (overestimates adherence by ~20%).
— Pill counts (cumbersome, gameable).
— Pharmacy refill records / MPR (medication possession ratio) — most practical; ≥80% typically defines adherence.
— Electronic monitoring (MEMS caps) — research-grade.
— Drug-level testing (e.g., antihypertensives in resistant HTN, immunosuppressants).
— 2–4 weeks after starting a new chronic medication: assess tolerability, side effects, early adherence.
— 3 months: check biomarker response (LDL for statins, A1c for diabetes meds, BP for antihypertensives).
— 6–12 months: long-term adherence reinforcement, dose adjustment.
— Simplification: once-daily > BID > TID > QID dosing.
— Combination pills: amlodipine/valsartan, polypills for cardiovascular prevention.
— Pharmacist-led medication therapy management.
— Automated refill reminders, text messaging.
— 90-day fills, mail-order pharmacy for chronic stable medications.
— Explain expected benefit using ITT-derived numbers ("If 100 patients like you start this drug, about X will avoid a stroke over 5 years").
— Acknowledge that benefit depends on continued use.
— Set expectations for side effects and a plan for managing them.
— Schedule explicit follow-up to reassess.
Board pearl: The single most cost-effective adherence intervention is once-daily dosing; the second is fixed-dose combinations. Both narrow the ITT-PP gap by improving real-world adherence to near-trial PP levels.
Step 3 management: Document adherence at every chronic-disease visit and address barriers proactively — non-adherence is the most common reason for "treatment failure" in primary care, far more common than true pharmacologic resistance.
— Pre-specify the primary analysis in the protocol and registration before unblinding.
— Switching from ITT to PP after seeing data is scientific misconduct if undisclosed.
— Report both per regulatory and journal guidelines (CONSORT, ICMJE).
— Participants must understand that their data will be analyzed by original group assignment, even if they stop the study drug or cross over.
— This is part of the bargain of randomization and must be explicit in consent.
— Participants can withdraw consent at any time.
— Withdrawal of consent for further data collection ≠ exclusion from ITT analysis of data already collected.
— Investigators should clarify in consent whether previously collected data may be retained.
— Crossover provisions in oncology trials reflect ethical equipoise — once progression occurs, equipoise may be lost.
— Allowing crossover is ethically necessary but analytically inconvenient for ITT.
— FDAAA 2007 and ICMJE require prospective trial registration.
— Results must be reported regardless of outcome — combats publication bias.
— Discrepancy between registered primary analysis and published primary analysis is a research integrity red flag.
— A patient discharged on a new evidence-based medication (e.g., post-MI beta-blocker) has a high risk of discontinuation within 30 days.
— Medication reconciliation, clear discharge instructions, early follow-up (within 7–14 days), and primary care handoff are essential.
— ITT trial data assume continued therapy; clinicians must operationalize that assumption.
Step 3 management: At every hospital discharge, perform explicit medication reconciliation, send a discharge summary to the PCP within 48 hours, and schedule follow-up within 7–14 days for high-risk patients (CHF, post-MI, post-stroke) — bridging the ITT trial-to-real-world gap.
Board pearl: Selective outcome reporting and post hoc analysis switching are forms of research misconduct under federal definitions and may invalidate publications.
Board pearl: If a Step 3 question describes nonadherence or crossover and asks which analysis is correct for the primary efficacy comparison of a superiority trial, the answer is almost always ITT.
Key distinction: ITT = effectiveness (real world); PP = efficacy (ideal use). Both are legitimate; the question dictates which matters.
— Stem: "1000 patients randomized; 100 dropped out; investigators analyzed all 1000 by original assignment. Which analysis?"
— Answer: Intention-to-treat.
— Stem: "Investigators excluded 150 nonadherent patients before analysis."
— Answer: Per-protocol — and the question typically asks why this is biased.
— Bias: selection bias / breaks randomization / overestimates effect.
— Stem: "A noninferiority trial showed ITT met the noninferiority margin but PP did not."
— Answer: Noninferiority is NOT established — both must agree.
— Stem: "ITT showed RR 0.95 (NS); PP showed RR 0.70 (significant). Which should guide practice?"
— Answer: ITT is the primary analysis; PP discordance suggests nonadherence diluted effect, but practice should not be based on PP alone for a superiority claim.
— Stem: "Which analytic population is most appropriate for adverse event reporting?"
— Answer: As-treated / safety population.
— Stem: "50% of placebo patients crossed over to drug at progression. ITT showed no survival benefit; as-treated showed large benefit. What's the best interpretation?"
— Answer: ITT is the unbiased primary analysis; as-treated overestimates due to confounding by crossover. Sensitivity analyses (RPSFT) may help but don't replace ITT.
— Stem: "Investigators switched from ITT to PP after seeing the ITT result was non-significant."
— Answer: This is post hoc analytic switching, a form of bias / research misconduct.
— Stem describes Coronary Drug Project findings.
— Answer: Adherence itself is prognostic; PP analyses are biased by this.
Board pearl: When in doubt on a Step 3 biostats question about analysis of an RCT, default to ITT — it is the correct answer the overwhelming majority of the time.
Intention-to-treat analyzes every randomized patient by their original group assignment to preserve randomization and estimate real-world effectiveness, while per-protocol analyzes only adherent completers to estimate biological efficacy — and ITT is the default primary analysis for superiority RCTs, with PP serving as a sensitivity analysis (and co-primary in noninferiority trials).
— "Once randomized, always analyzed" — the ITT mantra. ITT preserves randomization's protection against confounding; PP breaks it by conditioning on adherence (a post-randomization variable correlated with prognosis, per the Coronary Drug Project's healthy adherer effect).
— Superiority trial → ITT primary, PP sensitivity. Noninferiority trial → both ITT and PP must support noninferiority (ITT biases toward the null, which falsely favors noninferiority, so PP concordance is required). Safety analyses → always as-treated, because attributing side effects to patients who never took the drug is biologically meaningless.
— ITT estimates effectiveness (what happens when you offer this strategy to real patients, including nonadherence and dropouts); PP estimates efficacy (what happens when patients actually take it as designed). For clinical counseling and NNT calculations, ITT is the honest number because your real patients won't be perfect adherers — narrow the gap with once-daily dosing, polypills, and structured follow-up at 2–4 weeks, 3 months, and 6–12 months.
— Red flags on critical appraisal: dropout >20%, crossover >10%, modified ITT without justification, discordant ITT and PP results, post hoc analytic switching, PP-only reporting in a superiority trial, or industry-sponsored PP-emphasized headlines. Always verify the pre-specified primary analysis against the ClinicalTrials.gov registration.
Board pearl: When a Step 3 RCT question describes dropouts, crossover, or nonadherence and asks which analysis is correct for the primary efficacy comparison, the answer is intention-to-treat — unless explicitly told it's a noninferiority trial, in which case both ITT and PP must agree.