Chapter 4 — Intention‑to‑treat vs per‑protocol

The ‘you signed up for a real‑deal, but you switched to the cheat‑code’ dilemma

A trial randomised 300 septic shock patients to a new vasopressor infusion protocol (drug A) vs standard care (drug B). After a year of data, the primary outcome mortality was 22% (67/304) in the drug A group and 26% (78/302) in the control group. The authors proudly conclude: “Protocol deviation reduced mortality by 11% (ARR 0.09, NNT 11).”

You, however, read the fine print. The drug A protocol was, in fact, a hybrid of both drug A and drug B, used sequentially in the first 30 minutes to achieve MAP ≥ 65. The intention‑to‑treat (ITT) analysis reported an absolute difference that barely reached statistical significance (p = 0.043).

But the authors also performed a per‑protocol (PP) analysis, restricting the drug A arm to those who followed the full protocol exactly, and the control arm to those who never received drug A outside the protocol. In this subset, mortality fell to 18% vs 28%, p = 0.01.

Your question: which result should drive changes at the bedside? Do you treat the patients you enrolled, or the patients who adhered perfectly to the protocol?

The answer depends on what you are trying to estimate, and that is why the statistical community splits between two analysis strategies: ITT vs PP.

What the two strategies answer

You will see both in the same paper because they are mathematically different; each can answer a different clinical question. However, PP analyses are vulnerable to selection bias because the “protocol adherers” often differ systematically from those who deviated (sicker patients might be more likely to deviate, or healthier patients might be more likely to adhere).

Why you should trust ITT more in clinical decision‑making

Real‑world relevance

Practice is not a sterile clinical trial. Patients leave the ICU, transfer to other wards, receive concomitant therapies, and sometimes ignore protocol. The ITT analysis reflects that messiness, giving you an estimate of the effect of offering the treatment to your real patient population.

A study of early goal‑directed therapy for sepsis (the EGDT arm of the ProCESS trial) famously found no difference in mortality when the protocol was rolled out in real ICUs, despite a strong mortality benefit under the controlled Armstrong trial conditions. The ITT estimate is the one that matters for you when you read a protocol and consider implementing it.

Protection against selection bias

Consider a PP analysis of a new sedation protocol. Patients who were calmer, fewer comorbidities, or had better families may have stayed on the protocol; those who were agitated or had multiple complications often were taken off it. If you compare these two groups, you are comparing different populations, not the same population under different treatments. The mortality difference may simply reflect these baseline imbalances.

When PP is still valuable

1. Safety analyses – PP can show whether the treatment, when used correctly, is harmful. If a drug causes severe liver injury only when taken at maximum dose, and PP analysis detects that but ITT dilutes the signal, PP is the safer read.
2. Efficacy trials – In earlier phase II or phase III trials with strict inclusion/exclusion criteria, PP gives a clearer picture of what the drug can do under ideal conditions. Regulatory decisions often use this signal, but real‑world decisions still rely on ITT.
3. Biomarker or surgical studies – For procedures (e.g., “central venous catheter placement by experts”) or imaging biomarkers, PP may be the only feasible estimate, because the exposure is not a drug you can continue or discontinue. The protocol defines the exposure period; adherence determines who is “exposed.”

How to spot ITT/PP in the manuscript

Documentation in the methods

Look for language such as: > “The primary analysis was conducted on an intention‑to‑treat basis, including all randomised patients in the group assigned, regardless of subsequent treatment.”

or > “A per‑protocol analysis was also performed, excluding patients with major protocol violations (≥ 2 ml/kg fluid bolus outside the allowed window).”

Reporting in the results

• Tables usually present baseline tables and primary outcomes for both ITT and PP if both were done.
• Statistical tests for PP often include different denominators (e.g., “PP analysis included 178 patients in the treatment group and 183 controls”).
• Separate p-values: ITT may be non‑significant, PP may be significant — this is a red flag for bias.

What the CONSORT says

CONSORT 2020 emphasises that ITT is the preferred method for estimating treatment effects in RCTs and should be reported for primary outcomes. PP analyses are encouraged only when clearly justified and should be pre‑specified in the protocol.

Your bedside analysis checklist for each RCT

When you encounter an RCT, apply these four checkpoints:

1. Which analysis is the primary? — Most papers list “Primary outcome: mortality at 28 days (ITT analysis)” in the abstract; watch for “Per‑protocol analysis is reported as supportive.” If no clear primary analysis, favour ITT for clinical decisions.
2. Did the authors discuss the differences? — Do they mention the potential bias in PP, e.g., “PP analysis may be subject to selection bias because patients who deviated were often sicker.”
3. Is the effect direction the same? — If ITT shows a benefit and PP shows a harm, that is a red flag; you need to investigate why.
4. Which estimate aligns with your question? — If you are asking “Should I adopt this protocol now?” use ITT. If you are asking “What is the theoretical maximum benefit under ideal conditions?” PP may be of interest but not actionable.

Go deeper

• StatPearls – “Intention‑to‑Treat Principle in Clinical Trials” (NBK534302): concise definition, pros/cons, and example tables. https://www.ncbi.nlm.nih.gov/books/NBK534302/
• BMJ Statistics Notes – “Intention‑to‑treat analysis” (Altman & Bland): short on why ITT is protective and how to interpret PP results. https://www.bmj.com/content/bmj_stats_notes
• RCT methodology article, open‑access (doi:10.1186/s13063-020-04185-3): “Intention‑to‑treat vs per‑protocol analysis in randomised controlled trials: what are the implications?” https://doi.org/10.1186/s13063-020-04185-3 (free full text via PMCID: PMC7027125)

Next: Chapter 5 — Types of data & descriptive statistics