Intention-to-Treat vs Per-Protocol Analysis in Longevity Trials

Who This Is Useful For

This page is useful for readers comparing primary and sensitivity analyses in randomized studies of diet, exercise, behavioral programs, medications, or other sustained interventions related to ageing and healthspan. These studies can separate assigned exposure from achieved exposure, as illustrated by long-term calorie-restriction trials in which the prescribed and achieved restriction differed. [8] [9]

Intention-to-treat and per-protocol analyses are not merely two statistical routes to the same answer. They are attempts to answer different causal questions. Intention-to-treat usually asks what happened after participants were assigned to an intervention, while a per-protocol effect asks what would have happened under adherence to a specified intervention protocol. [3] [6]

This distinction is important because randomization occurs at assignment, not at later adherence. Participants who continue a demanding protocol can differ from those who discontinue it in health, adverse effects, motivation, or other prognostic characteristics. Removing one group after randomization can therefore turn a randomized comparison into an observational one. [2] [3] [10]

The Two Approaches at a Glance

1. What Intention-to-Treat Preserves

In an intention-to-treat analysis, participants remain in their randomized groups even if they stop, switch, or incompletely follow the assigned intervention. This preserves the trial's original allocation and reduces bias that could arise from excluding participants because of post-randomization events. [1] [2] [4]

The resulting estimate reflects the effect of an assignment strategy as it operated in that trial, including the trial's adherence, co-interventions, and treatment changes. It should not automatically be described as the biological effect that would occur if every participant sustained the planned exposure. [3] [7]

2. What Per-Protocol Tries to Estimate

A per-protocol analysis aims to estimate an effect under adherence to a pre-specified protocol. A simple version retains only participants who met an adherence threshold, but modern causal approaches may adjust for measured variables that predict adherence and outcomes over time. [3] [6]

The phrase “per protocol” is incomplete unless the protocol is defined. It may refer to taking a minimum dose, attending a specified proportion of sessions, avoiding prohibited co-interventions, or following a time-varying rule. Different definitions can target different effects and include different participants. [3] [6]

3. Why Excluding Non-Adherers Can Create Bias

Adherence is measured after randomization and can be influenced by early symptoms, adverse effects, frailty, competing demands, or perceived response. If any of those factors also predict the study outcome, adherers and non-adherers are no longer exchangeable simply because they began in randomized groups. [3] [6] [10]

This is why an unadjusted adherer-only analysis can overstate or understate an intervention effect. Statistical adjustment can reduce bias when the important predictors of adherence and outcome were measured adequately, but causal per-protocol estimates still depend on assumptions that cannot all be verified from the observed data. [3] [6]

4. Missing Outcomes Are a Separate Problem

Keeping participants in their assigned groups does not recover outcomes that were never measured. A report can use the intention-to-treat label while still excluding participants with missing endpoint data, so readers need the participant flow and the exact analysis population rather than the label alone. [2] [4]

Methods for missing data rely on assumptions about why observations are missing and what the unobserved values might have been. No single method repairs substantial missingness under every plausible mechanism; trial retention, transparent assumptions, and sensitivity analyses are therefore central to interpretation. [4] [5]

5. The Estimand Comes Before the Analysis Label

An estimand is a precise description of the treatment effect a study intends to estimate, including the population, endpoint, treatment conditions, summary measure, and handling of events such as treatment discontinuation or switching. The estimand framework is intended to align the trial question, design, conduct, and statistical analysis. [7]

Under this framework, intention-to-treat and per-protocol language is a starting point rather than a complete specification. Two analyses with the same label may handle rescue treatment, death, treatment switching, or missing measurements differently and therefore estimate different quantities. [6] [7]

6. Why Longevity Trials Make the Distinction Important

Longevity-related interventions may require sustained behavior and repeated follow-up while relying on biomarkers or risk factors rather than directly observed lifespan effects. This creates room for the assigned intervention, achieved exposure, and measured endpoint to diverge over time. [8] [9]

CALERIE illustrates the distinction. Participants were randomized to a two-year intervention designed to achieve 25% calorie restriction or to an ad libitum control, while the intervention group achieved a smaller average restriction than prescribed. The intention-to-treat results therefore describe the effect of assignment to that demanding program under achieved trial conditions, not the effect of a uniform 25% reduction sustained by every participant. [8] [9]

How to Read the Analyses

What This Does Not Mean

• It does not mean intention-to-treat always gives a smaller effect; the direction depends on adherence patterns, treatment switching, outcome missingness, and intervention effects. [3] [4]
• It does not mean per-protocol questions are invalid; they require a clearly defined protocol and methods suited to the post-randomization causes of adherence. [3] [6]
• It does not mean agreement between the two analyses proves absence of bias; similar numerical results can arise under different assumptions and target questions. [3] [7]
• It does not mean one analysis strategy fits every trial design; non-inferiority and equivalence trials have additional interpretation and reporting considerations. [11]

Practical Interpretation Examples

• If the intention-to-treat effect is modest but adherence was low: interpret it as the effect of assignment under the observed adherence pattern, then inspect whether the per-protocol analysis clearly defines adherence and addresses selection bias. [3] [6]
• If the per-protocol estimate is much larger: the difference may reflect greater exposure, selection of healthier adherers, model assumptions, or a combination of these explanations. [3] [10]
• If many outcomes are missing: neither keeping randomized labels nor restricting to protocol completers resolves the missing-data problem without additional assumptions. [4] [5]
• If the paper says “modified intention-to-treat”: check the exact inclusion rule because this term does not reliably identify which randomized participants were excluded. [2]

Related Reading

• Intervention Adherence in Longevity Trials
• Randomized Controlled Trials vs Observational Studies in Longevity Research
• How to Read the Methods and Limitations of a Longevity Study
• Selection Bias and Healthy User Bias in Longevity Studies

Summary

Intention-to-treat analysis estimates the effect of randomized assignment under the conditions observed in a trial. Per-protocol analysis targets an effect under adherence to a defined protocol, but it loses the simple protection of randomization unless post-randomization selection is handled with appropriate assumptions and methods. In longevity trials, careful reading therefore requires the target estimand, adherence definition, participant flow, missing-data strategy, and both effect estimates—not a choice between two labels in isolation. [2] [3] [6] [7]

References

1. Schulz, K. F., Altman, D. G., and Moher, D. (2010). CONSORT 2010 Statement: updated guidelines for reporting parallel group randomised trials. BMJ. https://www.bmj.com/content/340/bmj.c332
2. Moher, D., Hopewell, S., Schulz, K. F., et al. (2010). CONSORT 2010 Explanation and Elaboration: updated guidelines for reporting parallel group randomised trials. BMJ. https://pmc.ncbi.nlm.nih.gov/articles/PMC2844943/
3. Hernan, M. A., and Robins, J. M. (2017). Per-protocol analyses of pragmatic trials. New England Journal of Medicine. https://www.nejm.org/doi/10.1056/NEJMsm1605385
4. White, I. R., Horton, N. J., Carpenter, J., and Pocock, S. J. (2011). Strategy for intention to treat analysis in randomised trials with missing outcome data. BMJ. https://www.bmj.com/content/342/bmj.d40
5. Little, R. J., Cohen, M. L., Dickersin, K., et al. (2012). The design and conduct of clinical trials to limit missing data. Statistics in Medicine. https://pmc.ncbi.nlm.nih.gov/articles/PMC5944851/
6. Rudolph, J. E., Naimi, A. I., Westreich, D., and Kennedy, E. H. (2020). Defining and identifying per-protocol effects in randomized trials. Epidemiology. https://pmc.ncbi.nlm.nih.gov/articles/PMC7400733/
7. Kang, M., Kendall, M. A., Ribaudo, H., et al. (2022). Incorporating estimands into clinical trial statistical analysis plans. Clinical Trials. https://pmc.ncbi.nlm.nih.gov/articles/PMC9232859/
8. Ravussin, E., Redman, L. M., Rochon, J., et al. (2015). A 2-year randomized controlled trial of human caloric restriction: feasibility and effects on predictors of health span and longevity. The Journals of Gerontology: Series A. https://pmc.ncbi.nlm.nih.gov/articles/PMC4841173/
9. Kraus, W. E., Bhapkar, M., Huffman, K. M., et al. (2019). 2 years of calorie restriction and cardiometabolic risk: exploratory outcomes of a multicentre, phase 2, randomised controlled trial. The Lancet Diabetes & Endocrinology. https://pubmed.ncbi.nlm.nih.gov/31303390/
10. Shrank, W. H., Patrick, A. R., and Brookhart, M. A. (2011). Healthy user and related biases in observational studies of preventive interventions: a primer for physicians. Journal of General Internal Medicine. https://pmc.ncbi.nlm.nih.gov/articles/PMC3077477/
11. Piaggio, G., Elbourne, D. R., Pocock, S. J., Evans, S. J. W., and Altman, D. G. (2012). Reporting of noninferiority and equivalence randomized trials: extension of the CONSORT 2010 statement. JAMA. https://pubmed.ncbi.nlm.nih.gov/23268518/