ITT vs. Per Protocol

Keren Xu


Pragmatic trials are used to test healthcare innovations (e.g. drugs, devices, et al.). They provide evidence for clinical or policy decision making. Intention-to-treat (ITT) analysis keeps patients in that treatment group that was assigned to them, even if they deviated from their assigned group after randomization. One drawback of ITT is that it can be affected by selection bias that are caused by different degrees of adherence or loss to follow up. Thus, ITT analyses can only estimate the effect of treatment assignment in a particular trial, instead of the effect of treatment itself. Someone may argue that the effect of treatment assignment is the effect that truly matters in real life since not all patients would adhere to their prescription. However, the published trial results will change drastically patients perspective on the treatment they are given and thus change their degree adherence. The degree of adherence in a clinical trial will be different from that in real life. Miguel A. Hernán and James M. Robins outlined a alternative analytic approaches per-protocol analysis 1, which is a method to estimate the effect of the treatment if all the patients adhered to the trial protocol. They listed three major points in per-protocol analysis:

  1. data should not be censored just because patients stop the treatment, especially for clinical reasons. For example, in a study investigating the effect of statin therapy on CVD, trial participants who stop statin therapy due to side effects should not be censored in the analysis since they did not deviate from protocol.

  2. data should be censored when it is no longer certain that patients are receiving treatments.

  3. adjustment should be made for confounding (prognostic factors) that causes different degrees of adherence.

One drawback of per-protocol analysis is that it is not only affected by selection bias caused by loss to follow up and adherence, but also affected by confounding due to incomplete adherence to the assigned treatment. Both adherence and loss to follow-up can be affected by social and clinical factors.

G-methods can be used to adjust for postrandomization factors affected by prior treatment. They include inverse-probability weighting, g-estimation of structural nested models, the g-formula, and their extensions that are based on doubly robust estimation.

  1. Hernán MA, Robins JM. Per-protocol analyses of pragmatic trials. N Engl J Med. 2017 Oct 5;377(14):1391-8. Link