Intention to Treat and Per Protocol Analyses

The Intention to Treat (ITT) and the Per Protocol (PP) Analyses respectively include:

• - all patients enrolled even those who were lost while on trial,
• - only the patients who entirely completed the study.

This issue is important, because a trial with major differences in results between an ITT and PP analysis is not robust. For example, usual null hypothesis testing with the ITT population makes differences in a treatment comparison look smaller, but it mirrors, what will happen in practice (including the non-compliants). However, it, also, shifts the study towards a negative result. In contrast, with, for example, equivalence testing instead of null hypothesis testing, similarly differences will be smaller, and, again, the results will mirror, what will happen in practice. But, the ITT analysis will here shift the study towards a positive result. An adequate recommendation would be, therefore, to perform, with equivalence studies, both an ITT and a PP analysis. If the differences in results between the two are small, then the study was robust.

The above study is an example of a study with simultaneous publication of both ITT and PP analyses (Lancet 2015; 385: 1519-26).

In the above statistical analysis plan (SAP), protocol violators are excluded from the PP analysis, and the study is analyzed according to an ITT procedure, while a full analysis data set would have involved all randomized subjects. According to such study protocol, the preservation of the initial randomization is crucial.

Thus, also included in the ITT analysis were

• • protocol violators,
• • drop-outs,
• • withdrawals (unless informed consent had been withdrawn), and
• • those who were treated differently or crossed-over,

The PP analysis dealt with differential drop-out ..., (and it required at least one outcome measurement). The ITT strategy is, sometimes, called a conservative strategy, and it is close(r) to clinical practice.

A modifiedITT analysis is, currently, often applied in trial protocols. “Modified” means, that only those patients have been included, who received at least ‘one’ medication - dose/treatment. This messes up the randomization principle (but, fortunately, only marginally). Differential withdrawal/drop-out rates may be the cause of two treatment groups differing with respect to important prognostic variables (in the baseline variables).

With ITT analyses, exclusion due to failure of an exclusion/inclusion criterion is possible, but entry criteria must be measured prior to randomization, detection of failure must be completely objectively, all patients must be equally scrutinized, and all detected violators must be excluded.

A per protocol set (PP) should include ‘valid cases’, otherwise called the ‘efficacy sample’, or the ‘evaluable cases’, consisting of patients compliant with the protocol (including pre-specified minimal exposure criteria). It, also, should have available measurements of the primary outcome variable(s), and no major protocol violations (including violation of inclusion/exclusion criteria), and, finally, it should not use excluded medication, only include data with adequate compliance, and it should not include patients lost for follow-up, or, otherwise, missing data. The reasons for exclusion of patients must be documented before database-lock/data unblinding. It should compare treatment groups, with respect to the frequency and time to such occurrences. This analysis does, of course, give an optimistic effect estimate.

Tablel Use of intention to treat and other methods to analyse trial of coronary artery bypass surgery and medical treatment for stable angina pectoris in 768 men.² Mortality 2 years after randomisation is shown by allocated and actual intervention*

The above table gives results from the ITT and PP analyses of a parallel group study of 768 patients with stable angina pectoris, and shows how far final results may be different (European Coronary Surgery Study Group, Lancet 1979; 313: 889-93).

Another example is taken from the 1994 Eur J Gastro Hepatol study (1994; 6: 1135-9) of the Dutch omeprazole MUPS study group. Single gastric ulcers > 5 mm was the inclusion criterion. Randomization was to lansoprazole (n = 60) or omeprazole (n=66). Outcome was endoscopic healing. The main results of the two analyses are given.

Results: healed not healed

ITT PP

lansoprazole 56 0 4

omeprazole 52 2 10

PP 100 % versus 96 % healed (p > 0.05)

ITT 93 % versus 82 % healed (p=0.05)

The recommendation is, thus, to perform both analyses, with the hope for robustness of the conclusions. If otherwise, then discuss the differences. The European Medicines Agency and American Food and Drug Administration both expect access to all data and may audit them.

We should emphasize that superiority confirmatory trials & treatment-strategy trials are always ITT, and that equivalence trials and non-inferiority trials are PP, while ITT analysis is anti-conservative here. Anti-conservative here means that the treatment effects are minimized.

LOCF (last observation carried forward), BOCF (best.....), and WOCF

(worst.......) , are some traditional strategies for performing ITT regimens. Three more examples of studies with protocol violators after randomization are given underneath.

First Example (BMC 2007; 7: 3-10)

Second Example (Curr Med Res Opin 2008; 24: 2151-7)

Third Example (Neth J Med 2015; 73: 23-9)

In the above three randomized controlled trials, the numbers of protocol violators were 1-5 %, which is a small number. Obviously, the difference between the results of a PP and ITT analysis may, in practice, be pretty small.