I am involved in a trial comparing the withdrawal patients staying in an in-patient unit experience after discontinuing either of two drugs. The drug they were taking before they commenced abstinence will not be not randomised.

A reviewer on the ethics committee who are assessing the study enquired about whether we were following intention-to-treat principles.

I never even considered using intention to treat since:

(i) participants cannot switch drugs during treatment since they are only being measured after they have stopped (ii) participants were not randomised, the decision to use either drug was based on the preferences of the prescribing doctor and the participant.

But just say participants decide the withdrawal they experience is too intense and want to resume use of either the drug they were originally on or the other drug. If they agree to continue providing withdrawal symptom data should their withdrawal symptom data still be analysed as if they had never resumed use of the drug?

From what I understand from here and here intention to treat preserves the bias-removal effect inherent in randomisation.

But does intention to treat offer the same advantages when there is no randomisation?

As you can tell I am a bit confused and would welcome some clarity.


I think intention-to-treat should be considered in non-randomly assigned groups. Just thinking about a study of a programme to change behaviour (smoking). Group A nominated themselves for the programme, but 40 out of 50 of them did not complete it, it was sooo boring. It would be unfair to compare the 10 who did complete with the comparison group. The Intention-to-Treat principle is important because drop-outs skew your data, and if your intervention/drug is to be assessed comprehensively surely you need findings that account for the likelihood that people don't finish the intervention/drug prescription.

  • $\begingroup$ PS i would have to disagree with Jiwei on the problems with your non random assignment. You analyse the results in the same way as you do an RCT, results just come with the caveat - there may be a diff between these 2 groups. But your study surely adds insight, despite this $\endgroup$ – Gale Oct 8 '20 at 22:40
  • 1
    $\begingroup$ Thank you for your answer @Gale. Just trying to think this through: What if in your example the outcome measure was nicotine withdrawal during abstinence? Would you include the 40/50 people who resumed smoking in Group A's withdrawal data after they had resumed smoking for the remainder of the trial, assuming they continue to provide withdrawal data? If only 10 participants in Group B relapsed then an intention to treat analysis would conclude that Group A's intervention was associated with less severe withdrawal than Group B despite the fact that more people in Group A had relapsed... $\endgroup$ – llewmills Oct 8 '20 at 22:57
  • $\begingroup$ Jesus llewmills your a brain-melter! Let me try and get my head round that $\endgroup$ – Gale Oct 8 '20 at 22:59
  • 1
    $\begingroup$ ...Surely in this scenario including data after resumption of smoking would introduce more bias than a per-protocol analysis? I hasten to add I am not trying to say your suggestion was foolish or wrong. Far from it: I really appreciate the help. I am just trying to think it through. $\endgroup$ – llewmills Oct 8 '20 at 23:00
  • 1
    $\begingroup$ Intention-to-Treat melts my brain. Nicotine withdrawal reported on a 0-10 scale or something like that. $\endgroup$ – llewmills Oct 8 '20 at 23:01

I have only seen the term "intent to treat" in randomized trials, where you are analyzing the effect of treatment assignment. Not sure if it is the right term to use in this case. As for the ethical issues, perhaps they are worried these patients do not voluntarily discontinue?

Comparing the withdrawal patients from two different drugs in non-randomized setting sounds a very complicated analysis to me. You need to make sure the two groups are comparable in the first place after controlling for the confounding factors. Otherwise the results are difficult to interpret. Not sure if it is feasible in this case. Even if the two groups of patients are from a randomized trial for two drugs, the withdrawal patients are not comparable.

Should their withdrawal symptom data still be analysed as if they had never resumed use of the drug? Yes, I think so. If you are interested in the withdrawal symptoms, the data after resumed use of the drug may not be useful. Perhaps you can also do a time to failure analysis and consider the resumed use as a failure. That being said, I am not sure how to design this analysis properly in the first place.

  • $\begingroup$ Thanks for your comment @Jiwei He. The best we can get is some attempt at matching participants based on age, gender etc. It will be a pilot study whose main objective is to describe withdrawal from the two different drugs. Sample size will be too low for between-group inferences to have much meaning. Nevertheless longitudinal quantitative analysis will be done using repeated measurement of continuous withdrawal scales, so I need to think about ITT. We will most certainly be doing a survival analysis where failure of abstinence is the event, but this is not the primary quantitative analysis $\endgroup$ – llewmills Oct 8 '20 at 23:58
  • $\begingroup$ Sorry, I probably misread the question. I thought the question is asking about two groups of patients taking two different drugs respectively. If the group of patients were taking both drugs and discontinuing from either one of them, it makes sense to do a comparison between the two withdrawal groups, if you can do a good matching at the time of withdrawal (e.g. using PS related method). Perhaps patients who do poorly more intend to discontinue drug A versus drug B. Perhaps you should take into such factors into consideration. not just the baseline factors like age and gender. $\endgroup$ – hehe Oct 9 '20 at 2:50
  • $\begingroup$ No you are right that it is two groups of patients. Each patient takes only one drug or the other, not both. And which drug is not randomised. It is mainly a pilot study whose findings may justify applying for funding to run a well-powered and randomised design. $\endgroup$ – llewmills Oct 9 '20 at 5:01
  • $\begingroup$ I see. If that is the case, I do think analysis within each group makes more sense. Patients taking drug A versus drug B are two different groups with no overlapping. Not sure if this difference can be resolved by matching. $\endgroup$ – hehe Oct 9 '20 at 16:09
  • $\begingroup$ I take your point @Jiwei He, however, the fact that it is quasi-experimental rather than experimental research does not prevent running statistical models and estimating effects, but does prevent one from drawing any strong conclusions about causation. $\endgroup$ – llewmills Oct 10 '20 at 21:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.