This blog is closed to new posts due to inactivity. The post remains here as part of the network’s archive of useful research information. We hope you'll join the conversation by posting to an open topic or starting a new one.
Today yet again I was confronted with a heated discussion on the usefulness of intention to treat in different trial settings. ITT has been used as ethical and scientific standard practice but is not always applicable. Situations vary. Consider this; A trial evaluating the efficacy of two types of progesterone support in first trimester of pregnancy after Intravenous Fertilization is confouded by whether implantation will take or not. If percentages of implantation vary between the arms then ITT analysis done on the basis of randomization at enrolment is clearly useless or misleading. Therefore if the outcome is affected by an intermittent event occuring after randomization which we have no control over affecting both arms then an ITT analysis becomes invalidated. What do you think?
This has been a issue for some time now. It was recently discussed by very well known and experienced statisticians at a meeting on clinical trial guidelines Toronto and I don't believe a consensus was agreed upon.
As mentioned by Marcel, the true benefits of randomization can only be exploited using an ITT analysis. "Tweaking" the ITT no longer guarantees a true randomized estimate as other factors come into play. The other advantage that has been suggested is in a situation where subjects who adhere to the protocol maybe different those who don't in terms of outcome than. Excluding subjects based on this as is done in a PP analysis could introduce bias.
I have not come across a clear definition of modified ITT analysis. I have seen it being described as a PP analysis and also as an ITT where outcomes for lost to follow-up subjects were imputed.
I agree with Raymond as well in that I would like to see results from an ITT analysis followed by mITT or PP analysis. Consistent results would make the data more believable.
Interesting comments from Patrice, Raymond, Ropar, Marcel and Jagarwal. I believe your submissions stem from fair statistical practice and evidence based medicine.
It's however, as many authors have described, common to see clinical trials reporting on what they call MODIFIED INTENTION TO TREAT. I believe that this is an ambiguous statement or misnomer if you allow me because ITT can never be modified. The other interesting thing about these "modifications" is that they are usually done in such a way that they bring out certain aspects of the data. This manipulation of data analysis to fit certain results is called modified ITT. It may not be a primary end point but could be anything else in name. The fact that it's included in an analysis creates or plants seeds in the mind of the reader. This is fine but consideriing that modified ITT may not even be any close to PP means that modified ITT is a cleaner word for "biased sensitivity" analysis.
I am not an experienced statistician. I am a clinical researcher from India, but I recall from my training that ITT and PP both have their place. We should be reporting both on our papers and then all is clear. ITT makes the randomisation stronger but PP does seem to make more sense on occasions like the one Moses has given us.
In general, I think ITT should be the norm. ITT is the only analysis which exploits all the benefits of randomization.
The further we deviate from ITT, the closer we get to an observational study.
In your example, can it be absolutely excluded that the type of progesterone support the woman gets affects the chance of implantation? If no, ITT will be the way to go as it measures a "real-life" benefit. Stratifying your analysis by an intermittent post-randomization event (e.g. implantation or not) is a risky business.
You are randomizing. Thus if progresterone support doesn't affect the chance of implantation, your two groups will tend to have similar chances of implantation, and the ITT analysis would be the way to go (plus you can do a supportive PP analysis which excludes non-implantations). If you are very concerned about chance imbalances, the suggestion by Ropar makes sense to me, i.e. randomize only after implantation.
I also agree with Raymond Omollo that generally I would like to see both ITT and PP results to be convinced the data is robust. ITT would usually be the primary analysis but for non-inferiority and equivalence trials the situation is less clear.
interesting example you have posted there. i just wonder would it not be better to randsomize from the number of women in which implantation actually took place as opposed to having to getyour study population from those who used in vitro fertilization? getting the numbers that are needed would be a bit of a challenge but i do think that if you did your randomization based on implantation then ITT would be feasible in this case.
There are always two sides to the coin whichever way one wants to look at it. Normally ITT is prefered as it gives conservative estimates of the outcomes. My experience with studies involving long follow-up is to present both ITT and PP results (this needs to be stated before hand in an analysis plan) so that one is able to see the differences btw. the two. More often than not they tend to be similar.
I fully agree. In some study designs, ITT is useful, in some others it may give misleading results (if true ITT analysis is performed).
For example, in studies in which a long follow-up period is required to have an efficacy endpoint. Such studies usually have more deviations and ITT will usually consider them as failures, which may lead to excessively pessimistic results.
This is a difficult area, and seems to always cause the hottest debates. I guess it is a matter of sensibly assessing every trial and seeing whether PP or ITT will provide the best assessment of the question and the truest answer. I would like to hear what others think - any stats people out there?