>There's a web site by Amanda Burls that gives some
>explanation about the "intention to treat analysis".
>http://www.public-health.org.uk/casp/statistics.html
>but I don't have it clear either: intention to treat
>weakens the observed effect or not?.
>
>What means "using intention to treat analysis only weakens
>the observed effect of an intervention and does not make
>an intervention appear effective when it was not. That is
>to say, it does not undermine any observed association
>between an intervention and outcome -it makes it more
>believable"
I have a web page that talks about intention to treat (ITT). The page needs
some updating and some references, but the basic ideas are still worthwhile.
The issue of dropouts is very complicated, but you can understand the value
of ITT if you examine it solely from the context of compliance. ITT does two
things: it prevents bias and it gives a more pragmatic indication of the
effect of a therapy in the real world.
ITT prevents bias by preserving the randomization that you went to all the
trouble of generating. There are stories about doctors who take a sterilized
coin into the operating room to insure proper randomization.
Randomization implies that neither the patient nor the doctor have any
control over which treatment the patient receives. If you allow the
patient's choice about whether to comply or not to influence which group the
patient is assigned to, you lose randomization and all of the benefits that
accrue to randomization.
Bias is especially troublesome when one of the two treatments being examined
is more likely to have compliance issues.
Consider a trial which compares a surgical intervention with a non-surgical
intervention. Some patients might die prior to surgery. This could be
considered an extreme form of non-compliance. But if we excluded the
non-compliant patients from analysis, we would be taking the rapidly dying
patients out of the surgery arm of the trial but not out of the non-surgical
arm. There's an interesting published example of this happening, which I do
not have in front of me.
ITT can go both ways, but in general, the effect of a therapy in ITT
analysis is less than when you exclude patients who are noncompliant. The
reason for this is that noncompliant patients tend to have worse outcomes
than compliant patients.
The amazing fact is that patients who fail to comply with a placebo do worse
than patients who do comply with a placebo. There is a reference for this,
but I don't have it handy. Why does this happen? One theory is that
noncompliers with placebo also tend to have other bad habits. They don't
care for themselves as well as compliers would. So excluding noncompliers
would paint a rosier picture than you might expect.
The second issue is the pragmatic one. As a doctor, you do not have the
option of treating only the compliant patients. You have to take all
patients as they come to you and treat them as best you can. If a new
therapy is twice as effective as a standard therapy, but the side effects
cause two out of three of the patients to stop taking the drug, what would
your best course of action be? If you had ESP and knew in advance which
patients would be compliant, life would be easy. But if you can't predict
compliance, then you have to factor in the impact of noncompliance on
efficacy.
A good example is assessing the value of smoking cessation efforts. We know
that stopping smoking will reduce your risk of heart attacks, but how
effective is it for a doctor to recommend a smoking cessation PROGRAM? We
can't judge the effectiveness of such a program by only looking at the
successful graduates of that program (i.e., the compliers). Since some
patients in a smoking cessation program keep on smoking and some patients
outside a smoking cessation program quit on their own, this tends to dilute
the effect of smoking cessation. There's a second level of noncompliance.
Not every patient that you tell to attend a smoking cessation program will
actually attend.
The question becomes "what do you want to measure?" Do you want to measure
the effectiveness of smoking cessation? of the smoking cessation program? of
the advice to join a smoking cessation program?
The issues are more complex than I have described them here, but the ideas
are worth noting nonetheless. ITT seems very counterintuitive, but there are
actually some very good intuitive justifications for using ITT.
You can find my web pages on ITT on the "Ask Professor Mean" section, which
can be found at
http://www.cmh.edu/stats/ask.htm
I hope this helps.
Steve Simon, [log in to unmask], Standard Disclaimer.
STATS: STeve's Attempt to Teach Statistics. http://www.cmh.edu/stats
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|