Dear Professor Campbell:
The question of intention-to-treat analysis has come up before on this list. Below are some comments and references I provided earlier. I don't know what the specifics are in your particular RCT concerning either efficacy or costs; however, the consensus of many who have studied the intention-to-treat problem is that there are multiple questions that can be asked about a trial. Some can only be answered by an intention-to-treat analysis. Others require a separate analysis of patients who received and complied with a treatment. While the primary question may require an intention-to-treat analysis, there may be good reasons to carry out other analyses to answer other questions, and this is appropriate and desirable.
My previous comments:
Intention-to-treat analysis can best be thought of in terms of compliance.
There is no one correct way to analyze a trial; rather, one analyzes
different ways for different purposes. If one wants to know the average
result of applying the treatment to a population, then one uses
intention-to-treat analysis and includes all patients initially assigned to
the treatment or control group, regardless of whether they complied or
dropped out or crossed over. However, this obscures the effect of treatment
on those who comply, which is also a valuable thing to know. If I am a
patient, I want to know what to expect from the treatment if I respond to
treatment and comply. Averaging in a bunch of patients who don't respond or
don't comply confounds this result.
Statisticians tend to prefer intention-to-treat analysis, because it
maintains the purity of the original assignments. Clinicians and
manufacturers promoting an intervention tend to prefer to present results
only for those who respond and comply, to show the maximum effect. The
practical compromise is to present an intention-to-treat analysis as the
primary result, and to do a secondary analysis on those who respond and
comply. The proportion who respond and comply are themselves valuable
pieces of information that should be published.
The bottom line is that it is not necessary or desirable to do one or the
other; do both.
There was a set of articles discussing this in detail and proposing the "do
both" approach:
Efron B., 1998, Stat Med 17:259-50.
Urquhart J and De Klerk E, 1998, Stat Med 17:251-67.
Rubin D, 1998, Stat Med 17:371-85.
Cox D, 1998, Stat Med 17:387-89.
Also see:
Demissie K et al., 1998, J Clin Epidem 51(2):81-91.
David L. Doggett, Ph.D.
Senior Medical Research Analyst
Health Technology Assessment and Information Services
ECRI, a nonprofit health services research organization
5200 Butler Pike
Plymouth Meeting, PA 19462, USA
Phone: (610) 825-6000 x5509
FAX: (610) 834-1275
e-mail: [log in to unmask]
-----Original Message-----
From: M.J.Campbell [mailto:[log in to unmask]]
Sent: Friday, August 02, 2002 11:14 AM
To: [log in to unmask]
Subject: Intention to treat for cost minimization studies
Dear all,
Everyone knows that the optimal strategy for analysing an RCT to
compare two interventions is by Intention to Treat (ITT). This means
that you analyse the study by what treatment the subjects were
randomised to, and not what they actually received. Most also know
that if one is trying to prove equivalence of treatments rather than a
difference, then ITT is not a good idea, since protocol deviations
will bias the results towards what you want to prove.
Suppose you have a clinical trial with clinical and cost outcomes.
Your expectation is that the two treatments will be clinically
equivalent, but you want to prove that one is cheaper. Should you
analyse by ITT? For the clinical outcomes you have an equivalence
trial and for the cost outcomes a difference trial.
Have others come across this and what should one do?
Mike
Professor Mike Campbell
Institute of Primary Care and General Practice
University of Sheffield
Community Sciences Centre
Northern General Hospital
Sheffield S5 7AU
Tel 0114 271 5919
Fax 0114 242 2136
e-mail [log in to unmask]
http://www.shef.ac.uk/michaelcampbell/
http://www.shef.ac.uk/medical-statistics/
|