Wes' E-mail on Wed, 20 Sep 2006 11:43:38
Here is what I sent out to a dozen or so heavy hitters. Let's see what
they say.
Quick question: I'd like to know which of the following YOU would like
to see kept in an Intention-to-Treat (ITT) analysis for an RCT? No
strings attached or real agenda here. I just got into a healthy
discussion about this with our fellows this week and wanted to know your
opinion.
Three types of patients (answer YES or NO in the table below if you'd
keep them in an ITT analysis for an RCT):
1. Patient enrolled (consent signed) and never randomized with no
data collected due to death or immediate withdrawal by patient/family
2. Patient enrolled and randomized but no data collected or study
drug given
3. Patient enrolled and randomized and data collected but no study
drug given
Gordon' Reply
>
> 1. Patient enrolled (consent signed) and never randomized with no
>
> data collected due to death or immediate withdrawal by patient/family
No. Should be reported in CONSORT but not in ITT.
>
> 2. Patient enrolled and randomized but no data collected or study
>
> drug given
>
> 3. Patient enrolled and randomized and data collected but no study
>
> drug given
This depends a little on the study. TECHNICALLY, of course, these
should be in an ITT as the dropouts occured prior to randomization. If random assignment is not known to anyone (double blind) than dropouts AFTER RANDOMIZATION but before drug delivery cannot bias the results. If,eg, there is anyone, anyhow, anywhere, what arm the patient was randomized to (and this might be tough to exclude) then patients who drop out prior to randomization must be treated on an as randomized basis. THe same is tru for patients who are enrolled, randomized, no drug given but found
to be ineligible for the study. I guess as a practical matter I would analyze both ways and pray that it did not affect results.
Gordon
Taylar's reply
From: Thompson, Taylor,M.D. [mailto:TTHOMPSON1@PARTNERS.ORG]
Sent: Wednesday, September 20, 2006 12:15 PM
To: Ely, Wes
Subject: RE: Taylor, ITT question for you
Wes, Depends on the research question. See the attached from
http://www.tufts.edu/~gdallal/LHSP.HTM, which I think is a fair
description of the pros and cons. By the way, FACTT hand one patient
that fits your first scenario, and we dropped that patient from the ITT. Taylor
Article
Ely, Wes wrote:
>
Meredith, can you pull before you leave today and enter with Ref Man
>
numbers. Thanks.
>
author = {Stewart, William H.},
>
title = {Basing intention-to-treat on cause and effect
>
criteria
>
journal = Drug Information Journal,
>
year = 2004,
>
volume = 38,
>
pages = {361-369},
>
annote = {intent to treat;nice description of things
>
that can go wrong (e.g., protocol violations) and when it is reasonable
>
to exclude patients from ITT analyses} }
Wes forwarding Derek's reply
I think that some of you may have thought that Tim and Pratik and I were being funny about our thoughts related to the different studies and application of ITT in different study design...but maybe now you realize how much ambiguity there can be on this topic. I'll send you the final tally on votes tomorrow when a few more answer.
Wes
Original Message-----
From: Angus, Derek [mailto:angusdc@ccm.upmc.edu]
Sent: Wednesday, September 20, 2006 6:31 PM
To: Ely, Wes
Subject: coupla things ...
First, just got back to Paris after being in Pgh for a couple of days - without internet access (cancelled DSL line for the year). Anyways, got your note and am most most appreciative. When not running around, I'll send you back a longer response.
In the meantime, as to ITT: I don't really have an absolute view here. Rather, I think I gain information by understanding whether results change or not in sensitivity analyses around different versions of ITT. And, of course, the main objective is to have as few such cases as possible, so different versions should be an exercise with no meaningful consequence. Put another way: there's something fishy if the p-value changes materially when any of the 3 groups are included or excluded.
That said, if I had to draw a line in the sand, I'd probably include 3 in the ITT, accept that 1 is not in the ITT, and be ambivalent about #2 - obviously, if you include #2, the most conservative thing since no data were collected is to assume bad outcome for all treatment arm cases and good outcome for placebo cases. This action serves to keep the investigators on top of randomization and follow-up, for noone really wants to have #2 in the ITT!
And, I acknowledge that NONE of this is actually ITT. ITT is to include all 3 (or definitely everyone post-randomization, so therefore #2 and #3). Thus, it should be called MITT (modified ITT).
Derek C. Angus, MD, MPH
Final Pull
Team,
Here are the results of my informal poll/survey. I hope that
this helps. We have all discussed this very much lately and have some
good articles to place on the website along with this table. Several
mentioned the term MITT (modified intention to treat). Others stressed
that the correct answer depends on the research question being asked.
Still others stated that there just needs to be a distinction between
ITT and per protocol, which is actually a much simpler approach. In all
cases, it was mentioned that most times you elect to look more than one
way and that the most solid results are those in which the answer
doesn't change depending on the nuances of how you categorize /
re-categorize a few patients.
Thanks to all for this fun exercise!
Do other research teams take such time out for fun games? I
doubt it
Wes
Patient |
Enrolled |
Randomized |
Data collected |
Study drug given |
Yes or No for ITT |
1 |
Yes |
No |
No |
No |
13 No answers, 0 Yes answers 0 ? |
2 |
Yes |
Yes |
No |
No |
2 No answers, 8 Yes answers 3 ? |
3 |
Yes |
Yes |
Yes |
No |
1 No answers,10 Yes answers 2 |
Another e-mail from Wes
In Sept 14th NEJM, there are two interesting things related to our
recent RCT Stats discussions.
1. In the second article on drug eluting stents, there is a
reduction in mortality but not statistically significant. Much the same
as our VFDs in ABC, where the impressive 12 vs. 20 days are 0.07. In
this NEJM article, the CONCLUSION of the article in the abstract, the
most direct and pointed aspect of any article, states that the stents
"reduced the incidence of serious adverse cardiac events by 4 percent,
the reduction was not statistically significant." Tim, Pratik, read
this and see what you think for ABC and MENDS. Can you frame the
conclusion of your two abstracts? I would like to send them to NEJM,
JAMA, respectively, as first attempts to publish.
2. In the article on Echinocandins (excellent review of this
important area), a review of one of the papers in therapy showed that
for "patients who received even 1 day of therapy, a higher rate of
clearing of fungemia was found." Do you see which analysis this is? It
is "per protocol" for this drug study. In this case, all 3 of the
patients I used in the ITT survey that I sent out to you last night
would be excluded from the results.
Chance favors the prepared mind. Meaning that we just all have keep
learning, watching for good and bad examples of others. I think that
both of the above are perfectly acceptable ways of handling the data in
those circumstances, yet both are deviations on how most handle those
issues in slightly different circumstances.
Wes