Mark Boye wrote:
I am seeking to analyze time from tooth extraction to time to a 30%
reduction in pain with 4 active treatment arms and a placebo control in a 24
hr randomized, double-blind clinical trial. The dilemma is that after 1
hour, patients in sufficient pain are permitted to drop out of the study and
take rescue medication. Hence, this appears to be a case of informative
censoring.
Can list members suggest how to estimate an appropriate model in STATA (or
SAS)?
----------------------------------------------------------------------------
The late Lewis Sheiner had worked on this issue at one time (L. B. Sheiner,
A new approach to the analysis of analgesic drug trials, illustrated with
bromfenac data. _Clin Pharmacol Ther_ 56:309-22, 1994).
I take it that you would rather avoid assigning zero to pain-relief scores
after rescue (or assigning the corresponding maximum if you used pain
scores).
Would you be able to use -gllamm- in a mixed-response model, handling the
is-rescued status as binomial at each observation time point, and the pain
scores or pain-relief scores as ordered probit (or normal, if warranted by
the scoring system) at each corresponding observation time point,
missing-valued after rescue? Consider also the possibility of handling the
right-censored time-to-rescue in an exponential model in a mixed-response
set up in -gllamm- with, say, summed pain-relief scores until rescue or
censoring as the other response variable. See the user's manual and other
documents on the gllamm.org website. If you or your colleagues in the
company are more familiar with or feel more comfortable with SAS than Stata,
then someone on SASList ( www.listserv.uga.edu/cgi-bin/wa?A0=sas-l ) might
suggest something using PROC NLMIXED.
Perhaps the study's protocol contained a passage that defined a therapeutic
success as a case in which such and such is present but not such-and-such
else, and that defined a therapeutic failure as otherwise. If so, then with
the outcome as either success or failure, and time to 30% reduction in pain
redefined accordingly, you could use conventional methods for modeling
binary response over time, for example, -xtlogit , re- or -xtgee , i() t()
family(binomial) corr()-.
It seems to be a conventional study design for clinical evaluation of this
type of medication. If your prior attempts turned up empty, perhaps a
citation-trace search of the medical literature starting with Lewis
Sheiner's article above might lead you to prevailing approaches to modeling
in these circumstances, if there are any.
Joseph Coveney
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
