Back to the Top
In an human PK trial, a few subjects showed very low plasma concentrations (
~ ten times lower than the average). What is the best way to deal with the
situation? Is it acceptable to exclude these outlier subjects from the
average evaluation of PK parameters by non-compartmental analysis ? or is
there any acceptable criteria for exclusion of subjects?
Thank you for your help, Ahuva Cern
Back to the Top
I would advise against excluding these subjects from analysis. If the
compound in question exhibits polymorphic drug metabolism, then your
perceived non-compliance could be due to the ultra-rapid metabolizer
phenotype! Please provide more information on this example, in particular
the metabolic pathways, assay LLQ, inter and intra-subject variability,
etc). Regardless, I would not exclude them from analysis although you can
do sub-group analysis to determine distributions.
Back to the Top
Dear Ahuva,
Is it possible to be more specific: a few mean what?
(How many out of what total number of patients) Is
the profile different (e.g., one vs two compartment)
or just the magnitude of concentration? From your
e-mail sounds like the profiles are not different?
Should be very careful with this averaging business
and comparing with average as well (in some cases
averaging the data may mask detail observation or
result in miss interpreting the data). What is the
nature of the compound (e.g., lipohphilic,
hydrophilic) and administration route? I think you
should first try to describe why this is happening
(e.g., different enzyme levels, activity, possibility
for food or concomitant drug interactions, differences
in protein binding) any scientific reason/
speculations, rather than call getting to the outlier
discussion? Is it possible that you may have two
types of responder in the population? Hope this would
initiate some thoughts.
Rostam
Back to the Top
This message is in MIME format. Since your mail reader does not understand
Dear Ahuva:
The danger in excluding data arbitrarily is that you may be ignoring valid
information, as Raj & Rostam suggest. One might expect the data to vary
randomly about a single population mean, but there are examples of compounds
that have two pharmacokinetic populations, fast & slow metabolizers. If data
are eliminated because they don't meet pre-concieved expectations, one may miss
the true pharmacokinetic characteristics.
Data should only be excluded based on criteria determined a priori, a
mathematical test for outliers, and an investigation into the likely cause of
the "bad" data. It would seem that you are in the investigation phase. Even
after all this, proceed with caution when excluding data.
Peter Gingras
Apotex
Back to the Top
Dear Peter,
I do agree that one should be very reluctant with regard to the exclusion
of outliers. Indeed, these"outliers" might be the most interesting and
informative measurements of the whole experiment. I would suggest that data
are only excluded if the cause of the deviation is known, e.g. the subject
vomited after the p.o. application of the drug, an error occured in the
sample preparation or (chemical) analysis etc. The best thing to do in this
case would be to repeat the "outlying" experiments. If one is interested in
modeling a homogeneous subset of the data this is OK. The data excluded
should be clearly stated and the reasons for excluding them given.
Additionally, in my experience many people underestimate the variance of an
experimental system.
You suggested to exclude data based on "a mathematical test for outliers".
This is a tricky business and depends on the method you use, assumptions
about the type of distribution, the estimate of the variance etc. It might
help to generate a "clean" homogeneous subset of the data but can't tell
you what's wrong with the "outliers".
Best regards,
Martin
Back to the Top
Hello all,
If I may add a comment from out on the front lines too: Ignoring
certain outlier events may ultimately lead to post-marketing recalls,
black boxes on package inserts, or other clinical misadventures
(which sometimes make the prescribers and dispensers look a little
silly to their patients or the crew from 48 Hours). As a dispenser
I'm a real fan of full disclosure.
Cheers,
Richard Molitor, R.Ph.
http://www.angelfire.com/wa/pharmacist
Seattle, WA
Back to the Top
Richard M.
Thanks for the morning smile. That's funny.
Seriously though, one of the fellows said it best in a prior message,
it's OK to
exclude data, but only as a homogenous subset of your original analysis. Full
disclosure is absolutely required.
Sometimes in complicated experiments, everything doesn't always go exactly as
planned. Here is a thought for you. In some cases, it can be midleading to
include all of the data.
For example, in a crossover, bioequivalence study, your method allows you to
analyze one phase of one subjects blood samples per day. On one day, one QC
sample concentration is slightly out of specifications, however the sample
results are within the range of variance of the other phase, and other subjects
data.
What do you do?? Analyze the data including that subject, and then reanalyze
the data excluding that subject. What do you do if pass or fail depends on
which analysis you rely upon?? It's always a good idea to have a priori
criteria for excluding data.
Cheers
Peter Gingras
PharmPK Discussion List Archive Index page
Copyright 1995-2010 David W. A. Bourne (david@boomer.org)