Date: Fri, 10 Sep 2010 17:13:27 +0100
Reply-To: John Whittington <John.W@MEDISCIENCE.CO.UK>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: John Whittington <John.W@MEDISCIENCE.CO.UK>
Subject: Re: ESTIMATE statement in Proc MIXED
In-Reply-To: <81F8139F381BE844AE05CA6525FF2AAE01EDBE9A@tpwd-mx9.tpwd.sta
te.tx.us>
Content-Type: text/plain; charset="us-ascii"; format=flowed
At 09:55 09/09/2010 -0500, Warren Schlechte wrote:
>In surveys, we often think that those who return the surveys likely
>respond differently than those who do. As such, we try to correct for
>this non-response bias. if we can detect trends in the response rates
>associated with time of return, we can model the missing data
>(non-returns). I wonder if you couldn't do something similar here,
>where the missing data are imputed based on a model of what you do
>observe. For example, maybe patients stop coming in once they get well,
>so missing data reflect the resolved cases. Or, maybe they get too sick
>to come in, so they reflect the terminal cases. In either case, looking
>at the data to see if some model explains the missingness may help you
>model the missingness.
>
>I'm no expert in this field, but I recently saw a talk where the authors
>used the idea of having model hyperparameters within a Bayesian setting
>to help account for the missing data, where missing data were not MAR.
I think that one certainly has to do something about 'informative' (or
potentially informative) missing data (i.e. not MAR) - since to simply
ignore it (i.e.treat it as 'missing') can lead to very biased and
potentially very misleading results. As I wrote in response to Dale's
comments, this is a situation which I come across frequently in the context
of clinical trials, but it is much more difficult to deal with (and has a
smaller literature) than missing data which is MAR.
In the context of clinical trials, my personal inclination is often that,
although widely criticised (because of potential bias), use of the 'last
observation carried forward' (LOCF) approach may be the least of the
evils. A common situation is that in which treatments are given to control
or modify a measurable or assessable 'outcome' quantity (e.g. blood
pressure, blood sugar, pain level or whatever). It will often happen that
serial measurements of the outcome show a progressive deterioration up to
the point at which it is deemed necessary or appropriate to remove the
subject from the trial (and treat more effectively), with the effect that
all subsequent measures are missing. If one uses the LOCF approach, all
subsequent measurements will be deemed to be the same as the one which
resulted in the subject's discontinuation - which perhaps reasonably
reflects the real-world situation in which one would not leave a patient on
a treatment if they were progressing to even less acceptable situations.
In a situation such as I've described, it may well be possible to model the
progressive deterioration of the patient, and thereby to obtain
extrapolated estimates of what results would have been obtained had the
patient continued to receive the treatment - and, in one sense or another,
I guess that's what most imputation methods would be seeking to
achieve. However, if those extrapolated values/imputations are unrealistic
in terms of the real world (i.e. they would never be allowed to arise, and
may not even be compatible with life), I have to question whether this
approach is necessarily appropriate, even if it is less open to
statistical-theory-based criticism than is LOCF.
That's how I see it, anyway.
Kind Regards,
John
----------------------------------------------------------------
Dr John Whittington, Voice: +44 (0) 1296 730225
Mediscience Services Fax: +44 (0) 1296 738893
Twyford Manor, Twyford, E-mail: John.W@mediscience.co.uk
Buckingham MK18 4EL, UK
----------------------------------------------------------------
| 
