Dear all,
Thanks a lot to those who gave suggestions.
Here there is a summary of the responses I had to my original question:
__________________________
Dear All,
I hope to receive suggestions about the situation I'm going to explain.
Consider a clinical trial aiming to assess treatment efficacy (two arms) in
reducing two primary endpoints:
1. mortality
2. combined endpoint defined as first occurence of: death, hospitalization
for HF
Suppose that at some point, we are interested in analysing whether a
specific factor (physiological or biochemical) is predictive
of hospitalization for HF(only) . Therefore we are considering as EVENT
those patients who had an hospitalization for HF and as NON EVENT those who
hadn't. In this way, those patients who died without being Hospitalized for
HF will be NON EVENTS. I would like to draw the attention on the fact that
people who died (and therefore more seriously compromised) will be mainly
in the NON EVENTS group.
Do you think that we are under-estimating the predicitve value of the
factor considered ?
How would you consider the patients who died? Does it make sense to
consider them in the NON EVENT group?
Thanks
Simona
__________________________
Dear Simona,
My guess is that you are mainly interested in hospitalization of people who
are at risk to get hospitalized, i.e. who are alive. Following this you
might want to assess the prevalence of hospitalization given survival. So a
subgroup analysis for those who have survived seems sensible to me. Thereby
the worse treatment doesn't get more NON-EVENTS. Still, there is the
possibility, that the worse treatment causes death for all but the strong
ones, resulting in death ,or alive and well only, but this is more a matter
of interpreting the data than of analysis. Hope that helps.
Best wishes
Thomas
---
Thomas Hotz
Research Associate in Medical Statistics
University of Leicester
United Kingdom
Department of Epidemiology and Public Health
22-28 Princess Road West
Leicester
LE1 6TP
Tel +44 116 252-5410
Fax +44 116 252-5423
Division of Medicine for the Elderly
Department of Medicine
The Glenfield Hospital
Leicester
LE3 9QP
Tel +44 116 256-3643
Fax +44 116 232-2976
_________________________
Hi Simona,
I try to always include all deaths as poor outcomes as this is more
robust. But it isn't always popular with medics. You might find the
following reference useful...
Lubsen J,.Kirwan BA. Combined endpoints: can we use them? Statistics
in Medicine 2002;21:2959-70.
Steff
---------------------------------------------------
Stephanie C. Lewis, PhD MSc CStat
Medical Statistician
Bramwell Dott Building
Department of Clinical Neurosciences
Western General Hospital
Crewe Road
EDINBURGH Tel: +44 (0) 131 537 2932
EH4 2XU Fax: +44 (0) 131 332 5150
UK Email: [log in to unmask]
_____________________________
Dear Simona,
well, I would use a kummulative endpoint: "hospitalization for HF or worse"
(this includes the death definitely) or you use survival techniques and
death means censored: this patient cannot be hopitalized anymore. Then, the
dead patient doesn't give that much information as a healty fit guy.
(But I must admit: this kind of trials is not my special topic. )
Best wishes,
Steffen
____________________________
Hello Simona,
I'm afraid I don't have any immediate solutions - but I had the same
concerns myself over a comparable if slightly different analysis we were
doing. The main issue is whether the "censoring mechanism" is informative
and I suspect in the case you describe case that it is in some way, but
without knowing more of the medical background, can't comment further. In
any case, I'd be very interested to hear what advice you get, not least for
my own sake!
Regards,
Neil
Neil Walker
Westminster Primary Care Trust
50 Eastbourne Terrace,
London W2 6LX
_______________________________
Dear Simona,
I am currently struggling with a similar problem (and am still considering
death as censoring event for the event of interest as well, which is
unacceptable though possibly less influential in my work).
I have briefly investigated the issue and it seems it comes under
'illness-death' model (see e.g. Per Kragh Andersen, Stat in Med Res
2002;11:91-115 or Philip Hougaard 's book on Analysis of multivariate
survival data). I have also come across a related earlier citation
Abbott,R.D., Carroll,R.J., American Journal of Epid 1986; 123:728-35
Conditional regression models for transient state survival analysis.
I do favour the use of conditionality in the estimation because that is
what is happening.
At some point I thought that the task could be partially viewed as a
competing risk problem but I am less convinced now.
I suppose my biggest problem is that I can not find these methods
implemented in statistical packages as SAS/Stata though I am still
investigating.
I would appreciate if you could forward the suggestions you receive to me
as well.
Best of luck,
Boby
Borislava (Boby) Mihaylova
NHS Training Fellow
Health Economics Research Centre
Institute of Health Sciences
Headington
Oxford OX3 7LF, UK
Tel +44 01865 226753
Fax +44 01865 226842
[log in to unmask]
Simona Barlera, MS
Medical statistician
Department of Cardiovascular Research
Medical Statistics Unit
Pharmacological Research Institute MARIO NEGRI
Via Eritrea, 62
20157 MILANO, Italy
http://wwwcardio.marionegri.it/
Tel: +39-02-39014558
Fax: +39-02-33200049
e-mail: [log in to unmask]
___________________________________________
|