[R-sig-Epi] OT - study design advice sought

[Neil Shephard]

Thanks to all who have replied so far.  Its been a while since I did
anything similar to this (my normal area is statistical genetics), but
its good to refresh these things every now and again.

The replies (two of which I've included below for other readers of the
list) have helped clarify some of my thoughts, and dredged up some
memories of differences between incidence and prevalence.

To answer two of the points below, one of the considered methods of
analysis was conditional logistic regression to take account of the
matching which has been done retrospectively.

Thanks again for all suggestions,

Neil


On 11/22/06, Brown, Dr Nigel <Nigel.Brown at qml.com.au> wrote:
> Hi Neil
>
> From a medical angle I would see the occurrence during one's lifetime of
> cardiovascular disease (CVD) is being used as a marker for some
> underlying increased risk of CVD and it is this hidden risk variable(s)
> that is being 'linked' to AMD.
> If an underlying increased risk variable is being postulated perhaps
> rather than using lifetime CVD events as the surrogate marker only
> 'premature' CVD should be the criterion.
> As for the method of statistical analysis I can have no formal informed
> comment (not being a statistician) but I imagine a regression technique
> using premature CVD events as independent variables and AMD as dependent
> variable (? With Logistic regression methods) could be discussed.
> Of interest in terms of data dredging is factor analysis (do throw
> things yet) as an exploritory method just using default methods and
> varimax rotation to see if CVD and AMD load on the same factor and what
> else they load with. I've seen some interesting but soft outcomes of
> this type on analysis when applied to multivatiate medical numerical
> result data sets.
> Regards
> Nigel (Pathologist)



On 11/21/06, BXC (Bendix Carstensen) <bxc at steno.dk> wrote:
> Neil,
>
> It is not entirely clear how your matching was done.
>
> Are the controls matched to  cases at the date of AMD diagnosis (which I
> realize must be a pretty soft thing). And how was the data collection
> done --- around the time of diagnosis?
>
> Or was the study done a long time after AMD diagnosis (for some).
>
> As you describe it, the latter seems to be the case, i.e. you have
> matched prevalent cases of AMD to prevalent persons in the population.
> In that case what you get out of you study would be PREVALENCE
> odds-ratios associated with various risk factors. In that case it makes
> perfectly sense to have variables that are subsequent to disease. The
> prevalences you will be addressing are namely a function both of the
> disease incidence related to factors prior to disease and to survival
> related to factors subsequent to disease.
>
> The former sampling scheme (incidence-density sampling) would give you
> odds-ratios that were interpretable as hazard rates, and of course
> variables referring to time after diagnosis would not make any sense.
> But by the mere nature of the sampling scheme they would not be there
> anyway, since sampling was AT time of diagnosis.
> _______________________________________________
>
> Bendix Carstensen
> Senior Statistician
> Steno Diabetes Center
> Niels Steensens Vej 2-4
> DK-2820 Gentofte
> Denmark
> +45 44 43 87 38 (direct)
> +45 30 75 87 38 (mobile)
> +45 44 43 73 13 (fax)
> bxc at steno.dk   http://www.biostat.ku.dk/~bxc
-- 
"An intellectual is a person who has discovered something more
interesting than sex." - Aldous Huxley

Email - nshephard at gmail.com / neilshep at cyllene.uwa.edu.au
Website - http://slack.ser.man.ac.uk/
Photos - http://www.flickr.com/photos/slackline/