Thanks to all those who have replied.
Email me if you want any info.
Jon
[log in to unmask]
>
>
> Hello again,
>
> I'd like to follow up my original post (shown below)
> with another question.
>
> I now know that this is a legitimate method, Rothman and
> Greenland's book (Modern Epidemiology) describe how one can
> 'assess the percentage change in effect size' and they give
> an example where the unadjusted and adjusted risk ratios are
> compared with the calculation:-
>
> (1.44 - 1.33)/1.33 = 8% increase in effect
>
> The paper to which I refer below performed similar sums using
> odds ratios but assessed decrease in effect when a confounder
> was offered rather than an increase resulting from it's removal
> as shown above.
>
> My question/point is that in the above calculation, as well
> as in the paper, they are assessing %-change in odds or risk
> assuming a baseline of 0. Surely if a confounder reduces
> an odds ratio from 1.5 to 1.0 it has removed 100% of
> the effect not just 33% hence the correct calculation should be:-
>
> (0.44 - 0.33)/0.33 = 33% increase in effect.
>
>
> There may well be a glaringly obvious reason as to why I am wrong
> - I don't usually argue with textbooks. So any clarification would
> be apprecated.
>
>
>
> thanks
>
>
>
> Jon
>
>
> >
> >
> > Dear All,
> >
> >
> > I read a paper yesterday in which the amount of confounding
> > occurring was assessed by examining percentage change in Odds
> > Ratios between the exposure variable and the outcome
> > when each confounder was added.
> >
> > Sentences like, "confounder X accounted for 32% of the size
> > of effect" were then used.
> >
> > I am rather wary of this method - what happens when the OR
> > increases once a confounder has been controlled for? or worse,
> > what if a confounder reverses the direction of the effect?
> > 150% of effect accounted for? or some such rubbish.
> >
> > Having said this, In my current problem, I have a three-level
> > exposure variable so this method would be quite useful in
> > assessing the effect of confounding on the difference between
> > those three levels - e.g. this variable reduces the difference
> > between levels 1 and 2, but does not affect the difference
> > between 1 and 3 - something I don't think I could do purely by
> > looking at the likelihood.
> >
> > So my question is - is this a valid method or not?
> > Any thoughts/references would be gratefully recieved.
> >
> >
> > thanks
> >
> >
> > Jon
> >
> >
> > ============================
> > Dr Jon Heron
> > University of Bristol
> > ============================
|