dear Steve,
This answer comes too late for your teaching at Wednesday.
To make clinical sense out of sign. different means in outcome
on the level of a population can be done by judging the proportion of
patients that improve at least the mininal clinically relevant
difference. Confidence intervals can be recalculated to such
proportions using the normal distribution or by using the real
distribution. The latter seems best and balances outliers at the
right (are included) with those to the left (are excluded). To make
sure no harm is done one could also give the proportion of
patients that get worse by the experimental intervention.
A wild guess is that a mean difference of 0.5 - given a MIREDIF of
2.0 - will result in a low proportion, say 20%, of patients that
improve, the majority will show no difference and a small number will
deteriorate.
Depending on the health problem and the treatment (chemotherapy for
cancer versus the advice to go jogging for sinusitis (it helps!))
this is a good treatment or a worseless treatment.
This trick is similar to the suggestion I give to present data on an
ordinal scale.
Nico van Duijn, GP-epidemiologist
Depart. General Practice
Division Public Health
Academic Medical Centre
University of Amsterdam
Amsterdam
the Netherlands
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|