Teresa Benson posted:
> If you are using correlation coefficients to decide whether one
> clinical
> test can be substituted for another, particularly with non-dichotomous
> values (e.g., blood glucose), is there a certain minimum value you
> look
> for? (Assuming, of course, the new test is cheaper or easier.) Or do
> you always just look at things like sensitivity/specificity and
> predictive value, and disregard the correlation coefficients? The old
> User's Guide to the Medical Literature recommends a correlation
> coefficient of at least .8, but I'd like to know what others think
> about
> this.
>
I'd recommend sensitivity, specificity, positive & negative predictive values, LR+, LR-, and McNemar's chi-square test as being more meaningful than correlation coefficients for this purpose. For non-dichotomous tests, it also would be appropriate to examine goodness of fit throughout the entire reported value range (new test might perform well in certain zones but not as well as the other test in other zones). Probably wise to check the literature on diagnostic test assessment so as to have a reference supporting the approach you select.
David.
--
David Birnbaum, PhD, MPH
Adjunct Professor
School of Nursing
University of British Columbia
Principal, Applied Epidemiology
British Columbia, Canada
|