On 11/03/2013 17:05, Steve Simon, P.Mean Consulting wrote:
> On 3/10/2013 5:20 PM, Piersante Sestini wrote:
>
>> However, when it comes to
>> statistical tests (for example, to compare the accuracy of two
>> different commercial assays), they all use metaregression, that works
>> on Odds Ratios. Of course, one cannot compute an odds ratio with
>> three outcomes. What is the usual solution here?
>
> There is no "usual" solution here because the methodology for
> systematic overviews of diagnostic tests is still evolving. Here are
> some thoughts.
>
> First, the likelihood ratio is mathematically equivalent to a relative
> risk. There are some fairly established methods for pooling relative
> risks. There is, however, an issue of heterogeneity. The very nature
> of a relative risk means that it has bounds that are dependent on the
> baseline risk. It is impossible, for example, to get a relative risk
> greater than 1.25 if the probability is 0.80 in the baseline group.
> You can probably be okay here, but it is an issue that you need to
> think about.
>
> As I understand it, a diagnostic odds ratio is likely to be the best
> way to analyze the data for a diagnostic test. The odds ratio, for all
> the criticisms it gets, has one big advantage of the relative risk,
> absolute risk, or number needed to treat. The odds ratio has no
> constraints that are dependent on the baseline risk.
>
> So that leads to your comment that meta regression won't work here.
> There is nothing in a three group design that prevents you from using
> meta-regression. It's basically a logistic regression model. In your
> indeterminant group, you can compute the log odds of disease. Do the
> same for the positive and negative groups. Logistic regression is
> funadmentally a comparison of log odds across multiple groups. So you
> can do this. It's not trivial, but only because meta-regression itself
> is not trivial. Once you figure out how to do meta-regression for a
> diagnostic test with two outcomes, it will not be much harder to
> figure it out for three outcomes.
>
Thanks. It didn't realize that I could compute three odds ratios, for
positive, negative and indeterminate results, just as I could compute
separate likelihood ratios.
>
> I hope this helps. Good luck with the systematic overview. I'd be
> curious to hear how it all comes out.
>
Thanks. The more challenging point so far is try to distinguish between
truly prospective cross-sectional studies of diagnostic accuracy, where
patient are enrolled based o clinical suspect, and retrospective
studies where patients are identified on lab logs and clinical
characteristics are extracted from clinical records. Of course, in
retrospective studies the index test was known to the attending
physician and could have influenced clinical workout (and probably the
final diagnosis), introducing bias. Wording in the reports is often
ambiguous, and blinding is seldom mentioned. So far, it seems that there
is no difference in accuracy between studies which are clearly labeled
as retrospective and the rest, but this is not necessarily reassuring.
thanks again
Piersante Sestini
> Steve Simon, [log in to unmask], Standard Disclaimer.
> Sign up for the Monthly Mean, the newsletter that
> dares to call itself average at www.pmean.com/news
>
>
|