Is anyone done joint analysis of sensitivity and specificity when studies report multiple thresholds? More precisely, in which software? Whether there are such possibilities in STATA?
I already read the paper: Dukic and Gatsonis. "Meta-analysis of Diagnostic Test Accuracy Assessment Studies with Varying Number of Thresholds" and I see they use BUGS code. Someone with experience in this? Some tutorial or something similarly?
Can you not do it using -metandi- in Stata? Or is your concern more that you have multiple thresholds in each study, so you've got an extra level of variation? I guess you could use -xtmelogit-, which is what's called by -metandi- anyway. Or -mvmeta- might be somewhere else to look. Or else use the WinBUGS code, called from within Stata (there are several packages available to link Stata and WinBUGS).
To me, the issue after "pooling" publications about a certain test, even though the different papers give account of an array of cutpoints, is to whom your results will be useful?.
Attempting such a meta-analysis is a common comon place that poses two questions: a) why there is not a based-on-the-literature agreed cutpoint before -before¡- carrying out the study?. Don´t it suggests that there are frequent forms o "revision" bias?. There are many issues related to papers on test behaviour and their indexes. A meta-analysis -if possible- would deal with a lot of these issues, not always included inthe text.
@ Darren C Greenwood
My main concern is multiple thresholds in each study and in some cases different thresholds between studies (I am talking about index test, reference test is fixed).
What I have in mind is two different approaches:
1. simplified: to choose only one threshold per study, one whit the highest diagnostic odds ratio (DOR) as a criterion.
potential problem: Whether it is a good criterion? DOR is an important element in meta‐analytic model building but he has little direct clinical relevance.
2. I can use all available thresholds per study and I can extend the HSROC model to analyze all thresholds at once.
potential problem: clear intuitive interpretation for clinicians; lack of my personal experience ("technical how to" problem) and generally low experience with such models.
p.s. To be more precise I identify diagnostic studies with variety of diagnostic strategies involving multiple sampling with different timing for rule-out and rule-in (0-1h, 0-2h, 0-3h ... using delta values between last and first test as threshold for rule-out and rule-in). Some of them using different absolute delta, some diferent relative delta. The question is whether reducing the minimal sampling interval significantly reducing diagnostic accuracy? (searching the optimal timing).
@Patricio Herrera
I am trying to present the results to be comprehensive (and helpful) to clinicians. But it often comes to a trade off between rigorous methodological approach and clinical relevance of the results (as in this case). Let say I can use highest diagnostic odds ratio (DOR) as a criterion in first approach but even DOR is an important element in meta‐analytic model building it has little direct clinical relevance. On the other hand I can use "sum of the number of false negative and false positive results" which is much more interested to clinicians but that will "cost me" a lot methodologicaly speaking.
My concern is based on empirical evidence: same authors use PCR with different thresholds for different "diagnostic" situations (what means, among other issues, different gold standards). Why is that?. Besides the methodological issues, is the cost of false positive and false negatives what is important to clinicians. In my country, each clinical setting has its "own" cutpoint. Publications from that source do not exist but just a few. Yet this test is performed almost to every child with fever. I can undersatnd a different issue: In public settings this test means an incremental cost. In private settings it leads to incremental benefit.
- Badges
- Science topic