Diagnostic test accuracy meta-analysis
Sensitivity, specificity, summary point and diagnostic odds ratio
Use this module when each study gives a 2x2 accuracy table: true positives, false positives, false negatives and true negatives. Follow the three-step flow, but step 2 is much shorter and step 3 is the most different of all the modules.
Step 1, the data
| Column | What it is |
|---|---|
study |
the study label |
TP |
true positives |
FP |
false positives |
FN |
false negatives |
TN |
true negatives |
In the comparative module there is one more column, test, with the name of the test, and each study appears twice, one row per test.
Extra columns are not used in this workflow. The Data check verifies the required ones, counts the studies and flags negative or blank cells.
Why this module is different
In diagnostic meta-analysis the outcome is bivariate: sensitivity and specificity move together and are correlated. A more sensitive test is usually less specific, because the threshold moves. That is why the model is the bivariate binomial GLMM, and not two separate models. That is the model the Cochrane Handbook for DTA Reviews v2.0 specifies in chapter 10.
This has a practical consequence: looking at pooled sensitivity alone gets you nowhere. The main result is the operating point, sensitivity and specificity together, with their confidence intervals.
Step 2, the parameters
This is the shortest step 2 in MetaVidence, because most of the choices in a bivariate model are not configurable field by field.
Single test module
Outcome name goes into titles and file names.
Analysis model (default Random-effects) affects the sensitivity, specificity and diagnostic odds ratio forest plots.
The Reitsma bivariate model is always random-effects, by construction. Switching to Fixed-effect changes the forest plots, but neither the summary point nor the summary operating point.
There is no sm, Method, Tau method, prediction interval or group name here: none of those fields applies to the bivariate model.
Comparative module
Besides Outcome name, there are Test A and Test B, which fill themselves with the names found in the test column. They define which test is compared against which, and the order the two appear in the figures.
Step 3, export and results
Single test
Sensitivity forest plot and Specificity forest plot, one for each, with sensitivity computed as TP / (TP + FN) and specificity as TN / (TN + FP).
Summary point, with the studies, the pooled sensitivity and specificity, a 95% confidence region and a 95% prediction region. There is no curve and no AUC: those belong to the HSROC model, which the handbook notes cannot be fitted in R with lme4. The bivariate model estimates a point, so a point is what the plot shows.
The forests deliberately carry no pooled diamond. Averaging sensitivity and specificity separately ignores the correlation between them, which is what section 10.2.3 of the handbook rules out. Every pooled number in this module comes from the bivariate fit, and there is only one of each.
Diagnostic odds ratio, with its own forest plot. The DOR summarizes accuracy in a single number, combining sensitivity and specificity. It is useful for comparing tests, but it says nothing about where the balance between the two errors lies, so it does not replace the sensitivity and specificity pair.
Threshold effect, the correlation between sensitivity and the false-positive rate across studies, with the Deeks test for small-study effects. A positive correlation suggests the studies used different positivity thresholds, which is the confounder specific to diagnostic accuracy.
To the bivariate model a subgroup and a head-to-head test comparison are the same thing: a covariate. The handbook says so plainly, calling them both “group”. That model lives in the comparative workflow, so splitting one test by a column belongs there rather than being duplicated as a weaker version here.
The app warns when there are fewer than 4 studies: the bivariate model estimates 5 parameters and converges anyway, but the estimate becomes unstable. Converging is not the same as being trustworthy.
Comparative
Sensitivity by test and Specificity by test, the forest plots split by test.
Comparative summary points, both tests on the same plot, each with its own confidence region. Both come from one joint model, grouped by study, so a paired design keeps its pairing.
Overall comparison, Sensitivity comparison and Specificity comparison: three likelihood-ratio tests, one for overall accuracy and one for each component. Having the overall test significant and the two specific ones not significant is common and informative: the tests differ in the balance between the two errors, not in either one alone.
Each study must have assessed both tests in the same population. Studies with only one of the tests stay out of the joint comparison. Comparing tests assessed in different populations means comparing populations, not tests.