1
$\begingroup$

I am trying to figure out how to perform a meta-analysis of diagnostic test accuracy studies and I have a doubt that is driving me crazy. I am using the package mada written in R and following the article "Meta-Analysis of Diagnostic Accuracy with mada".

My doubt is related to the bivariate meta-regression. The article presents an example comparing two subsets of data to investigate the efficacy of self-administered and interviewer-administered questionnaires. They used the factor type (SAQ or IAQ) as a covariate in diagnostic meta-regression and presented the output obtained:

enter image description here

According to the article, the z-value for the regression coefficient for the sensitivities is significant while the point estimate for the false-positive rates does not indicate any effect. However, I couldn't conclude the same by looking at the output and I don't know how to interpret it properly. The p-value for tsens.typeSAQ was 0.066 while for tfpr.typeSAQ it was 0.023. Doesn't it indicate the exact opposite of what the article concluded? Wouldn't the correct interpretation be that a significant difference was observed for the false-positive rates, but not for sensitivities (considering a significance level of 0.05)?

I am not really good with statistic and this doubt may seem trivial, but I've been trying to understand it for a while without success.

Thank you very much!

$\endgroup$
  • $\begingroup$ Note to future readers the typo in the vignette to which the OP linked has now been corrected as of mada version 0.5.10 $\endgroup$ – mdewey May 25 at 16:08
0
$\begingroup$

I think your point is correct, in the sense that nominal significance is met (p=0.023) for the false positive rate (FPR), which appears positively associated with self-administered questionnaires (SAQ). Conversely, sensitivity is positively associated with marginal significance (p=0.066) with SAQ.

This overall difference betweene questionnaire types is confirmed visually by inspecting the summary receiver operating caracteristic curves (SROC), with interview-administered questionnaires (IAQ) clearly outpeforming SAQ.

enter image description here

However, my perspective is that sensitivity and FPR (ie 1-specificity) are not independent features of the dataset, but of course highly correlated. This means that if you already reject the null hypothesis for the former, you may reject it for the latter (or the other way around, if it applies). Accordingly, the conclusion of the authors (despite their little inaccuracy) stays correct.

If you are interested, here you can find some other useful resources on diagnostic meta-analysis:

https://link.springer.com/article/10.1007/s12350-018-01485-y

https://www.springer.com/gp/book/9783319789651

https://www.teachepi.org/courses/meta-analysis-of-diagnostic-test-accuracy/

https://www.stata.com/meeting/spain13/abstracts/materials/sp13_plana.pdf

| cite | improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.