Best Meta-Analysis Methods for Diagnostic Test Accuracy Studies – How to Choose

accuracydiagnosticmeta-analysisrstata

I am conducting a meta-analysis of diagnostic test accuracy studies focusing on myocardial perfusion imaging.
I have used first Meta-Disc, but only for descriptive purposes, as it is clear that univariate approaches such as those provided by this package are biased (eg Takwoingi et al). Then, I have found the following bivariate methods, and used several of them:

  1. Bayesian bivariate model using bamdit in R;
  2. Bayesian bivariate model using meta4diag in R;
  3. Bayesian HSROC using HSROC in R;
  4. frequentist bivariate model using metamisc in R;
  5. frequentist bivariate model using metandi in Stata;
  6. frequentist copula mixed model using CopulaREMADA in R;
  7. frequentist hierarchical summary receiver operating characteristic (HSROC) model using metandi in Stata;
  8. frequentist proportional hazard model using mada in R;
  9. frequentist Reitsma model using mada in R;
  10. frequentist Reitsma model using Metatron in R.

Results are similar across many of these methods, albeit obviously not identical. Yet, I would favor the Reitsma model as available in mada as it gives me more comprehensive analytical and graphical results.

My questions stem from this actual project but are quite more general.

Is there a method which is best for meta-analysis of diagnostic test accuracy studies? Or are they more or less similar? Is there any other method not listed above which is better still?

Best Answer

@Giuseppe Biondi-Zoccai, okay, I got your question now. There are no simple criteria for choosing the best model among the many out there. There will always be many factors to be considered and to assess the models regarding different performance measures; one needs to do a simulation study. To my knowledge, the model of Reitsma et al. (2005) (the standard model to date) has repeatedly been compared with the other newly emerging models using simulations. And the articles report their results conditional on the scenarios being considered, including but not limited to sample size, number of studies and between-study variances. Hence, I would go into the packages and see which method they use, read the corresponding articles to see what they report in their simulation study to come up with the (closest) scenario for my data at hand and finally fit the best model to my data.

Related Question