Three types of dichotomised clinical trials, and how to combine them

I have just uploaded a new BayesCamp white paper on this subject to http://www.robertgrantstats.co.uk/papers/Taxonomy-MA-dichotomised-studies.pdf.

Abstract:

Outcomes in clinical trials are sometimes reported after dichotomising participants at some threshold into responders and non-responders. This can be a barrier to meta-analysis and evidence-based medicine, if some papers report mean outcomes, while others use various thresholds. This paper proposes a taxonomy of thresholds, parameterised in terms of a latent bivariate distribution. Bayesian models for these thresholds were tested by a simulation study and applied to a Cochrane
review of antidepressants in children. The simulation study found that bias and coverage were similar to a DerSimonian-Laird random effects meta-analysis, but the Bayesian model had much narrower credible intervals as the number of dichotomised trials grew. A further comparison was made with a model incorporating baseline statistics but omitting the dichotomised trials, which demonstrated that improvements largely arise from including all trials, not the baseline statistics. However, bias and coverage can be adversely affected by prior distribution locations and by shrinkage of study-level random effects, requiring sensitivity analysis. The re-analysis of the Cochrane review highlights a number of practical considerations and challenges.

All data, code and original image files are in a zip file at http://www.robertgrantstats.co.uk/code/bayesian-dichotomised-meta-analysis.zip.

This work was previously presented at JSM 2014 (Boston, MA, USA) and IWCEE 2017 (Rome, Italy).