gms | German Medical Science

Meta-analysis of clinical studies reporting diagnostic results is a classical problem in medical statistics. The aim of this work is to develop a modeling framework for the typical case where studies report sensitivity and specificity, but study characteristics, e.g. study population or diagnostic set up information may be incomplete. We follow a full Bayesian modelling approach: test results are modelled as multinomial random variables, between study variability as a bi-variate normal distribution [Ref. 3]. The diagnostic odds ratio (DOR) and a summary of diagnostic threshold (S) are linked with incomplete covariate information by a system of regression equations.

Prior distributions for regression coefficients are taken as independent double-exponential distribution with common precision parameter. This strategy implicitly incorporate a L1-constrained fitting [Ref. 2]. Because the heavy tails of these priors the analysis produces posteriors of regression parameters concentrated around zero or posteriors with little probability mass on zero, making results more interpretable. Statistical computations are performed with Markov chain Monte Carlo (MCMC) methods.

Our approach is illustrated with a systematic review evaluating the potential diagnostic benefits of computer tomography scans in diagnostic of appendicitis [Ref. 1].