Meta-analysis of diagnostic studies: SROC modelling - the profile log-likelihood approach based upon the proportional hazard model

Tags: Poster
Charoensawat S1, Böhning} D2
1Udon Thani Rajabhat University, Thailand, 2University of Reading, United Kingdom

Background: The number of meta-analysis of diagnostic studies is increasing and models that deal with the summary receiver operating characteristics (SROC) approach have become popular. These models have become complex and difficult to use which led us to develop a simple model, the proportional hazard model (PHM).

Objective: To improve SROC modelling based upon PHM by finding the most efficient estimate of the diagnostic accuracy parameter θ.

Methods: A simulation study was done to investigate the performance of the profile log-likelihood (PMLE), the adjusted profile log-likelihood (APMLE) and the iterative weighted least square (IWLS) approaches. The performances of approaches were found from score function and the Fisher information I(θ). The heterogeneity was considered and the APMLE with the random effect was developed. Maximum likelihood estimation was used to find the estimate of θ and tau2, two parameters in the model. Also, the variance-covariance matrix was used to determine the inverse of I(θ) and I(tau2) and to find the estimate of θ and tau2 of I(θ). Meta-analysis of diagnostic studies of exercise echocardiography (ECHO) and exercise single photon emission compute tomography (SPECT) imaging in the diagnosis of coronary artery disease was exemplified under homogeneity and heterogeneity.

Results and Conclusion: PMLE turned out to be an easy approach to eliminate nuisance parameter from the model but a normalizing constant was missing. APMLE has been suggested for correcting the shortcoming of the PMLE. In the simulation study, likelihoods behaved similarly, but the APMLE was the most efficient estimator. In our example of meta-analysis of diagnostic studies, θ of ECHO was smaller which meant it had higher diagnostic accuracy than the SPECT device, since the smaller the θ the higher the diagnostic power. p - values became non-significant when tau2 was incorporated.