Structural and parametric uncertainties in full Bayesian and graphical lasso based approaches: beyond edge weights in psychological networks

摘要

Uncertainty over model structures poses a challengeudfor many approaches exploring effect strength parameters atudsystem-level. Monte Carlo methods for full Bayesian modeludaveraging over model structures require considerable computationaludresources, whereas bootstrapped graphical lasso and itsudapproximations offer scalable alternatives with lower complexity.udAlthough the computational efficiency of graphical lasso basedudapproaches has prompted growing number of applications, theudrestrictive assumptions of this approach are frequently ignored,udsuch as its lack of coping with interactions. We demonstrateudusing an artificial and a real-world example that full Bayesianudaveraging using Bayesian networks provides detailed estimatesudthrough posterior distributions for structural and parametricuduncertainties and it is a feasible alternative, which is routinelyudapplicable in mid-sized biomedical problems with hundreds ofudvariables. We compare Bayesian estimates with correspondingudfrequentist quantities from bootstrapped graphical lasso usingudpairwise Markov Random Fields, discussing also their interpretationaluddifferences. We present results using synthetic data fromudan artificial model and using the UK Biobank data set to exploreuda psychopathological network centered around depression (thisudresearch has been conducted using the UK Biobank Resourceudunder Application Number 1602).