Modelling publication bias from a random effects model formulated as a mixture model

Abstract

Publication bias (PB) is one of the main threats to contemporary science. It refers to the effects of any factor that affects the representativeness of the published studies, with respect to the studies carried out. If studies with statistically significant results are more likely to be published, then the accessible results give a distorted picture, usually consisting of overestimating effect sizes. Since the meta-analysis, several techniques have been developed to detect, assess, and correct for the effects of BP. Almost all of these techniques assume a random effects model that has some flaws and weaknesses, already pointed out in other sources. In this communication we present some ways of modeling PB using as a base an alternative formulation to the classical random effects model. It is a formulation based on a mixture model that overcomes the flaws and weaknesses of the classic EA model. From the mixture model, PB can be understood through both inappropriate and undesired influences in the mixing distribution and intrusions from interfering distributions of an unknown nature.