Nos tutelles

Nos partenaires

Rechercher




Accueil du site > Séminaires LNC > séminaires à venir (upcoming seminars)

Agustin Lage (Centro de Neurociencias de Cuba, La Habana)

Causes for low reproducibility in fMRI experiments : evidence in favor of type II errors

Several hypotheses have been prposed in order to explain the low reproducibility of fMRI data (Ioannidis, 2005). Initially, the lack of rigorous statistical procedures such as multiple testing corrections was considered the main cause. In this line of thought, non-independent analysis and double dipping have been pointed out as dishonest procedures to avoid stringent multiple comparison corrections (Kriegeskorte, W, Bellgowan, & Baker, 2009 ; Vul, Harris, Winkielman, & Pashler, 2009). Next, the large analytical flexibility of fMRI analysis pipelines has also been addressed as a potential source of false positive findings (Carp, 2012). These explanations consider that the cause of low reproducibility is type I error : false positives (Nichols, 2012). In a different perspective, Yarkoni (2009) identified type II errors as an alternative explanation for non-reproducible results. The argument is that low sample size combined with stringent multiple comparison corrections lead to effect size inflation and to the impossibility of discovering activations of medium and small effect sizes. As consequence only high specific task-region associations are reported, that is only the “top of the iceberg” of a pattern of distributed medium and small effects size activations (Lieberman & Cunningham, 2009). This low-power scenario supports the localizationist view of brain function and undermines the identification of extended networks of brain-task relationships, thereby promoting an oversimplified view of cognitive processing. In this talk we provide new evidence in favor of the “low power” explanation. Recently proposed frameworks for automatic meta-analysis of large data sets and the use of public databases of fMRI experiments are presented.