Power calculations for multicenter imaging studies controlled by the false discovery rate

Abstract

Magnetic resonance imaging (MRI) is widely used in brain imaging research (neuroimaging) to explore structural and functional changes across dispersed neural networks visible only via multisubject experiments. Multicenter investigations are an effective way to increase recruitment rates. This article describes image-based power calculations for a two-group, cross-sectional design specified by the mean effect size and its standard error, sample size, false discovery rate (FDR), and size of the network (i.e., proportion of image locations) that truly demonstrates an effect. Minimum sample size (for fixed effect size) and the minimum effect size (for fixed sample size) are calculated by specifying the acceptable power threshold. Within-center variance was estimated in five participating centers by repeat MRI scanning of 12 healthy participants from whom distributions of gray matter were estimated. The effect on outcome measures when varying FDR and the proportion of true positives is presented. Their spatial patterns reflect within-center variance, which is consistent across centers. Sample sizes 3-6 times larger are needed when detecting effects in subcortical regions compared to the neocortex. Hypothesized multicenter studies of patients with first episode psychosis and control participants were simulated with varying proportions of the cohort recruited at each center. There is little penalty to sample size for recruitment at five centers compared to the center with the lowest variance alone. At 80% power 80 participants per group are required to observe differences in gray matter in high variance regions.

The neuro/PsyGRID calibration experiment

“…Moreover, underpowered studies are difficult to justify ethically. Maps of within-center variance derived from calibration experiments of the type described here are suitable for power calculations that predict key parameters of trial design [Suckling et al, , 2010. Furthermore, choice of acquisition protocol or data processing pipeline can be made based on outcomes of the power calculations.…”

“…Furthermore, choice of acquisition protocol or data processing pipeline can be made based on outcomes of the power calculations. By making these calculations on a voxelwise basis, the predictions are also adaptable to specific neurobiological hypotheses that may include regions with a range of within-center variances that in turn impact on organisational decisions [Suckling et al, 2010].…”

“…The first goal has been previously considered [Suckling et al, 2010] with image-based power calculations drawing upon within-center variance estimates of grey matter from tissue segmentation of T 1 -weighted, high-resolution structural MRI, where it was concluded that multicenter recruitment strategies have no detrimental effects on sample sizes. The same approach could also be applied to metrics from any imaging modality.…”

Multivariate classification of autism spectrum disorder using frequency-specific resting-state functional connectivity—A multi-center study

“…So multicenter datasets are needed. Recent data sharing projects such as the 1000 Functional Connectomes Project (Biswal et al, 2010), the International Neuroimaging Data-sharing Initiative (INDI) (Mennes et al, 2013) and the Autism Brain Imaging Data Exchange (ABIDE) (Di Martino et al, 2014) have revealed that resting-state fMRI datasets obtained from multi-sites are fruitfully aggregated for discovery and replication (Tomasi and Volkow, 2012) and is regarded that analysis using multi-center data can facilitate recruitment, increase study power, and overcome between-scanner variance to produced more generalizable results revealing common regions that contribute to classification consistently within each dataset (Biswal et al, 2010;Gradin et al, 2010;Moorhead et al, 2009;Suckling et al, 2010).…”

Brain morphology is individual-specific information

“…Many previous studies have evaluated the effects of using different scanners and/or scanner upgrade on morphometric results [11,[25][26][27][28][29][30][31][32][33][34][35][36][37][38]. Regarding volumetric measurements, there is generally greater inter-scanner variability than intra-scanner variability.…”