Characterizing individual differences in functional connectivity using dual-regression and seed-based approaches

Highlights

• Sex differences are expressed in connectivity patterns with multiple networks.

• Seed-based analysis (SBA) does not accurately represent connectivity with networks.

• Dual-regression analysis (DRA) accurately represents connectivity with networks.

• Individual differences in functional connectivity are characterized better with DRA.

Abstract

A central challenge for neuroscience lies in relating inter-individual variability to the functional properties of specific brain regions. Yet, considerable variability exists in the connectivity patterns between different brain areas, potentially producing reliable group differences. Using sex differences as a motivating example, we examined two separate resting-state datasets comprising a total of 188 human participants. Both datasets were decomposed into resting-state networks (RSNs) using a probabilistic spatial independent component analysis (ICA). We estimated voxel-wise functional connectivity with these networks using a dual-regression analysis, which characterizes the participant-level spatiotemporal dynamics of each network while controlling for (via multiple regression) the influence of other networks and sources of variability. We found that males and females exhibit distinct patterns of connectivity with multiple RSNs, including both visual and auditory networks and the right frontal–parietal network. These results replicated across both datasets and were not explained by differences in head motion, data quality, brain volume, cortisol levels, or testosterone levels. Importantly, we also demonstrate that dual-regression functional connectivity is better at detecting inter-individual variability than traditional seed-based functional connectivity approaches. Our findings characterize robust—yet frequently ignored—neural differences between males and females, pointing to the necessity of controlling for sex in neuroscience studies of individual differences. Moreover, our results highlight the importance of employing network-based models to study variability in functional connectivity.

Introduction

Individuals are remarkably diverse, exhibiting variation across a host of behaviors and phenotypes. Psychologists have long recognized the importance of including individual variability in cognitive models (Underwood, 1975), and neuroscientists have begun to identify underlying structural and functional variability in specific brain regions (Braver et al., 2010, Hariri, 2009) and how that variability relates to individual differences in a range of domains: motivation (Clithero et al., 2011, Mobbs et al., 2009, Strauman et al., 2013), reward sensitivity (Beaver et al., 2006, Carter et al., 2009), trait anxiety (Bishop, 2009, Etkin et al., 2004), and working memory capacity (Osaka et al., 2003, Todd and Marois, 2005).

Yet, many computations are distributed across networks of regions rather than being restricted to a specific region (Friston, 2009). Accordingly, studies of functional connectivity of the brain at rest have converged on the idea that the brain is organized into multiple, overlapping resting-state networks (RSNs) (Beckmann et al., 2005, Smith et al., 2009). Some of these networks, including the default-mode network (Buckner et al., 2008, Raichle et al., 2001), are observed in multiple species (Hayden et al., 2009, Lu et al., 2012, Vincent et al., 2007), which highlights the fundamental nature of their role in neural organization. Although RSNs represent a primary target of recent work on individual differences, even relatively straightforward questions regarding sex differences have led to equivocal results (Biswal et al., 2010, Filippi et al., 2012, Wang et al., 2012a, Weissman-Fogel et al., 2010). The lack of consensus across these studies could be due to a number of factors, including small sample sizes (Yarkoni, 2009) and the inability of traditional analysis approaches to accurately represent the distributed computations that occur across RSNs (Cole et al., 2010).

Characterizing the neural bases of sex differences could provide a crucial first step toward understanding the mechanisms of psychopathologies that are linked to sex (Rutter et al., 2003). We therefore investigated whether sex differences are expressed in patterns of functional connectivity during the resting state. We recruited a large sample of participants (N = 188), which we partitioned into split samples for an internal replication. For each dataset, we computed a spatial independent component analysis (ICA) that parceled the functional data into a set of independent spatial maps (Fig. 1), some reflecting artifactual spatial structures and others reflecting well-characterized RSNs (Smith et al., 2009). We then employed a dual-regression functional connectivity analysis, which quantifies connectivity with an entire RSN—rather than a representative node of the RSN, a limitation of traditional seed-based approaches (Cole et al., 2010)—while controlling for the influence of other RSNs (Filippini et al., 2009, Leech et al., 2011, Leech et al., 2012). Our analyses revealed two key results. First, functional connectivity patterns between distinct brain regions and multiple RSNs reliably predicted sex differences. Second, functional connectivity estimates derived from dual-regression analysis were better at classifying males and females than similar estimates obtained from a seed-based analysis, suggesting that dual-regression analysis provides a superior representation of the distributed computations that occur within RSNs.