Minimal Relationship between Local Gyrification and General Cognitive Ability in Humans. Samuel R Mathias et al. Cerebral Cortex, bhz319, February 9 2020. https://doi.org/10.1093/cercor/bhz319
Abstract: Previous studies suggest that gyrification is associated with superior cognitive abilities in humans, but the strength of this relationship remains unclear. Here, in two samples of related individuals (total N = 2882), we calculated an index of local gyrification (LGI) at thousands of cortical surface points using structural brain images and an index of general cognitive ability (g) using performance on cognitive tests. Replicating previous studies, we found that phenotypic and genetic LGI–g correlations were positive and statistically significant in many cortical regions. However, all LGI–g correlations in both samples were extremely weak, regardless of whether they were significant or nonsignificant. For example, the median phenotypic LGI–g correlation was 0.05 in one sample and 0.10 in the other. These correlations were even weaker after adjusting for confounding neuroanatomical variables (intracranial volume and local cortical surface area). Furthermore, when all LGIs were considered together, at least 89% of the phenotypic variance of g remained unaccounted for. We conclude that the association between LGI and g is too weak to have profound implications for our understanding of the neurobiology of intelligence. This study highlights potential issues when focusing heavily on statistical significance rather than effect sizes in large-scale observational neuroimaging studies.
Discussion
In the present study, we analyzed data from two samples of related
individuals to examine the association between gyrification and general
cognitive ability. We used a popular automatic method to calculate LGI
across the cortex from MRI images (
Schaer et al. 2008), and calculated
g from performance on batteries of cognitive tests. We estimated the heritability of height, ICV, and
g, as well as the heritability LGI, area, and thickness at all vertices. We estimated phenotypic, genetic, and environmental LGI–
g correlations, as well as partial LGI–
g
correlations with height, ICV, area (at the same vertex), and thickness
(at the same vertex) as potential confounding variables. We estimated
the amount of phenotypic variance of
g explained by all LGIs
together via ridge regression, and examined the across-sample
consistency of neuroanatomical specificity in heritability of LGI, area,
and thickness, as well as LGI
–g correlations. Finally, we tested whether heritability estimates and LGI–
g correlations were stronger in regions implicated by the P-FIT, a model of the neurological basis of human intelligence (
Jung and Haier 2007).
A
novel finding of the present study was that LGI was heritable across
the cortex, extending a previous study that established the heritability
of whole-brain GI (
Docherty et al. 2015).
This finding was not particularly surprising because many features of
brain morphology are heritable. Nevertheless, it was necessary to
establish the heritability of LGI before calculating genetic LGI–
g correlations, which are only meaningful if both LGI and
g
are heritable traits. The previous study estimated the heritability of
GI to be 0.71, which is much greater than most of the heritability
estimates for LGI observed in GOBS or HCP. This result is also not
surprising, because GI is likely to be contaminated by less measurement
error than LGI. Heritabilities of all other traits were consistent with
those published in previous studies.
The present study represents a
replication of previous work and provides several important extensions
to our understanding of the relationship between gyrification and
cognition. First, we replicated previous work by finding positive and
significant phenotypic LGI–
g correlations (e.g.,
Gregory et al. 2016). Furthermore, we found that genetic LGI–
g
correlations were positive and significant (but only in HCP),
suggesting that the relationship between gyrification and intelligence
may be driven by pleiotropy. Since environmental LGI–
g
correlations were not significant, their net sign differed across GOBS
and HCP, and their spatial patterns showed no consistency across
samples, it is reasonable to conclude that they mostly reflected
measurement error rather than meaningful shared environmental
contributions to LGI and
g.
In our view, the most important finding from the present study is that all LGI–
g correlations, even the significant ones, were weak. Phenotypically, LGI at a typical vertex poorly predicted
g. Even when the predictive ability of all LGIs was considered together via ridge regression, at least 89% of the variance of
g remained unaccounted for. Phenotypic and genetic LGI–
g correlations were weaker than ICV–
g correlations in the same participants, and about the same as area–
g correlations. Partialing out ICV or area further reduced LGI–
g correlations.
The
volume of cortical mantle is often computed as the product of its area
and thickness, but at the resolution of meshes typically used to
represent the cortex, the variability of area is higher than the
variability of thickness such that surface area is the primary
contributor to the variability of cortical volume (
Winkler et al. 2010),
and therefore of its relationship to other measurements; the same
holds, more strongly even, for parcellations of the cortex in large
anatomical or functional regions. This means that the association
between overall brain volume and cognitive abilities reported by
previous studies (e.g.,
Pietschnig et al. 2015) is probably primarily driven by area–
g correlations (
Vuoksimaa et al. 2015). LGI is strongly correlated with area (
Gautam et al. 2015;
Hogstrom et al. 2013), which explains why partialing out either ICV or area reduced phenotypic and genetic LGI–
g
correlations in the present study. Thus, we conclude, based on our
results, that the association between gyrification and cognitive
abilities to a large extent reflects the already well-established
relationship between surface area and cognitive abilities, and that the
particular association between the unique portion of gyrification and
cognitive abilities is extremely small.
The above conclusion is consistent with that of a previous twin study (
Docherty et al. 2015),
which examined genetic associations between overall cortical surface
area, whole-brain GI, and cognitive abilities. The authors concluded
that the genetic GI–
g correlation could be more or less fully explained by the area–
g
correlation. It has been argued previously that focusing on whole-brain
GI may miss important neuroanatomical specificity; however, our
findings suggest that Docherty et al.’s conclusion holds for both local
and global gyrification.
The P-FIT is a popular hypothesis concerning which brain regions matter most for human cognition (
Jung and Haier 2007).
The P-FIT was initially proposed to explain activation patterns
observed during functional MRI experiments, but has been extended to
aspects of brain structure. Previous studies have suggested that the
association between gyrification and cognitive abilities may be stronger
in P-FIT regions than the rest of the brain (
Green et al. 2018;
Gregory et al. 2016).
However, when we tested this hypothesis, we actually found evidence to
the contrary. Since neuroanatomical patterns of phenotypic and genetic
LGI–
g correlations were consistent across GOBS and HCP, this
unexpected finding was unlikely to have been caused by a lack of
specificity, such as if LGI–
g correlations were distributed randomly over the cortex. Instead, while LGI–
g
correlations exhibited a characteristic neuroanatomical pattern, this
pattern did not match the P-FIT. A potential limitation of the present
study in this regard is that there is no widely accepted method of
matching Brodmann areas (used to define P-FIT regions) to surface-based
ROIs (used to group vertices). Therefore, one could argue that our
selection of P-FIT regions was incorrect. While our selection was based
on that of a previous study (
Green et al. 2018),
we nevertheless reperformed our analysis several times with different
selections of P-FIT regions, and the results remained the same.
Importantly, although we argue that the P-FIT is not a good model for
the association between gyrification—a purely structural aspect of
cortical organization—and cognitive abilities, our results should not be
used to criticize the P-FIT as a hypothesis of the brain’s functional
organization, because function does not necessarily follow structure.
Most
of our results were consistent across samples. However, estimates of
heritability and genetic correlations were generally weaker in GOBS than
HCP. Notably, some genetic LGI–
g correlations were strong
enough to surpass the FDR-corrected threshold for significance in HCP,
but not GOBS. Such differences could be related to study design. One
limitation of all family studies is that polygenic effects are
susceptible to inflation due to shared environmental factors, which
would cause overestimation of both heritability and genetic
correlations. It could be argued that extended-pedigree studies, such as
GOBS, are less susceptible to this kind of inflation than twin studies,
such as HCP, because there are usually fewer shared environmental
factors between distantly related individuals than twins (
Almasy and Blangero 2010);
this reduction in inflation comes at the expense of a reduction in
power to detect polygenic effects, which could also explain the lack of
significant genetic correlations in GOBS. It is unlikely that
differences in results between samples were caused by differences in
scanner or scanning protocol (
Han et al. 2006).
Furthermore, while GOBS and HCP participants completed different
cognitive batteries, both were comprehensive in terms of measured
cognitive abilities, ensuring that
g indexed a similar construct in both samples.
With
the recent emergence of large, open-access data sets and international
consortia, neuroimaging and genetics studies have entered a new era
characterized by samples comprising many thousands of participants. In
such large studies, trivial effects may be labeled as statistically
significant. This observation is not new (
Berkson 1938) and numerous solutions have been proposed, such as adopting more stringent significance criteria (
Benjamin et al. 2018), scaling criteria by sample size (
Mudge et al. 2012), testing interval-null rather than point-null hypotheses (
Morey and Rouder 2011), and, most radically, abandoning the notion of statistical significance altogether (
McShane et al. 2019).
One could argue that these solutions suffer from their own drawbacks
and are unlikely to be adopted by the scientific mainstream in near
future. Therefore, in the meantime, we believe that it is imperative to
judge, at least qualitatively, whether the sizes of statistically
significant effects are large enough to justify one’s conclusions,
particularly when these conclusions may have broad, overarching
implications. This idea is not new either (
Kelley and Preacher 2012)
but deserves to be restated. Based on the results of the present study,
we are inclined to believe that gyrification minimally explains
variation in cognitive abilities and therefore has somewhat limited
implications for our understanding of the neurobiology of human
intelligence.