Differential Effects of Common Variants in SCN2A on General Cognitive Ability, Brain Physiology, and messenger RNA Expression in Schizophrenia Cases and Control Individuals

Importance  One approach to understanding the genetic complexity of schizophrenia is to study associated behavioral and biological phenotypes that may be more directly linked to genetic variation.

Objective  To identify single-nucleotide polymorphisms associated with general cognitive ability (g) in people with schizophrenia and control individuals.

Design, Setting, and Participants  Genomewide association study, followed by analyses in unaffected siblings and independent schizophrenia samples, functional magnetic resonance imaging studies of brain physiology in vivo, and RNA sequencing in postmortem brain samples. The discovery cohort and unaffected siblings were participants in the National Institute of Mental Health Clinical Brain Disorders Branch schizophrenia genetics studies. Additional schizophrenia cohorts were from psychiatric treatment settings in the United States, Japan, and Germany. The discovery cohort comprised 339 with schizophrenia and 363 community control participants. Follow-up analyses studied 147 unaffected siblings of the schizophrenia cases and independent schizophrenia samples including a total of an additional 668 participants. Imaging analyses included 87 schizophrenia cases and 397 control individuals. Brain tissue samples were available for 64 cases and 61 control individuals.

Main Outcomes and Measures  We studied genomewide association with g, by group, in the discovery cohort. We used selected genotypes to test specific associations in unaffected siblings and independent schizophrenia samples. Imaging analyses focused on activation in the prefrontal cortex during working memory. Brain tissue studies yielded messenger RNA expression levels for RefSeq transcripts.

Results  The schizophrenia discovery cohort showed genomewide-significant association of g with polymorphisms in sodium channel gene SCN2A, accounting for 10.4% of g variance (rs10174400, P = 9.27 × 10⁻¹⁰). Control individuals showed a trend for g/genotype association with reversed allelic directionality. The genotype-by-group interaction was also genomewide significant (P = 1.75 × 10⁻⁹). Siblings showed a genotype association with g parallel to the schizophrenia group and the same interaction pattern. Parallel, but weaker, associations with cognition were found in independent schizophrenia samples. Imaging analyses showed a similar pattern of genotype associations by group and genotype-by-group interaction. Sequencing of RNA in brain revealed reduced expression in 2 of 3 SCN2A alternative transcripts in the patient group, with genotype-by-group interaction, that again paralleled the cognition effects.

Conclusions and Relevance  The findings implicate SCN2A and sodium channel biology in cognitive impairment in schizophrenia cases and unaffected relatives and may facilitate development of cognition-enhancing treatments.

Schizophrenia is a heritable neurodevelopmental disorder characterized by disturbed patterns of behavior and abnormalities of brain function.^1,2 Genomewide association studies (GWASs) are beginning to yield insights into the genetic architecture of schizophrenia, although effect sizes for individual genes are modest.^3-5 However, to our knowledge, few GWASs have examined behavioral or biological traits associated with the disorder, which may reflect more penetrant effects of common genetic variation.

Broad cognitive impairment is common in schizophrenia.^6-8 Subtle cognitive differences are often measurable years before psychotic symptoms or exposure to medications^9-13 and impairment is seen in an attenuated form in unaffected relatives,^6,7,14-16 suggesting that impaired cognition is an intermediate phenotype related to the genetic risk for schizophrenia.¹⁷ Studies in nonclinical groups,^18-20 and in patients with schizophrenia,^6,21,22 have indicated that cognitive data are characterized by a hierarchical structure, in which individual measures group into domain-specific cognitive factors (eg, working memory), which underlie a higher-order construct referred to as general cognitive ability or g. General cognitive ability is reliably indexed with standard measurement tools,²³ stable over time,^24,25 and associated with life outcomes from academic and vocational success^26-30 to health and mortality.^31,32 Physiologically, g is closely related to the efficiency of the prefrontal cortex (PFC),^33,34 an important focus of schizophrenia research.³⁵

The heritability of g has been estimated at between 40% and 80%,^25,36-38 but genetic associations with cognitive performance in nonclinical samples have been difficult to find and replicate^27,39 likely owing to the interaction of multiple genetic and environmental influences on brain development and function. Gene-cognition associations within clinical groups present additional complexities because of the potential role of illness epiphenomena (eg, medication) but may be enriched for illness-specific mechanisms of cognitive impairment (eg, apolipoprotein-E4 in Alzheimer disease samples). A fast-emerging, but inconsistent, literature has explored the association of cognitive performance with suspected genetic markers of schizophrenia.^40-46 One twin study suggested significant overlap in the genes that contribute to cognition and schizophrenia,⁴⁷ whereas another concluded that overlap was more limited.⁴⁸ Thus, it remains unclear to what degree the set of genes that leads to schizophrenia risk also impacts brain systems that underlie cognitive performance.

Here, we report a GWAS of cognition in Americans of European ancestry with DSM-IV schizophrenia and community control participants from the Clinical Brain Disorders Branch (CBDB)/National Institute of Mental Health (NIMH) Study of Schizophrenia Genetics (D.R.W., principal investigator). In the sodium channel gene, SCN2A (gene identification: 6326)—previously associated with seizure disorders, intellectual disability, and autism^49-53—we have identified single-nucleotide polymorphisms (SNPs rs10174400 and rs10182570) that show GWAS-significant association with general cognitive ability in schizophrenia. We found consistent evidence in a sample of the unaffected siblings of these probands and in independent schizophrenia samples. Further support comes from analyses of blood oxygen level–dependent (BOLD) functional magnetic resonance imaging (fMRI) during working memory and of RNA sequencing in postmortem PFC tissue samples.

The GWAS discovery sample included 363 community control individuals and 339 people with DSM-IV schizophrenia^54,55 after exclusions and genotyping quality control (QC) (Table 1). The main findings were tested further in a sample of full siblings of 147 of these probands (eTable 1 in Supplement provides details regarding inclusion and exclusion of participants). All research participants were competent adults and provided written informed consent pursuant to the National Institutes of Health neuroscience institutional review board–reviewed and –approved protocols.

Cognitive phenotypes were composites of individual measures constructed to represent verbal memory, visual memory, N-back, processing speed, card sorting, working memory span, and g (eTable 2 in Supplement). All composites were unweighted and were calculated in exactly the same way for probands, control participants, and unaffected siblings.⁶

DNA samples were genotyped using Illumina HumanHap550K/610Quad Bead Chips, according to the manufacturer’s protocol (eAppendix 1 in Supplement). After QC procedures (eAppendix 1 in Supplement), 495 089 high-quality autosomal SNPs were available for analysis. Quality control of individual genotyping results (eAppendix 1 in Supplement) left a total of 933 individuals with good genotype information. Of these, 339 probands and 363 control individuals had cognitive test data and were retained for discovery analyses (g could not be calculated for 5 probands because of missing data).

For siblings, SCN2A rs10174400 genotypes were determined using the 5′ exonuclease TaqMan assay. Single-nucleotide polymorphisms probe and primer sets were acquired from Applied Biosystems. Genotype accuracy was assessed by regenotyping within a subsample and reproducibility was routinely greater than 99%.

We performed multidimensional scaling on the matrix of genomewide identity by state pairwise distances using PLINK⁵⁶ version 1.07 and, to control for population stratification, included the first 4 multidimensional scaling axes as covariates in GWAS analyses. Analyses of the associations of 495 089 SNPs with 7 cognitive variables were performed in PLINK, assuming an additive genetic model and also controlling for age and sex. We did not control for education because it is confounded with illness and with g.⁵⁷ Analyses in unaffected siblings were conducted using PASW Statistics version 18.0 (IBM).

Study design details for the multisite CATIE (Clinical Antipsychotic Trials of Intervention Effectiveness) schizophrenia antipsychotic effectiveness trial have been published including details of cognitive assessments, genotyping, and genotype QC methods.^58-60 (Details related to the current comparison sample are in eAppendix 1 in Supplement.) Details of data collection for the Japanese sample have been previously published.⁶¹ The cognitive battery was comparable with the CBDB/NIMH battery. Genotyping and QC are described in eAppendix 1 in Supplement. Genetic and cognitive data were available for 95 people (eTables 1 and 2 in Supplement). The German sample consisted of 294 clinically stable individuals of European ancestry with DSM-IV schizophrenia, as described previously (details are in eAppendix 1 in Supplement).⁶²

Genotype-cognition association analyses in independent schizophrenia samples were conducted using PASW Statistics version 18.0. We performed unidirectional tests (ie, 1 tailed), assuming a minor allele disadvantage in schizophrenia, and using an additive genetic model, controlling for age and sex. For meta-analysis of effect sizes across schizophrenia samples, we calculated sample-weighted effect sizes with a bias correction for the small number of samples combined.

To test the relationship between SCN2A rs10174400 and cognition-related activation patterns as measured by BOLD fMRI, we studied 397 control participants and 87 schizophrenia cases from the CBDB/NIMH sibling study who were genotyped and completed the N-back working memory task while scanned at 3 T (details are in eAppendix 1 in Supplement). After quality screening and correction for covariates of no interest (eg, head motion), we used analyses of covariance controlling for age and sex to test SCN2A genotype within each diagnostic group and the interaction of diagnosis by genotype. Genotype groups within diagnoses did not differ in terms of demographic and performance variables. Thus, differences in activation are thought to reflect neural efficiency (ie, less activation at similar performance implying greater efficiency)—such differences representing a familial and heritable phenotype.^63-66 Given our interest in prefrontal information processing efficiency, we used a prefrontal region of interest with small-volume statistical correction (familywise error).

RNA sequencing was performed on postmortem PFC gray matter from 61 adult control participants (51 males; mean [SD] age, 44 [14.6] years) and 64 adult probands (51 males; mean [SD] age, 44.3 [14.8] years), all of European ancestry. Detailed brain-tissue collection methods used by the Lieber Institute and CBDB/NIMH have been published⁶⁷ and details of RNA sequencing are described in eAppendix 1 in Supplement. The relative abundances of the 3 common SCN2A RefSeq transcripts—NM_21007, NM_001040142, and NM_001040143—were estimated by Cufflinks version 2.0.2 and compared with Illumina iGenome gene annotation. The 3 transcripts can be differentiated based on unique 5′ exons, thus allowing a reliable estimation of relative abundance of each specific transcript. We used analyses of covariance, with age and sex covariates, to investigate main effects and interactions among diagnosis, SCN2A rs10174400 genotype, and SCN2A transcript levels for the 3 transcripts. Analyses were also corrected for postmortem interval and RNA integrity number. We calculated Cohen d effect sizes. With the low number of rs10174400 minor allele homozygotes (8 probands and 7 control participants), we combined heterozygotes with minor allele homozygotes (T carriers) for analyses based on genotype.

In Supplement, eAppendix 1 describes covariate sensitivity analyses (ie, medication, chronicity, age at onset, and family socioeconomic status); analysis of the potential role in current findings of low-frequency exonic SNPs; and tests of the association of g with SNP sets representing the whole SCN2A gene, other sodium channel genes, and the whole sodium and calcium channel gene families.

The GWAS in the schizophrenia sample identified a strong association signal (Figure 1A). Two linked, intronic SNPs in SCN2A surpassed GWAS significance (ie, P = 5.0 × 10⁻⁸) for association with g (rs10174400, P = 9.27 × 10⁻¹⁰; rs10182570, P = 2.56 × 10⁻⁹; Table 2; eFigure 1 in Supplement)—accounting for 10.4% of g variance—with no evidence of inflation of test statistics due to population effects (λ_{genomic control} = 1; Figure 1B, eAppendix 2 in Supplement, and eTables 3-6 in Supplement). Performance was least impaired in participants homozygous for the major C allele, intermediate in heterozygotes, and most impaired in participants homozygous for the T allele (Figure 2). In nonindependent analyses, the SCN2A rs10174400 genotype was also associated with the 6 cognitive domain variables in schizophrenia (Table 2). Each of the domains showed directionally consistent and, at least nominally significant, association with rs10174400 genotype, but none met the GWAS threshold.

For control individuals, no SNP association approached GWAS significance (eAppendix 2, eFigure 2, and eTable 7 in Supplement) and the SCN2A rs10174400 genotype was not a predictor of case/control status (Table 2). Unexpectedly, the allelic trend for the control association with g was in the direction opposite the schizophrenia association (Figure 2), and an analysis of the interaction of rs10174400 genotype by group was also GWAS significant (P = 1.75 × 10⁻⁹, Table 2).

Although not independent of proband results, the sibling analyses addressed the concern that the proband association might be primarily related to illness characteristics (eg, ongoing symptoms) or medications. In unaffected siblings, there was a robust, directionally parallel association between rs10174400 genotype and g, accounting for 3.4% of performance variation, and a significant genotype-by-group interaction (Table 2).

In healthy populations, g has been shown to predict educational attainment,^18,27 so a genotype that predicts g might be associated with education. In 147 unaffected siblings, the rs10174400 genotype accounted for 5.7% of the variance in years of education completed (P = .003), with T-allele carriers showing clearly reduced educational attainment compared with C-allele homozygotes (eFigure 3 in Supplement). This association was not present in the full schizophrenia sample (P = .38), likely because of the confounding effect of illness on educational attainment.⁵⁷

In 279 schizophrenia cases from the CATIE trial, regression analysis confirmed the association of an rs10174400 proxy to the CATIE neurocognitive composite,⁶⁸ a general cognitive ability index similar to g, again showing directionality parallel to the discovery analyses (Table 2). Genotype associations to subsidiary composites for processing speed and working memory were also significant and parallel. In 95 Japanese schizophrenia cases, regression analysis yielded a directionally consistent significant association of the same proxy SNP with g, accounting for 3.4% of the variance in performance (Table 2). Post hoc analysis using a recessive model showed an even more pronounced effect, and we observed a similar pattern for a verbal memory composite. Finally, we examined gene/cognition associations in 295 Germans with schizophrenia. Regression analyses failed to replicate the association of rs10174400 with g in schizophrenia in this sample (Table 2). However, there was a parallel genotype association with the working memory span composite in the German cases, which was the strongest domain-specific effect in the discovery sample. Together, the 3 replication samples included 652 people with schizophrenia and g. Across the 3 groups, the rs10174400 genotype accounted for 1.0% of the variance in g (sample-weighted mean effect size). Including the discovery sample with the replication samples (N = 983), genotype accounted for 3.0% of g variance in schizophrenia on average.

Looking beyond performance, we tested for genotype effects at the level of brain physiology using an N-back working-memory paradigm that robustly engages prefrontal cortical circuitry. The rs10174400 genotype was differentially associated with PFC efficiency in cases and control individuals, analogously to the cognitive results pattern. Among control individuals, TT homozygotes were most efficient; among schizophrenia cases, they were least efficient, and the interaction effect was significant (Montreal Neurological Institute coordinates: x = −36, y = 27, z = 33; familywise error–corrected P = .02; eFigure 4 in Supplement). There were also main effects of SCN2A genotype in both diagnostic groups consistent with the direction of this interaction and with the cognitive associations (eAppendix 2 in Supplement).

Analysis of RNA sequencing data from postmortem PFC gray matter tissue samples showed significantly reduced expression of SCN2A mRNA in the schizophrenia sample relative to control individuals for 2 of 3 RefSeq transcripts and significant genotype effects and interactions for these 2 transcripts (Table 3 and eFigure 5 in Supplement). The effect sizes for significant findings were small to medium in magnitude. The directions of genotype effects were opposite for the 2 groups and the diagnosis-by-genotype interactions were significant—patterns remarkably similar to those in the cognitive and imaging data.

Our main findings showed little change in analyses with additional covariates (ie, medication, age at prodrome onset, chronicity, positive and negative symptoms, or family socioeconomic status). Analysis of low-frequency exonic SNPs was inconclusive. Tests of the association of g with SNP sets were an initial step in determining whether the association of sodium channel biology with general cognitive performance extended beyond the influence of the 2 GWAS-significant SNPs (eAppendix 2 and eTables 8-12 in Supplement).

In our GWAS analyses of general cognitive ability in patients with schizophrenia, 2 linkage disequilibrium–linked SNPs in SCN2A showed GWAS-significant association. The effect accounted for 10.4% of the variance in overall cognitive performance. A parallel association of the rs10174400 genotype with g in 147 unaffected siblings indicated that the schizophrenia association cannot be attributed solely to illness epiphenomena (eg, medication). Notably, in the siblings, educational attainment also varied with the rs10174400 genotype, accounting for 5.7% of sibling education variance. We found evidence for weaker, but parallel, genotype/cognition associations in independent schizophrenia samples. Across these 3 replication samples, totaling 652 probands, genotype accounted for 1.0% of g variance. Control participants showed a trend for genotype association with allelic directionality opposite to the schizophrenia association, and the rs10174400 genotype-by-group interaction was also GWAS significant.

Neuroimaging findings and RNA sequencing data from postmortem PFC samples provided a measure of biological validation for the behavioral association findings. Analyses of prefrontal information processing efficiency during working memory revealed a genotype-by-diagnosis interaction. The rs10174400 minor (T) allele conferred efficiency advantages for control individuals but maximal inefficiency in schizophrenia. In postmortem RNA sequencing experiments, the schizophrenia sample showed reduced expression of mRNA for 2 of 3 common alternative transcripts and genotype-by-diagnosis interactions analogous to the imaging results. Thus, the pattern of differential rs10174400 genotype associations for cases vs control individuals that was hinted at in the behavioral data (ie, a clear allele dose-dependent effect on cognitive performance in schizophrenia and a weak opposite trend in control individuals) came more clearly into focus in biological analyses. In sum, congruent evidence spanning behavior, physiology, and mRNA expression suggests an interaction between SCN2A genetic markers and schizophrenia-associated phenomena.

Our discovery sample effect was dramatic and likely reflects the winner’s curse seen in some other genetic association studies of relatively small samples. Evidence from independent schizophrenia samples suggested that the SCN2A effect on cognition may generally be smaller—on average, genotype accounted for 1.0% of variance in our replication samples, although in 2 of these 3 samples, the effect was in the range of 1.5% to 3.4%. While smaller, these effects in independent samples were directionally consistent with discovery sample effects—notwithstanding considerable differences in ascertainment, genotyping, and phenotyping. Additionally, the magnitude of the main schizophrenia finding may have reflected enrichment of the CBDB/NIMH sample for a particular form of schizophrenia risk–associated cognitive impairment owing to uniform, restrictive inclusion criteria (eg, IQ>70 and no substance abuse). Across the discovery and replication samples (N = 986), the mean sample-weighted association effect size was 3.0% of g variance. Neuroimaging findings, and mRNA expression findings in wholly independent postmortem brain-tissue samples, offered further, directionally consistent support for the main finding. At the same time, the parallel findings in siblings, although nonindependent, suggested that the schizophrenia findings were not determined by illness epiphenomena. Altogether, the data alleviate concerns that these are not true genotype effects on SCN2A biology. A better understanding of the magnitude of these effects will require further analyses in other samples.

The findings are also plausible both biologically and in terms of known clinical associations. SCN2A encodes the α2 subunit of a voltage-gated sodium ion channel that is widely expressed in the brain and contributes to the initiation and propagation of action potentials.^69,70 Na(v)1.2 (the protein encoded by SCN2A) is abundant in parvalbumin-positive gamma-aminobutyric acid–related inhibitory interneurons, at least in the hippocampus and temporal lobe.⁷⁰ Gamma-aminobutyric acid system abnormalities have been a particular focus of cognitive impairment research in schizophrenia.^71,72 Multiple mutations in SCN2A have been associated with childhood epilepsies, sometimes combined with intellectual disability and/or autismlike symptoms,^69,73 and antiepileptic medications that block sodium channels (eg, topiramate) have adverse cognitive effects.⁷⁴ Notably, each of 3 recent whole-exome sequencing studies focused on nonsyndromic intellectual disability found de novo coding mutations in SCN2A (3 of 55 sequenced individuals in one study,⁵² 1 of 12 in the second,⁵¹ and 1 of 100 in the third⁴⁹). Results from a large exome-sequencing study of autism recently identified 279 independent de novo mutations and highlighted SCN2A as the single gene disrupted by 2 of these.⁵³

The hypothesis that cognition is an intermediate phenotype for schizophrenia implies that rs10174400 should discriminate cases from control individuals, at least to some degree.¹⁷ No case/control signal was observed in the discovery sample. Therefore, our results suggest that a strong, directionally specific SCN2A association with impaired cognition may emerge in the context of the complex genetic risk architecture of schizophrenia, which is shared by patients and family members, although there is little or no association of SCN2A with cognitive performance in the general adult population. We have very limited evidence as to possible mechanisms, but the involvement of sodium channel biology—and its apparent effect at the most general level of cognitive performance—suggests mediation through low-level and widely acting neural systems. Dysfunction in gamma-aminobutyric acid–related inhibitory systems could fit this description, although there are many possibilities. The findings in unaffected siblings may be quite important in further refining hypotheses about mechanisms. The sibling results clarified that the genotype association to cognition was not driven mainly by illness-specific phenomena. The association was not unique to family members with a schizophrenia diagnosis and was not tightly linked either to positive or negative symptoms, illness chronicity, or antipsychotic medication (eTable 8 in Supplement). Although impaired cognition and psychotic symptoms are defining characteristics of the schizophrenia syndrome, the sibling results reported here frame the question whether these characteristics may be related to distinct genetic components. At the same time, the current study was dramatically smaller than case/control samples that have shown high P value SNP associations with diagnosis. It may be that, with sufficient samples sizes, associations of SCN2A SNPs with the schizophrenia diagnosis will emerge. In the latest published Psychiatric Genomics Consortium analysis of more than 60 000 participants (>21 000 cases),⁵ several SNPs in SCN2A showed association with schizophrenia at P = 5.0 × 10⁻³.

Despite ample evidence of heritability for widely used cognitive measures,²⁴ in control individuals, no common variant reached genomewide significance or approached the magnitude of the rs10174400 effect in schizophrenia. Our results echo findings in earlier, larger cognition GWASs.^75,76 Perhaps especially for traits as conserved and fundamental as nondisordered cognition, the causal effects of individual, common genetic markers cannot be detected at present amid the complex interaction of genetic, environmental, and random influences that affect individuals over decades of development.⁷⁷

We have identified common variants in SCN2A that, in the context of schizophrenia and risk for schizophrenia, show substantial and consistent associations with broad cognitive performance, brain physiology, and mRNA expression in the brain. These findings intersect with prominent lines of schizophrenia research and suggest testable hypotheses about the biological roots of cognitive impairment in schizophrenia and avenues for new treatment development.

Corresponding Author: Daniel R. Weinberger, MD, Lieber Institute for Brain Development, and Departments of Psychiatry, Neurology, and Neuroscience, and Institute of Genetic Medicine, Johns Hopkins University School of Medicine, 855 N Wolfe St, Baltimore, MD 21205 (drweinberger@libd.org).

Submitted for Publication: June 6, 2013; final revision received December 23, 2013; accepted January 2, 2014.

Published Online: April 9, 2014. doi:10.1001/jamapsychiatry.2014.157.

Author Contributions: Drs Dickinson and Weinberger had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.

Study concept and design: Straub, Trampush, Gao, Takeda, Hyde, Berman, Weinberger.

Acquisition, analysis, or interpretation of data: Dickinson, Straub, Trampush, Gao, Feng, Xie, Shin, Lim, Ursini, Bigos, Kolachana, Hashimoto, Baum, Rujescu, Callicott, Hyde, Berman, Kleinman, Weinberger.

Drafting of the manuscript: Dickinson, Straub, Trampush, Gao, Feng, Ursini, Kolachana, Hashimoto, Baum, Hyde, Weinberger.

Critical revision of the manuscript for important intellectual content: Dickinson, Straub, Trampush, Feng, Xie, Shin, Lim, Ursini, Bigos, Hashimoto, Takeda, Rujescu, Callicott, Hyde, Berman, Kleinman, Weinberger.

Statistical analysis: Dickinson, Straub, Trampush, Gao, Feng, Shin, Lim, Ursini, Bigos, Hashimoto, Baum, Callicott.

Obtained funding: Takeda, Rujescu, Berman, Weinberger.

Administrative, technical, or material support: Dickinson, Trampush, Gao, Bigos, Kolachana, Hashimoto, Takeda, Rujescu, Hyde, Berman, Kleinman, Weinberger.

Study supervision: Dickinson, Straub, Trampush, Gao, Hyde, Berman, Weinberger.

Conflict of Interest Disclosures: Dr Dickinson works for the National Institutes of Health. Drs Hyde, Kleinman, and Weinberger were previously employed by the National Institutes of Health. Dr Weinberger directs the Lieber Institute for Brain Development. No other disclosures were reported.

Funding/Support: This work was supported by the Division of Intramural Research Programs, National Institute of Mental Health, National Institutes of Health, Bethesda, Maryland, as direct funding of the Clinical Brain Disorders Branch (Dr Weinberger, principal investigator) and the Lieber Institute for Brain Development, Baltimore, Maryland.

Role of the Sponsor: The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

Correction: This article was corrected online May 8, 2014, for an error in the Results section.