Early Childhood Behavioral Inhibition Predicts Cortical Thickness in Adulthood

Participants

Participants were derived from an ongoing longitudinal study of BI in which behavioral data were available beginning at 4 months of age.^3,30 Participants were originally recruited from the suburban Washington, DC area through commercially available mailing lists. Potential participants were screened over the telephone and excluded on the basis of prematurity, low birth weight, or perinatal complications. A home or lab visit was then scheduled for 433 infants (from 2 separate cohorts) within 2 weeks of their 4-month birthday, and infants were assessed for motor reactivity and emotional reaction to novel stimuli. Infants at the extreme ends of these measures were invited to participate in the full longitudinal study on the basis of selecting infants at high and low risk of developing behaviorally inhibited temperament.^3,31 A small number of participants (5 in the sample used in the current study) were added to the study in early childhood to serve as unfamiliar peers in a social lab assessment (4 or 7 years of age).

A total of 163 were potential participants for the current study on the basis of having participated in BI assessments during childhood as well as anxiety assessments during adolescence or young adulthood. Of these 163 participants, 25 refused participation in the young adult study, and an additional 105 were either excluded or not approached on the basis of the following: contraindication to magnetic resonance imaging (MRI), psychotropic medication use, psychopathology requiring immediate clinical attention, recent completion of other studies, and/or participant request not to be contacted for a period of time after which the current study was completed.¹² Acceptable neuroimaging data were therefore available for a subset of participants (n = 53, mean age 20.5 years, 29 male). Table 1 lists demographics for this subset and participation rates for the seven time points where data were acquired (ages 4 months; 14 months; 24 months; 4 years; 7 years; early adolescence, mean age 14.7 years; and young adulthood, mean age 19.8 years).

Children were observed for an assessment of BI at 14 and 24 months of age and for an assessment of social reticence at 4 and 7 years of age.^1,32,33 BI at 14 and 24 months was assessed by measuring infants’ reactions to novel objects and people.³ Mothers also reported their child’s social fear at 14 and 24 months using the Toddler Behavior Assessment Questionnaire.³⁴ Social reticence is considered a marker of BI during the early school-age years. For the assessments of social reticence at ages 4 and 7 years, children’s reticent behavior with three unfamiliar peers was measured with Rubin’s Play Observation Scale.³³ Mothers also rated their child’s social fear at 4 and 7 years using the shyness subscale of the Colorado Child Temperament Inventory.³⁵ For each of the four time points, observed behavioral and maternal report measures were standardized in the full cohort and then averaged to create a single measure of early childhood BI.^9,36,37 For our primary analyses, we used this continuous composite measure of early childhood BI and social reticence. This composite provides a stable measure of early childhood BI and has been used extensively in prior work.^11,12,36,37 For a secondary analysis to compare results with other studies, we categorized participants on the basis of their reactivity profiles at 4 months of age. This analysis attempts to replicate findings from the only other longitudinal structural brain study of young adults with histories of BI in infancy.²⁸ All 53 participants had the primary composite measure of BI available. For the secondary analysis, 43 had complete data for the 4-month reactivity phenotype available (of which 19 were classified as high negative reactive and 11 as low reactive).

In addition to examining the influence of early childhood BI, we wished to test whether anxiety and/or cognitive control during adolescence moderated the relation between early childhood BI and brain structure in young adulthood. This analysis minimized the number of statistical comparisons and derived a measure of anxiety incorporating as much data from the current study as possible. Accordingly, a single composite measure of anxiety over adolescence and young adulthood was created for each participant. This composite measure was created on the basis of three measures obtained during adolescence (mean age 14.7 years): parent report from Screen for Child Anxiety Related Disorders (SCARED), child report from SCARED,³⁸ and the Anxious/Depressed raw score on the Youth Self-Report (YSR);³⁹ and three measures obtained during young adulthood (mean age 19.8 years): Beck Anxiety Inventory,⁴⁰ Liebowitz Social Anxiety Scale,⁴¹ and the Anxious/Depressed raw score on the YSR. Standardized scores were created for each measure for each participant on the basis of all participants with available data on these anxiety measures (n = 163), not just the subset imaged for this current study. The use of standardized scores allows an equal weighting of each measure in the composite. The composite anxiety measure was an average over these six standardized measures (Cronbach’s alpha = 0.71). The adolescent and young adult composites were correlated (Spearman’s rho = 0.48, p<.001), suggesting trait-like features. Data were available in over 80% of the imaged participants for each individual scale, 92.5% of participants had at least one measure from adolescence, and 98.1% had at least one measure from adulthood. All participants had at least two of the six measures available.

The Eriksen Flanker task was used to measure cognitive control in the current sample. A subset of the cohort with imaging data in young adulthood had completed the Eriksen Flanker task in adolescence (n = 25, mean age 14.8 years, 9 female). The Flanker task consisted of equal numbers of congruent (HHHHH and SSSSS) and incongruent (HHSHH and SSHSS) trials in which participants had to indicate the central letter with a button press.⁴² Cognitive control was measured with the congruency effect, computed as the difference in reaction time for incongruent versus congruent trials, divided by the reaction time on congruent trials. Higher scores indicate poorer cognitive control.

Neuroimaging

All MRI data were acquired on the same scanner, a GE Healthcare MR750 3.0 Tesla scanner with a 32-channel head coil. Each scanning session included a single high-resolution, T1-weighted structural imaging sequence (MPRAGE; sagittal acquisition; 176 slices; 1 mm³ isotropic voxels; 256 × 256 matrix; flip angle = 7°; repetition time [TR] = 7.7 ms; echo time [TE] = 3.42 ms; inversion time [TI] = 425 ms).

Image Processing

FreeSurfer 5.3 (http://surfer.nmr.mgh.harvard.edu/) was used to generate subcortical segmentations,^43,44 cortical parcellations using the Destrieux et al. atlas,⁴⁵ and pial and white matter surfaces from the T1-weighted images.^46–48 After the parcellations and surfaces were generated, they were visually reviewed by trained research assistants and, where necessary, manually edited and regenerated. Regional volumes and cortical thicknesses were obtained from the edited parcellations and surfaces.^45,48

Regional a priori Analyses

Three bilateral regions of interest (ROIs) from the Destrieux et al. atlas⁴⁵ were selected a priori for initial cortical thickness analyses: middle anterior part of the cingulate gyrus and sulcus (dACC), short insular gyrus (aI), and the subcallosal gyrus (sgACC). These ROIs are depicted in Figure 1A. Bilateral volumes of the amygdalae and hippocampi were examined in a priori analyses as well. In an effort to replicate Schwartz et al.,²⁸ two additional ROIs were explored exclusively in the secondary analysis when examining the relations between four-month temperament and cortical thickness: left orbitofrontal and right ventromedial regions were hand-drawn on an average FreeSurfer volume based on Figures 1 and 2 from Schwartz et al.²⁸ These regions were then projected to individual participants to generate thickness values for the left orbitofrontal and right ventromedial regions for each participant.

Prefrontal Cortex Exploratory Analyses

We additionally performed exploratory analyses in the prefrontal cortex.²⁸ General linear modeling programs available through FreeSurfer (Qdec) were used for these vertex-wise analyses. A prefrontal cortex mask was generated by combining the following parcellations from the Desikan atlas:⁴⁹ superior frontal, rostral, and caudal middle frontal, pars opercularis, pars triangularis, pars orbitalis, lateral and medial orbitofrontal, frontal pole, and rostral and caudal anterior cingulate cortices. To correct for multiple comparisons, Monte Carlo simulations available in the FreeSurfer software were used to determine that an area of cortex 68.1 mm² with each individual vertex p<.001 was required to meet prefrontal cortex-wide cluster-wise significance of p<.05. Clusters meeting multiple comparison correction are reported in Montreal Neurological Institute (MNI) coordinates.

Statistical Analyses

All statistical tests were performed using SPSS version 20 (Armonk, NY) with the exception of the prefrontal cortex vertex-wise analyses, which were carried out using FreeSurfer. T-tests were used to compare BI and anxiety in the subset of participants that provided versus did not provide imaging data; chi-square tests were used to test for potential group differences in gender. Analogous tests compared participants in the current study that did versus did not provide Flanker task data in adolescence.

The relations among temperament, anxiety, and cortical thickness were examined with repeated measures analyses of variance (ANOVAs) predicting thickness. Specifically, for each a priori region, cortical thickness values from left and right hemispheres were included in a single model, with hemisphere treated as a repeated measure. Additional factors for the primary models comprised childhood BI, anxiety in adolescence/young adulthood, the interaction between BI and anxiety, gender, and whole-brain average cortical thickness. Variables were mean centered before computing interactions for all analyses. When significant effects emerged in primary models, secondary models added IQ and maternal education as covariates. Three participants each were missing IQ and maternal education data, and these missing values were replaced with group means; excluding these participants did not change the significance of results. Only results from secondary models are reported; primary models were included to ensure that results were not driven exclusively because of the addition of multiple covariates. For amygdala and hippocampus analyses, volumes were used instead of thicknesses, and whole-brain volume was used as a covariate instead of whole-brain average thickness. Bonferroni correction was used to protect against false positives for the five models based on the a priori regions.

Prefrontal cortex vertex-wise analyses were conducted separately for each hemisphere using the primary model with FreeSurfer. One participant’s composite BI measure was three standard deviations from the mean and another participant’s composite anxiety measure was three standard deviations from the mean. These outliers were handled by Winsorizing for all analyses; no outliers were detected in other measures.

Analyses examining the relations between cognitive control and cortical thickness of dACC were based on Westlye et al.⁵⁰ The congruency effect on the Eriksen Flanker task for each participant was computed as the difference in reaction time for incongruent versus congruent trials divided by the reaction time on congruent trials. Higher scores indicate poorer cognitive control, and the mean score across the sample was 0.10 with a standard deviation of 0.051. To attempt replication, the relation between congruency and cortical thickness was examined with a repeated measure ANOVA with hemisphere as a repeated measure and congruency, gender, and IQ as covariates. This model was followed up by additionally controlling for whole-hemisphere thickness. Finally, we included congruency, early childhood BI, and the interaction between BI and congruency in a single model to determine whether congruency moderated the relation between BI and cortical thickness.

Power analyses were computed using G*Power 3.1.⁵¹ Using a single predictor in a regression model to detect a moderate effect size (partial η²=0.15), power in the full sample of 53 was 0.84, power in the sample using just the 4-month phenotype (n=30) was 0.59, and power in the sample using the Flanker data (n=25) was 0.51. Power to detect a larger effect size (partial η²=0.25, on the order of the effect size of the main result in this study) was 0.98 in the full sample, 0.86 for the 4-month phenotype data, and 0.79 for the Flanker data.

Participants

Compared to participants from the larger longitudinal study that were not imaged, participants that were imaged for the current study had significantly lower anxiety scores over adolescence/young adulthood (−0.18, SD 0.57 vs. 0.16, SD 0.91, t=2.5, p=.012, d=0.45), but there were no differences in early childhood BI scores (−0.06, SD 0.73 vs. 0.012, SD 0.60, t=0.71, p=.48, d=0.11) or gender (χ² = 1.5, p=.22). Follow-up analyses revealed that imaged participants had significantly lower levels of anxiety compared to excluded participants (−0.18 vs. 0.17, t=2.6, p=.012). There was no difference in anxiety scores between the imaged participants and participants that refused participation (−0.18 vs. −0.07, t=0.7, p=.49). Note, however, that the absolute difference in anxiety between imaged participants and the total cohort was less than 0.2 SDs from the cohort mean, and the imaged participants were representative of the original longitudinal cohort in terms of BI scores. Within the current study sample, there were no significant differences in anxiety, BI, or gender (t=1.39, p=.17, d=0.38; t=0.24, p=.81, d=0.07; χ² = 0.23, p=.63, respectively) between the subset that participated versus the subset that did not participate in the Flanker task in adolescence. Early childhood BI and anxiety in adolescence/young adulthood were not significantly related in the sample used in this study (r= −0.043, p=.76).

Behavioral Inhibition, Anxiety, and Cortical Thickness: Regional Analysis

The relations among BI during early childhood, anxiety in adolescence/young adulthood, and cortical thickness in young adulthood were examined in three bilateral a priori cortical regions of interest: the dorsal anterior cingulate (dACC), anterior insula (aI), and subgenual anterior cingulate (sgACC). These regions are depicted in Figure 1A. Parallel analyses examined relations with hippocampus and amygdala volumes. As depicted in Figure 1B, early childhood BI predicted thinner cortex in the dACC with a large effect size (F[1,45] = 16.2, p<.001 uncorrected, partial η²=0.26, survived Bonferroni correction for five a priori regions). This analysis controlled for multiple potential confounding factors, including gender, maternal education, IQ, whole-brain average cortical thickness, anxiety in adolescence/young adulthood, and the interaction between BI and anxiety. Of note, similar results emerged in bivariate analyses examining only the relations between BI and dACC thickness. Early childhood BI similarly predicted thinner cortex in the sgACC (F[1,45] = 5.2, p=.027 uncorrected, partial η²=0.10), although with a smaller effect size. Unlike findings for the dACC, this result did not survive Bonferroni correction for five comparisons. Neither anxiety nor the interaction between BI and anxiety was significantly related to dACC or sgACC cortical thickness, suggesting that anxiety did not moderate the relation between early childhood BI and thickness of the dACC in adulthood. There were no significant relations among childhood BI, anxiety in adolescence/young adulthood, or the interaction between BI and anxiety with anterior insula thickness or hippocampus or amygdala volumes.

Behavioral Inhibition, Anxiety, and Cortical Thickness: Prefrontal Cortex Exploratory Analysis

To complement the a priori ROI analysis, a parallel analysis examined the relations among BI during early childhood, anxiety in adolescence/young adulthood, and cortical thickness in young adulthood in an exploratory fashion across prefrontal cortex. In contrast to the regional analysis, the prefrontal cortex vertex-wise analysis did not detect any regions surviving multiple comparison correction that were related to early childhood BI alone. A whole-brain vertex-wise map, uncorrected at p<.05, demonstrating the relation between early childhood BI and cortical thickness in young adulthood, is provided in Figure S1, available online. This map is consistent with the regional analysis, as the thickness of a large swath of cortex near the dACC was related to early childhood BI; as above, however, this cluster did not survive multiple comparison correction across prefrontal cortex. The prefrontal cortex vertex-wise analysis likewise did not detect any regions surviving multiple-comparison correction that were related to anxiety in adolescence/young adulthood.

The prefrontal cortex vertex-wise analysis did, however, detect a region in which thickness in young adulthood was related to the interaction between early childhood BI and anxiety in adolescence/young adulthood. As illustrated in Figure 2, cortical thickness in an 83 mm² patch of the right ventrolateral prefrontal cortex (VLPFC; centered at +48.2 +10.0 +13.9 in MNI coordinates, within the pars triangularis) was predicted by an interaction between childhood BI and anxiety in adolescence/young adulthood, controlling for gender and whole-hemisphere thickness (p=.024, corrected for multiple comparisons across prefrontal cortex). To explore the source of this interaction, a median split divided participants into those with low versus high levels of early childhood BI, and a similar median split was performed for anxiety. Note that effect sizes for this median split analysis are biased by the circular nature of the analysis, which computes statistics on a region initially identified in an exploratory test. Effect sizes are reported solely to interpret the interaction. In participants with low early childhood BI, a high level of anxiety during adolescence/young adulthood predicted thicker VLPFC compared to a low level of anxiety (d=1.56). In participants with high early childhood BI, however, anxiety during adolescence/young adulthood was unrelated to VLPFC thickness in young adulthood (d= −0.46).

Temperament at Four Months

The characterization of early childhood BI discussed above was a composite of measures from age 14 months to 7 years. Analyses in the only other study using a similar design, Schwartz et al.,²⁸ relied on data from 4 months of age. In secondary analyses, we used phenotypes similar to this prior study in the small subset (57%) of our participants with either high reactive negative (n = 19) or low reactive (n = 11) temperament at age 4 months. This measure did not predict cortical thickness in young adulthood in any of the a priori regions tested including the left orbitofrontal and right ventromedial regions derived from Schwartz et al.²⁸ Prefrontal cortex vertex-wise analyses similarly did not reveal any significant relations between 4-month temperament and regional cortical thickness. Using a less stringent threshold of p<.05 uncorrected also did not reveal any relations between cortical thickness near the regions derived from Schwartz et al.²⁸ and temperament at 4 months.

Cognitive Control

We examined whether congruency effects on the Eriksen Flanker task during adolescence moderated the relations between early childhood BI and dACC thickness in young adulthood. Higher congruency effects on the Eriksen Flanker task during adolescence (reflecting poorer cognitive control) predicted thinner dACC during young adulthood (F[1,23] = 4.66, p=.042, partial η²=0.17), controlling for gender and IQ. There was no relation, however, between BI in early childhood and the congruency effect in young adulthood among the current study sample (r=−0.01, p=.94). The interaction between BI and congruency likewise was unrelated to dACC thickness (F[1,21] = 2.12, p=.16, partial η²=0.09), indicating that the congruency effect (one measure of cognitive control) did not moderate the relation between early childhood BI and young adult dACC thickness.