At baseline, based on self-report from the Customary Drinking and Drug Use Record (CCDR) (Brown et al., 1998), 31 participants in the analysis sample reported alcohol use that exceeded age-adjusted National Institute on Alcohol Abuse and Alcoholism (NIAAA) guidelines for risky drinking (see Brown et al., 2015), and are considered “exceeds threshold drinkers” (ETD).

As with other sites, the NCANDA Pittsburgh sample was recruited such that approximately half of the participants were at increased risk for problematic alcohol use. The current sample utilizes participants with increased substance use risk based on one or more of the following: early substance use onset (EOS; n = 9), family history of substance use disorder (FH; n = 21), externalizing symptoms (EXT, n = 31), or internalizing symptoms (INT, n = 26). In the current analyses, EOS, FH, EXT, and INT were categorical variables (Risk vs. No Risk) and were defined at the baseline visit as in Brown et al. (2015).

EOS was defined as consuming the first full drink of alcohol prior to age 15, based on the child's report on the CDDR (Brown et al., 1998). FH was defined as having at least one biological parent with a history suggesting a substance use disorder (SUD), based on parent report using the Family History Assessment Module (Rice et al., 1995). EXT was defined as having endorsed at least one symptom of conduct disorder or antisocial personality disorder on the Computerized Semi-Structured Assessment for the Genetics of Alcoholism (SSAGA, Bucholz et al., 1994; Hesselbrock et al., 1999) or having an Achenbach System of Empirically Based Assessment (ASEBA; Achenbach, 2009) externalizing age- and gender-adjusted t-score above 60. Similarly, INT was defined as having endorsed two or more symptoms or having an ASEBA internalizing t-score above 60. In the current analysis, we did not examine EOS because of the few participants meeting criteria in the Pittsburgh sample (n = 9).

In addition to risk factors described by Brown et al. (2015), we also examined associations between AS performance and BOLD activation and trait impulsivity as measured by the short version of the UPPS-P (SUPPS-P, Cyders et al., 2014). Trait impulsivity is associated with increased vulnerability to SUD (c.f., Verdejo-García et al., 2008) and has been conceptualized as having shared underlying features with antisocial behavior and substance use (Krueger et al., 2007). Although trait impulsivity and response inhibition have both been considered measures of inhibitory control, it has been suggested trait personality measures of impulsivity likely represent more global differences in impulsive choice whereas behavioral measures of response inhibition reflect specific cognitive processes (Reynolds et al., 2006). We focused on the SUPPS-P scales of negative and positive urgency as the urgency domain has been associated with substance use (Zapolski et al., 2009) and BOLD activation in a rewarded stop signal task (Wilbertz et al., 2014). Participants completed the SUPPS-P at both baseline and follow up visits.

As part of the NCANDA test protocol, participants completed the Penn Computerized Neurocognitive Battery (WebCNP: Gur et al., 2010) and traditional “pencil and paper” neuropsychological tests (Sullivan et al., 2016). We utilized a composite score, generalized ability accuracy (GA; see Sullivan et al., 2016), as an outcome variable and covariate when examining substance use risk factors in our analyses. The GA measure was derived from several domains of neuropsychological function, allowing us to examine whether associations among substance use risk factors, AS performance, and BOLD activation were specific to response inhibition or instead, reflected reduced generalized cognitive ability.

The socioeconomic status (SES) variable was determined by both parental education and income (see Brown et al., 2015). SES was expressed as a standard score (mean = 100, SD = 15) and included as a covariate in secondary analyses.

In order to harmonize our analytic approach with previous work with NCDANDA baseline data (c.f., Brown et al., 2015; Müller-Oehring et al., 2017), at both baseline and follow up we utilized scores from study initiation for the primary substance use risk categories (FH, EXT, INT) and exceeds threshold drinking (ETD) as outlined in Brown et al. (2015). This approach allows our results to be interpreted more readily with other projects describing cognitive differences with baseline risk group definitions (c.f., Sullivan et al., 2016). Continuous primary measures (participant age and trait-impulsivity) were treated as time-varying covariates, with the unique scores entered at baseline and follow up.

Based on the random effects structure of our modeling framework (see below), we were able to utilize participants that only had data at one visit (baseline or follow up). However, some participants were missing primary variables of interest. In the final behavioral sample (see below), SUPPS-P data was missing for two participants at baseline and eight participants at follow up. In the final neuroimaging sample two participants were missing SUPPS-P at baseline. As SUPPS-P was collected at both baseline and follow up, missing sessions (subject at visit) were excluded from analyses of SUPPS-P variables. One subject included in both behavioral and neuroimaging analysis was missing GA data. Six participants were missing SES data. As GA and SES were defined at baseline these participants were excluded from all analyses using these variables. No participants were missing FH, EXT, INT, or ETD data.

Stimuli from the AS task were presented on a flat screen behind the MRI scanner with E-Prime software (Schneider et al., 2002) and made visible to the subject with a mirror mounted on the head coil. Eye-tracking was performed using a long-range optics eye-tracking system (Applied Science Laboratories, Bedford, MA) and eye-position was measured through corneal reflection. Video monitoring was performed to ensure task compliance. Prior to the test session, a 9-point calibration was performed for each subject.

Eye movements in the response epoch were a scored as a correct AS if the first eye movement during the response epoch had a velocity greater than or equal to 30°/s (Gitelman, 2002) and was made in the mirror location of the peripheral cue and extended at least 2.5° visual angle from central fixation. In contrast, eye movements were marked as an incorrect AS if the first saccade in the response epochs was made toward the peripheral cue and extended at least 2.5° visual angle from central fixation, but then later directed toward the correct location, suggesting task compliance. Trials in which no saccade occurred or if the eye-tracker lost fixation were excluded from all analyses. Scan sessions were excluded from all behavioral and fMRI analysis if the proportion of excluded trials was greater than 33%.

Within the final analysis sample (N = 116), the overall proportion of excluded trials was 11.4% at baseline and 12.1% at follow up. Utilizing linear mixed effects models (see below for methods), the proportion of excluded trials was positively associated with the trait impulsivity measure of negative urgency [t = 2.01, χ(1)2 = 4.02, p = 0.045], suggesting that those with higher levels of negative urgency had more excluded trials. Those with a family history of substance use had a significantly lower proportion of excluded trials [t = −1.97, χ(1)2 = 3.87, p = 0.049]. Age at visit, EXT, INT, ETD, and positive urgency were not associated with the proportion of excluded eye tracking trials (see Supplemental Table 1).

Associations among participant characteristics were analyzed for sample description and covariate selection (Table 2). In order to provide standardized associations across all variable types, Pearson (continuous-continuous), polyserial (continuous-categorical), and polychoric (categorical-categorical) correlation values (Fox, 2010) are reported (see Table 2). Significance testing utilized Pearson correlation for continuous-continuous associations, Welch's t-test for continuous-categorical associations and chi-square testing for categorical-categorical associations.

Behavioral analysis was performed using R 3.1.2 (R Core Team, 2014). Mixed effects models (lme4 package; Bates et al., 2013) were used to examine associations between antisaccade (AS) correct response rates (accuracy) and latency and exceeds threshold drinking (ETD), substance use risk factors (FH, INT, EXT), and trait impulsivity (positive and negative urgency) and whether these associations were moderated by reward.

AS correct response rates were analyzed with generalized mixed-effects models with a logit link function as trial level data was binomially distributed (correct vs. incorrect). AS latencies were analyzed with linear mixed effects models and only included correct trials. Both linear and generalized-linear mixed-effects models used maximum likelihood estimation. Random intercepts were estimated for each subject. Reward condition (reward/neutral), visit (baseline/follow up), age at visit, exceeds threshold drinking (ETD), substance use risk factors (FH, EXT, INT), trait impulsivity measures (positive and negative urgency), generalized ability (GA), and socioeconomic status (SES) were included as fixed effects in three types of models. In the first type (Type A), ETD, FH, EXT, INT, and impulsivity measures were examined as predictors of AS performance in separate models. Second (Type B), ETD, FH, EXT, INT, and impulsivity measures were examined as joint predictors of AS performance in the same model. Third (Type C), GA and SES were included as additional covariates to the joint model. Reward, visit, and age at visit were included in all three model types. Secondary analysis included gender as a covariate in analysis of positive urgency (see below). Significance values for fixed effects were obtained through the car package in R (Fox et al., 2016; Wald chi-square test). Given the correlation between our variables of interest and our goal of defining sensitive and specific associations between substance use risk factors and AS performance, we highlight predictors that are significant across all three model types (A-C).

Twelve sessions were excluded from all fMRI analyses due to poor EPI coverage across imaging runs (n = 4) or technical errors (n = 8). Accordingly, the final neuroimaging sample consisted of 171 scan sessions distributed across 111 participants (participants at baseline, n = 95; participants at follow up, n = 76; participants with data at both visits, n = 60). Subject-level fMRI analyses were performed with Analysis and Visualization of Functional Neuroimages (AFNI, Bethesda, MD) software. In order to estimate the BOLD response at the trial- and individual AS epoch-level (cue, preparation, response), two level 1 GLM analyses were run for each subject at each visit using AFNI's 3dDeconvolve tool.

Reward and neutral trial-level BOLD responses were estimated with a 4,500 ms boxcar convolved with a gamma function (AFNI's block 4) and scaled to have an amplitude of 1. Separate regressors were included for correct, incorrect, and dropped/poor eye-tracking AS trials. Partial trials were modeled in the same manner but with 1,500 ms (cue partial trial) or 3,000 ms (cue and preparation partial trial) boxcars. Six rigid-body head motion parameters and their derivatives and run-wise 0 through 3rd order polynomials were used as nuisance regressors. The current and preceding TR were censored if the Euclidean norm head motion distance surpassed 0.9 mm, based on suggestions for task-based neuroimaging outlined in Siegel et al. (2014). Parameter estimates were examined both as main effects (Task > Fixation) and in comparisons with reward type (Reward > Neutral).

As reward-related changes in BOLD activation have been shown to vary according to demand characteristics (c.f., Geier et al., 2010), we also estimated BOLD activation in individual task-epochs (cue, preparation, and response). Epoch-level BOLD responses were estimated in a second model with individual regressors for cue, preparation, and response epochs. Relevant partial trials were included as examples of the epoch in question to aid in epoch-specific HRF estimation. All three epochs (cue, preparation, and response) were modeled with a 1,500 ms boxcar convolved with a gamma function (AFNI's block 4). The same nuisance regressors and motion censoring were used as in GLM-1.

We first examined associations of BOLD activation with ETD and substance use risk (FH, EXT, INT, impulsivity) in a priori regions of interest (ROIs) (Table 3). Regions were selected based on their association with the AS task in previous voxelwise analyses (Velanova et al., 2009; Geier et al., 2010), including the frontal and supplementary eye fields (FEF and SEF), pre supplementary motor area (pre-SMA), and executive function more generally (c.f., Wesley and Bickel, 2014), including posterior parietal cortices (PPC), dorsolateral prefrontal cortex (DLPFC), ventrolateral prefrontal cortex (VLPFC), inferior frontal gyrus (IFG) and dorsal anterior cingulate cortex (dACC). Additionally, given the focus on interactions with reward, we also selected regions of interest from the striatum, including the caudate, putamen, and nucleus accumbens (NAcc), which have previously been associated with reward effects and externalizing symptoms (Bjork et al., 2010b) and rewarded response inhibition tasks (Wilbertz et al., 2014).

Cortical ROIs were taken from a previous longitudinal neuroimaging study utilizing an AS task in adolescents (Ordaz et al., 2013), where the full methods are provided. Briefly, central coordinates were identified using topic and term searches in Neurosynth (http://neurosynth.org/), a meta-analysis tool for functional neuroimaging. Minor corrections to coordinates were made to ensure final ROIs overlapped with canonical eye movement regions (c.f., Munoz and Everling, 2004). From each central coordinate, spheres were grown, with the radius determined based on anatomical size and to avoid overlap (see Table 3). Striatum ROIs were taken from the Harvard-Oxford Atlas distributed through FSL software (Jenkinson et al., 2012). All ROIs were eroded such that they only included voxels with a 50% or greater probability of being gray matter in the MNI-152 template.

Our final list of ROIs included several pairs of bilateral regions (Table 3). In order to reduce the number of comparisons and because our hypotheses were not hemisphere specific, our primary analyses utilized left and right ROI pairs in one model with hemisphere as a within-subject factor. No ROI pair had a significant reward by hemisphere interaction (Table 3).

Nonzero mean BOLD activation (parameter estimates from GLM-1 & 2; see above) from correct trials were extracted from ROIs for each subject at each visit. As in the AS behavioral analysis, linear mixed effects models were utilized to examine the association of BOLD activation with ETD and substance use risk factors (FH, EXT, INT, impulsivity) and whether these associations were moderated by reward. Random intercepts were estimated for each subject and fixed effects were included in three phases (Type A, B, C). Additionally, session-wise motion estimates (proportion of censored volumes due to head motion) were used as a covariate in all models. Secondary analysis of positive urgency included gender as a covariate, as positive urgency was higher in males than females (see below). Scan sessions (subject at a particular visit) were excluded from analysis of an ROI if it did not have at least 90% epi coverage of the ROI or ROI pair. This exclusion never resulted in more than three scan sessions being omitted. In order to maximize the precision of individual subject estimates, we first examined main effects of ETD and risk factors (FH, EXT, INT, impulsivity) and their interactions with reward on trial-wise BOLD responses (GLM-1). Subsequently, we examined whether reward effects varied by AS epoch (reward by epoch interaction) using individual BOLD estimates of cue, preparation, and response from GLM-2. In order to assist in the interpretation of the direction of effects with BOLD activation, we additionally examined the association between BOLD activation and antisaccade correct response rate (accuracy). In all cases, significance values across ROIs were corrected for multiple comparisons using the false-discovery rate (FDR; q < 0.05).

To ensure the selection of ROIs did not bias our results we also performed voxelwise linear mixed effects analysis (3dLME; Chen et al., 2013) with age at visit and session-wise motion as covariates (model Type A). As in ROI analysis, voxelwise analysis only used data from correct trials. Voxelwise testing was limited to voxels that met each of the following criteria: 50% or greater probability of being gray matter in the MNI-152 template, full EPI coverage in all participants across all runs, and a significant simple effect of the task in reward or neutral trials (task vs. fixation), suggesting activation to the AS task significantly differed (positively or negatively). When examining ETD and risk factors (FH, INT, EXT, impulsivity), multiple comparison correction within this voxelwise space was performed using the intersection of voxelwise FDR correction (q < 0.05) and cluster size. AFNI's 3dClustsim program was used to determine cluster size threshold through a Monte Carlo simulation with parameters derived from mean spatial autocorrelation parameters from GLM-1 residuals.

Overall, antisaccade (AS) correct response rate (accuracy) (mean = 75.31%, SD = 16.67%) was comparable with previous work (Geier et al., 2010). Consistent with previous work (Geier et al., 2010), participant age was a significant positive predictor of AS accuracy [z = 5.36, χ(1)2 = 28.70, p < 0.001], with increased performance in older participants. Furthermore, as in Geier et al. (2010), AS accuracy was significantly higher in reward trials compared to neutral [z = 9.01, χ(1)2=81.16, p < 0.001; Figure 2].

Externalizing risk [z = −2.34, χ(1)2=5.50, p = 0.019] and higher levels of the trait impulsivity measure of positive urgency [z = −4.19, χ(1)2 = 17.53, p < 0.001] were associated with lower AS accuracy while controlling for age and reward condition (model type A; Figure 2). Externalizing risk and positive urgency remained significant predictors when ETD and the other substance use risk factors were entered into a multivariate model (model type B; Table 4), suggesting that EXT and positive urgency were each independently associated with AS correct response rate controlling for other variables. INT, FH, ETD, and negative urgency were not significantly associated with AS accuracy (see Table 4).

Neither externalizing risk nor positive urgency had significant interaction terms with the AS reward condition [externalizing risk by reward: z = 0.66, χ(1)2 = 0.43, p = 0.510; positive urgency by reward: z = 0.32, χ(1)2 = 0.12, p = 0.750], suggesting equivalent performance effects in reward and neutral trials (Table 4, Figure 2). Supporting this, simple effects testing revealed externalizing risk and positive urgency were significant predictors of AS accuracy in both neutral and reward trials [externalizing neutral: z = −2.43, p = 0.015; externalizing reward: z = −2.10, p = 0.036; positive urgency neutral: −4.06, p < 0.001; positive urgency reward: z = −3.73, p < 0.001]. Reward interactions were also not significant for INT, FH, ETD, or negative urgency (Table 4).

SES was not associated with AS accuracy [z = 0.82, χ(1)2 = 0.68, p = 0.410]. Generalized cognitive ability (GA) was positively associated with AS accuracy [z = 2.93, χ(1)2 = 8.57, p = 0.003]. When covarying GA (i.e., model type A + GA), the association between externalizing risk and AS accuracy was reduced to a trend [z = −1.69, χ(1)2 = 2.84, p = 0.092], but positive urgency remained a significant negative predictor of AS accuracy [z = −3.88, χ(1)2 = 15.09, p < 0.001].These results were unchanged in the full model with all other predictors (model Type C; see Table 4). Secondary analysis further showed positive urgency remained a significant negative predictor of AS accuracy while covarying gender [type A + gender: z = −4.30, χ(1)2 = 18.52, p < 0.001]. Gender was not a significant predictor in this model [z = 1.17, χ(1)2 = 1.17, p = 0.279].

AS latency on correct trials (mean = 440.86, SD = 59.62) had a negative relationship with participant age [t = −3.32, χ(1)2 = 11.01, p = 0.001]. AS latency was significantly shorter in rewarded trials compared to neutral trials [t = −6.99, χ(1)2 = 48.89, p < 0.001].

Of the substance use and risk measures, only positive urgency was significantly associated with AS latency, where higher positive urgency scores were associated with shorter latencies [t = −2.80, χ(1)2 = 7.84, p = 0.005] (see Table 4). The relationship between positive urgency and AS latency did not vary by reward condition [t = 0.42, χ(1)2 = 0.18, p = 0.676].

Neither SES [t = 0.55, χ(1)2 = 0.30, p = 0.583] nor GA [t = 0.24, χ(1)2 = 0.06, p = 0.809] was associated with AS latency. Gender was associated with AS latency, where males (least-squares mean: 421.74 ms) were faster than females (least-squares mean: 457.55 ms) on correct trials [Type A: t = −3.55, χ(1)2 = 12.57, p = 0.0003]. When entered as joint predictors (Type A + positive urgency and gender), positive urgency [t = −2.47, χ(1)2 = 6.13, p = 0.013] and gender [t = −3.22, χ(1)2 = 10.35, p = 0.001] remained significant predictors of AS latency.

The proportion of censored volumes due to head motion was negatively associated with age [t = −2.07, χ(1)2 = 4.287, p = 0.038] and positively associated with both trait impulsivity measures [positive urgency: t = 2.86, χ(1)2 = 8.18, p = 0.004; negative urgency: t = 2.80 χ(1)2 = 7.86, p = 0.005]. There was also trend for those in the externalizing risk group to have more motion [t = 1.94 χ(1)2 = 3.77, p = 0.052]. In contrast, the proportion of censored volumes due to motion was not associated with exceeding threshold drinking [t = 0.35, χ(1)2 = 0.12, p = 0.728], internalizing risk [t = −1.16, χ(1)2 = 1.35, p = 0.244], or family history of SUD [t = 0.50, χ(1)2 = 0.25 p = 0.617]. As detailed in the method section, the proportion of censored volumes due to head motion was used as a covariate in all fMRI analyses.

Robust activation in response to the task (task > fixation) was observed in canonical AS and cognitive control regions, including bilateral frontal eye fields (FEF), posterior parietal cortices (PPC), as well as in the anterior cingulate cortex (ACC) and striatum (see Figure 3). All selected ROIs had significant positive BOLD activation (task > fixation) in either the neutral or reward condition (Table 3). Increased activation to reward was observed at the voxelwise level in the striatum, bilateral PPC, and the right middle frontal gyrus/dorsolateral prefrontal cortex (DLPFC; Figure 3). The majority of ROIs had greater activation in reward trials, compared to neutral (Table 3).

Within a priori ROIs, positive urgency was associated with reduced BOLD activation in the putamen [t = −4.68, χ(1)2 = 21.86, p < 0.001, corrected], FEF [t = −2.64, χ(1)2 = 6.96, p = 0.031, corrected], and inferior frontal gyrus [IFG; t = −3.00, χ(1)2 = 8.99, p = 0.015, corrected] (Table 5). Family history of SUD was associated with reduced activation in the posterior parietal cortex, although this was only a trend after multiple comparison correction [t = −2.74, χ(1)2 = 7.49, p = 0.006 uncorrected, p = 0.068 corrected]. These effects were unchanged with ETD and all substance use risk measures entered in a multivariate model (Type B) or when including SES and GA as covariates (Type C). Furthermore, the association between positive urgency and BOLD activation in the putamen, FEF, and IFG, remained significant while covarying gender. No other measures had significant or trending, corrected main effects.

At the voxelwise level, positive urgency had a negative relationship with BOLD activation in the left inferior frontal gyrus (IFG) and right frontal eye field [FEF; voxelwise threshold (p) = 0.0012, q < 0.05, number of contiguous voxels (faces touching) > 24] (Figure 4), confirming results from a priori ROIs. Negative urgency had two voxels that reached FDR corrected significance (q < 0.05). EXT, INT, FH, ETD, and negative urgency did not have significant voxelwise BOLD main effects (minimum q's > 0.072).

With trial-wise BOLD responses (GLM-1), ETD and substance use risk measures did not have a significant, corrected interaction with reward in a priori ROIs (Table 6) or at the voxelwise level (minimum q's > 0.876).

Significant BOLD reward interactions for AS task epochs in ROIs are presented in Table 7. See Supplemental Tables 2–4 for full list. No reward interactions were observed in the cue epoch. In the preparation epoch, there was a significant interaction between subject age and reward condition in the NAcc ROI [t = 2.85, χ(1)2 = 8.10, p = 0.048, corrected]. Post hoc testing revealed a significant simple effect of age in reward trials [t_(204.03) = 2.38, p = 0.018] but a non-significant simple effect of age in neutral trials [t_(204.03) = −0.45, p = 0.651]. A significant reward interaction was also observed in the preparation epoch for exceeds threshold drinking in the FEF ROI [t = −2.92, χ(1)2 = 8.54, p = 0.038, corrected] with post-hoc testing demonstrating that those who exceeded threshold drinking had higher BOLD activation than those not meeting criteria in the neutral trials [t_(150.98) = 3.47, p < 0.001], but not reward trials [t_(150.98) = 1.41, p = 0.158] (Figure 5).

In the response epoch, there was a significant reward interaction for positive urgency in the SEF [t = 3.06, χ(1)2 = 9.35, p = 0.024, corrected], with post-hoc testing demonstrating a significant simple effect of positive urgency in reward trials [t_(323.86) = 2.24, p = 0.026] but a non-significant simple effect in neutral trials [t_(323.86) = −1.10, p = 0.274] (Figure 5). A reward interaction was also observed for externalizing risk in the IFG during the response epoch [t = 2.96, χ(1)2 = 8.75, p = 0.034, corrected], but post-hoc testing revealed non-significant simple effects in both reward and neutral trials (p's > 0.0998). However, no reward interactions were significant at the voxelwise level (minimum q's > 0.107).