Mental illness generates a major burden on health systems and patients (Kessler et al., Reference Kessler, Aguilar-Gaxiola, Alonso, Chatterji, Lee, Ormel and Wang2009; Patel et al., Reference Patel, Saxena, Lund, Thornicroft, Baingana, Bolton and UnÜtzer2018). Unfortunately, psychiatric diagnoses still rely on symptoms not causes, and so treatments remain poorly targeted (Krueger et al., Reference Krueger, Kotov, Watson, Forbes, Eaton, Ruggero and Bach2018). Currently, clinicians diagnose disorders categorically, most recently with the Diagnostic and Statistical Manual for Mental Disorders, 5th edition, text revised (DSM-5-TR; American Psychiatric Association, 2022) and International Classification of Diseases, 11th revision (ICD-11; World Health Organization, 2021). Despite the use of symptoms as a base for diagnosis, patients with the same diagnosis may share only a few or no symptoms with each other. There is a growing consensus that such heterogeneity limits progress in understanding the neurobiological mechanisms in psychiatric disorders (Latzman & DeYoung, Reference Latzman and DeYoung2020).

A major problem for the categorical approach is that psychiatric illnesses form spectra rather than categories. Symptoms vary continuously across individuals, and there is no clear separation of health from disorder (Krueger et al., Reference Krueger, Kotov, Watson, Forbes, Eaton, Ruggero and Bach2018, Reference Krueger, Hobbs, Conway, Dick, Dretsch, Eaton and Latzman2021; Krueger & Tackett, Reference Krueger, Tackett, Beauchaine and Hinshaw2015). In spectrum models, disorders that share common features and that are frequently comorbid are clustered together (Kessler et al., Reference Kessler, Ormel, Petukhova, McLaughlin, Green, Russo and Alonso2011; Krueger et al., Reference Krueger, Hicks, Patrick, Carlson, Iacono and McGue2002, Reference Krueger, Markon, Patrick, Benning and Kramer2007). Notably, the features tend to form latent dimensions similar to traits in personality (Latzman et al., Reference Latzman, Krueger, DeYoung and Michelini2021), with chronic symptoms of disorders reflecting extremes of the traits (Latzman et al., Reference Latzman, Krueger, DeYoung and Michelini2021). Of particular relevance to the current study, disorders in which individuals commonly show high trait disinhibition (Mullins-Sweatt et al., Reference Mullins-Sweatt, DeShong, Lengel, Helle and Krueger2019; Patrick, Venables et al., Reference Patrick, Venables, Yancey, Hicks, Nelson and Kramer2013; Venables et al., Reference Venables, Foell, Yancey, Kane, Engle and Patrick2018) (e.g., attention-deficit/hyperactivity disorder, ADHD; substance use disorders, SUDs; conduct disorder, CD; antisocial personality disorder, APD) form an “externalising” spectrum (Kessler et al., Reference Kessler, Ormel, Petukhova, McLaughlin, Green, Russo and Alonso2011; Krueger et al., Reference Krueger, Hicks, Patrick, Carlson, Iacono and McGue2002, Reference Krueger, Markon, Patrick, Benning and Kramer2007). The latent factor underlying this spectrum reflects common vulnerabilities, and so, when used as a target for studying biological mechanisms, allows more precision and reliability than single categorical diagnoses (Kotov et al., Reference Kotov, Cicero, Conway, DeYoung, Dombrovski, Eaton and Wright2022; Latzman & DeYoung, Reference Latzman and DeYoung2020). Since the same latent factor also appears to underlie trait disinhibition (Krueger et al., Reference Krueger, Markon, Patrick, Benning and Kramer2007; Latzman et al., Reference Latzman, Krueger, DeYoung and Michelini2021; Patrick, Kramer et al., Reference Patrick, Kramer, Krueger and Markon2013), in addition to clinical diagnoses of externalising disorders (EXTs), neural activity that has been associated with trait disinhibition is also implicated in externalising.

Broadly speaking, the neurobiological correlates of trait disinhibition/externalising include: (a) poor functioning in frontal-subcortical circuits (Knutson et al., Reference Knutson, Dal Monte, Schintu, Wassermann, Raymont, Grafman and Krueger2015; Krueger et al., Reference Krueger, Hobbs, Conway, Dick, Dretsch, Eaton and Latzman2021); (b) neurotransmitter imbalances, particularly dopamine (Beauchaine et al., Reference Beauchaine, Zisner and Sauder2017; Pattij & Vanderschuren, Reference Pattij and Vanderschuren2020); and (c) reduced brain potentials, especially P300 (Bowyer et al., Reference Bowyer, Joyner, Latzman, Venables, Foell and Patrick2020; Krueger et al., Reference Krueger, Hobbs, Conway, Dick, Dretsch, Eaton and Latzman2021; Pasion et al., Reference Pasion, Fernandes, Pereira and Barbosa2018; Venables et al., Reference Venables, Foell, Yancey, Kane, Engle and Patrick2018). These correlates result from different levels and forms of analyses from multiple domains with no clear theoretical connections at the neural level and so the observed neural changes are only loosely connected with each other. Most analyses of brain functioning uses group averages and/or categorical diagnoses. Therefore, we have no clear idea of variation of neural activity with variation in individual differences in key latent factors (Latzman et al., Reference Latzman, Krueger, DeYoung and Michelini2021; Pasion & Barbosa, Reference Pasion and Barbosa2019). So, here, we operationalised externalising as a higher order latent factor by clustering symptoms of disorders across the externalising spectrum; and then assessed how its variation across individuals correlated with electroencephalographic (EEG) indicators of the subsystems in the revised reinforcement sensitivity theory (RST, Corr & McNaughton, Reference Corr and McNaughton2008; Gray & McNaughton, Reference Gray and McNaughton2000).

The RST is an extension of the state biology of animal learning, motivation, and emotion to traits as individual differences in human personality (Corr & McNaughton, Reference Corr and McNaughton2008). RST provides a biologically detailed theoretical framework for integrating existing neuroscience with the spectrum perspectives of psychopathology (McNaughton & Corr, Reference McNaughton and Corr2022). The subsystems (Gray & McNaughton, Reference Gray and McNaughton2000) of RST control goal approach mediated by the behavioural activation system (BAS); goal repulsion mediated by the fight/flight/freeze system (FFFS); and goal conflict mediated by the behavioural inhibition system (BIS). Each system shows trait-like, stable patterns of reactivities, which at extreme levels are associated with psychopathology (Corr & McNaughton, Reference Corr and McNaughton2008). BAS and FFFS are activated by cues of upcoming reward and punishment that elicit approach and active avoidance respectively. The BIS responds to conflict between concurrently activated goals, e.g., conflict between approach and avoidance generated by the co-activation of the BAS and FFFS. When goal conflict is detected, the BIS inhibits ongoing behaviours and initiates assessment of the risks of entering a potentially punishing situation by triggering scanning of the environment and memory. Resolution of the conflict is achieved by information gathering, particularly risk assessment, and by amplifying risk aversion (an increase in the strength of avoidance) until either: (a) avoidance occurs or (b) a safety signal is detected during the concurrent risk assessment.

EXT dysfunctions could result from high BAS sensitivity (and so excessive approach behaviours), low BIS sensitivity (and so failure of inhibition of inappropriate behaviours), and low FFFS sensitivity (and so failure of avoidance behaviours), or combinations of the three. Depending on the supposed underlying psychopathologies, the predicted pattern of contribution at the disorder-specific level varies (Bijttebier et al., Reference Bijttebier, Beck, Claes and Vandereycken2009; Corr & McNaughton, Reference Corr, McNaughton, Beauchaine and Hinshaw2016). However, consistent with their links to trait disinhibition, high BAS sensitivity and low BIS sensitivity are likely common factors (Corr & McNaughton, Reference Corr, McNaughton, Beauchaine and Hinshaw2016).

Trait disinhibition is a broad personality construct. It includes facets of impulsivity, risk-taking, distractibility, low perfectionism, and irresponsibility (Krueger et al., Reference Krueger, Derringer, Markon, Watson and Skodol2012). Even at the facet level, the constructs are still complex and do not map onto neural mechanisms neatly (Yarkoni, Reference Yarkoni2015). With this caveat, we note that high BAS maps onto impulsivity via elevated sensitivity for reward (Gray, Reference Gray, Stahl, Iversen and Goodman1987; Smillie & Jackson, Reference Smillie and Jackson2006). In both animal and human models of impulsivity, increasing dopamine levels reduced impulsive symptoms (Beauchaine et al., Reference Beauchaine, Zisner and Sauder2017; London, Reference London2020; Pattij & Vanderschuren, Reference Pattij and Vanderschuren2020). On the premise that the release of dopamine is associated with reward signalling (Schultz, Reference Schultz2016; Schultz et al., Reference Schultz, Dayan and Montague1997), it has been suggested that impulsive individuals engage in excessive reward-seeking in order to maintain a functional level of dopamine (Corr & McNaughton, Reference Corr, McNaughton, Beauchaine and Hinshaw2016). Externalising individuals also exhibit excessive risk-taking (Krueger et al., Reference Krueger, Derringer, Markon, Watson and Skodol2012). They appear to be insensitive to the negative consequences associated with risky reward-seeking behaviours. Within the framework of RST, such insensitivity is most likely mediated by low BIS sensitivity, which leads to poor behavioural inhibition and excessive risk-taking, rather than low FFFS sensitivity, which controls behaviours when escape from punishment is the dominant goal (Corr & McNaughton, Reference Corr, McNaughton, Beauchaine and Hinshaw2016; Smillie & Jackson, Reference Smillie and Jackson2006).

A challenge in testing the RST in relation to human personality has been to separate BIS (passive avoidance, inhibition) from FFFS (active avoidance, action). The two are often confounded in extant work dominated by psychometric testing (see McNaughton & Corr, Reference McNaughton and Corr2022; Standen et al., Reference Standen, Firth, Sumich and Heym2022) that subsumed FFFS in the construct of BIS, making it hard to evaluate the results on BIS–psychopathology relationships. More recently, there have been attempts to separate FFFS from BIS with questionnaires. But, care must be taken with their meaning since there is now a wide range of different RST questionnaire systems (Krupić et al., Reference Krupić, Corr, Ručević, Križanić and Gračanin2016) albeit with some statistical overlap (Leue et al., Reference Leue, Reuter, Corr and Ettinger2022), and none of the scales involved has ever been neurally validated.

The bedrock of RST generally and the BIS in particular is the action of selective anxiolytic drugs (Gray, Reference Gray1970; Gray & McNaughton, Reference Gray and McNaughton2000) that are not antidepressant or panicolytic. They have a consistent effect on rodent hippocampal 4–12 Hz rhythmicity (McNaughton et al., Reference McNaughton, Kocsis and Hajós2007), and we have shown that all three major chemical classes also have a consistent effect on human right frontal 4–12 Hz rhythmicity elicited by goal conflict in a stop-signal task (McNaughton et al., Reference McNaughton, Swart, Neo, Bates and Glue2013; Shadli et al., Reference Shadli, Glue, McIntosh and McNaughton2015) showing that this is a biomarker of BIS sensitivity homologous to the rodent assay for anxiolytic drugs (McNaughton et al., Reference McNaughton, Kocsis and Hajós2007). Consistent with BIS theory, this right frontal goal-conflict-specific rhythmicity also varies with Spielberger trait anxiety (Spielberger et al., Reference Spielberger, Gorusch, Lushene, Vagg and Jacobs1983) in healthy participants and among groups of anxiety disorder patients (Shadli et al., Reference Shadli, Ando, McIntosh, Lodhia, Russell, Kirk and McNaughton2021).

In the context of the stop-signal task, one can apply distinct statistical contrasts to distinct times within a trial to extract rhythmicity specific not only to goal conflict but also to the conflict of outcome in contrast to intention (i.e., error, generating future avoidance, via activation of the FFFS). “For error, measurements are made locked to the timing of the mouse click. In the SST, an error occurs when a participant fails to stop clicking on a STOP trial. So, on STOP trials, a click represents a mistake, in contrast to a click in GO-trials. To extract processing specific to error, we contrasted failed STOP with successful GO trials, ignoring stop-signal delay. The majority of the EEG studies on error processing have focused on the ERN (Pasion & Barbosa, Reference Pasion and Barbosa2019). As mentioned earlier, the ERN is a response-locked potential, but is also associated with midfrontal theta rhythmic activity (Cavanagh et al., Reference Cavanagh, Zambrano-Vazquez and Allen2012; Taylor et al., Reference Taylor, Stern and Gehring2007; Yeung et al., Reference Yeung, Bogacz, Holroyd, Nieuwenhuis and Cohen2007). A Minnesota Multiphasic Personality Inventory-3 (MMPI-3) measure of the higher order spectrum of internalising psychopathology was positively correlated with action-error theta” (Neo et al., Reference Neo, Shadli, McNaughton and Sellbom2023).

Neural measures linked to BAS sensitivity are also available. Dopamine has been linked to approach motivation (DeYoung, Reference DeYoung2013; Wacker & Smillie, Reference Wacker and Smillie2015; Zisner & Beauchaine, Reference Zisner, Beauchaine, Beauchaine and Hinshaw2015) and to high extraversion and high disinhibition (Depue & Collins, Reference Depue and Collins1999; DeYoung, Reference DeYoung2013; Smillie et al., Reference Smillie, Jach, Hughes, Wacker, Cooper and Pickering2019; Zisner & Beauchaine, Reference Zisner, Beauchaine, Beauchaine and Hinshaw2015). Dopamine release is linked to signals of unexpected reward (Bromberg-Martin et al., Reference Bromberg-Martin, Matsumoto and Hikosaka2010; Fiorillo, Reference Fiorillo2013; Matsumoto & Hikosaka, Reference Matsumoto and Hikosaka2009). Reward prediction error generates an early (70–100 ms) orienting/salience component and a late (200–300 ms) motivational value component (Redgrave & Gurney, Reference Redgrave and Gurney2006; Schultz, Reference Schultz2016; Schultz et al., Reference Schultz, Stauffer and Lak2017) that are linked “the psychopathology-related constructs of low extraversion (social anxiety) and high disinhibition (impulsivity), respectively, making the evoked potential components biomarker candidates for indexing aberrant processing of unexpected reward” (Neo et al., Reference Neo, McNaughton and Sellbom2021, p.1)

In the current study, we used our established BIS biomarker (McNaughton et al., Reference McNaughton, Swart, Neo, Bates and Glue2013; Shadli et al., Reference Shadli, Ando, McIntosh, Lodhia, Russell, Kirk and McNaughton2021; Shadli et al., Reference Shadli, Glue, McIntosh and McNaughton2015), together with these EEG proxies of BAS (Neo et al., Reference Neo, McNaughton and Sellbom2021) and FFFS (Neo et al., Reference Neo, Shadli, McNaughton and Sellbom2023), to concurrently test if a latent factor of EXTs was associated with: (1) high BAS and (2) low BIS sensitivity. Critically, we tested for, but did not expect, an association with FFFS. As an active control for our latent measure of externalising, and as a contrast to the RST measures, we also tested for a negative association with P300, a brain potential that has consistently been associated with disinhibition/externalising (Bowyer et al., Reference Bowyer, Joyner, Latzman, Venables, Foell and Patrick2020; Krueger et al., Reference Krueger, Hobbs, Conway, Dick, Dretsch, Eaton and Latzman2021; Pasion et al., Reference Pasion, Fernandes, Pereira and Barbosa2018; Venables et al., Reference Venables, Foell, Yancey, Kane, Engle and Patrick2018). In doing so, we also considered whether the RST systems would increment P300 in the prediction of the latent externalising spectrum.

1. Methods

1.1. Participants

The data came from an archival database where, as part of a broader study, participants were recruited from the community with Facebook advertisements and flyers. In addition to a general advertisement looking for individuals to take part in a brain and personality study, a specific advertisement sought individuals with drug, alcohol, or anger management problems – or with ADHD or impulsivity. A total of 230 participants were tested and administered structured clinical interviews. EEG data for 206 and 160 participants were available for the stop-signal task and the gold bar-lemon task, respectively. After exclusions for EEG artefacts, and inconsistent or deviant responding detected using well-validated validity scales from the MMPI-3 (Ben-Porath & Tellegen, Reference Ben-Porath and Tellegen2020), the sample included 144 participants in the stop-signal task (for both BIS and FFFS measures) and 137/152 participants in the gold bar-lemon task (for BAS/P300 measures). Age range and gender profile for the 206 participants consisted of 92 males, 114 females; mean age = 36; SD = 9; range = 18–56). Forty-nine percent of the final sample met diagnostic criteria for at least one DSM-5 externalising disorder: ADHD (any type, 27%); history of conduct disorder (23%); antisocial personality disorder (21%); alcohol use disorder (18%); and cannabis use disorder (12%). These rates are higher than a typical community sample (Kessler et al., Reference Kessler, Ormel, Petukhova, McLaughlin, Green, Russo and Alonso2011). This is expected because we overweighted our advert of externalising problems during recruitment. The University of Otago Ethics Committee (Health) approved our study, and the approval number is H16/031.

1.2. Structured clinical interviews and MMPI-3 questionnaires

We administered the Structured Clinical Interview for DSM-5 Disorders – Research Version SCID-5, First et al. (Reference First, Williams, Karg and Spitzer2015) and Structured Clinical Interview for DSM-5 Personality Disorders (SCID-5-PD, First et al., Reference First, Williams, Benjamin and Spitzer2016) to assess DSM-5 EXTs. Specifically, we used the modules for attention-deficit hyperactivity disorder (ADHD), substance use disorders, intermittent explosive disorder from the SCID-5 and modules for antisocial personality disorder and conduct disorder (history) from the SCID-5-PD. Each EXT disorder was scored according to the number of criteria that were met for each disorder (i.e., count variables).

1.3. General EEG materials

We sampled EEG at 512 Hz referenced to CPz and recorded from FP1, FPz, FP2, F7, F3, Fz, F4, F8, T7, T8 C3, Cz, C4, P7, P3, Pz, P4, P8, M1, and M2 with an Advanced Neuro Technology (ANT) amplifier and ANT caps with AgCl electrodes. We tested and reduced impedances to below 10 KΩ with ANT software (eego) and used the Matlab 2019a plugin, EEGlab (version 2019_0) to process the data.

1.4. Stop-signal task (SST)

1.1.1. Task description

The stop-signal task (SST) (Logan et al., Reference Logan, Cowan and Davis1984), originally designed to assess motor inhibition, was modified to assess BIS goal conflict as described in (Shadli et al., Reference Shadli, Glue, McIntosh and McNaughton2015). Briefly, as shown in Figure 1, participants left/right mouse-clicked as quickly as possible when they see a left/right arrow (GO trials) but had to inhibit their clicking if they later hear a tone (STOP trials). On STOP trials, onset of the tone followed the arrow. The delay was determined by stop-signal delays (SSDs) controlled by separate staircases to deliver short, medium, and long SSDs. 99 STOP, and 296 GO trials were spread over three blocks (one STOP every four trials, counter-balanced in a fixed sequence). Participants practiced on 30 GO trials prior to the actual test.

Figure 1. Sequence of events in a trial. Onset of the tone from the time of the arrow presentation (SSDs) in a Stop trial are variable. A Go trial follows the same event sequence but without the onset of the tone. A smiley/frowny face is presented for a successful/unsuccessful withholding of a mouse click in a Stop trial. In a Go trial, a smiley/frowny face is presented for correct/incorrect responses. ms: milliseconds. ITI: 500 ms to 4000 ms; SSDs: see Table 1; A FI: Go Correct = 1000 ms; Stop Fail = 1500 ms; Stop Correct = 1700 ms.

1.1.2. EEG processing

All recorded EEG were re-referenced (M1/M2 average) and filtered (1–36 Hz, eegfiltnew) before we applied independent component analysis (runamica15) to extract eye blink and movement components, which were then identified and removed with an automatic function (pop_icflag) using a 90% similarity threshold. Next, we extracted two 1.5 s EEG segments centred, respectively, at the stop-signal and mouse-click from each trial. Trials with EEG > ±70 µV in any sensors were excluded (pop_eegthresh). For each 1.5 s segment, we applied two consecutive 1s Hanning windows that overlapped by 0.5 s, which delivered a 1 s epoch centred at the event of interest. The Hanning window is a cosine wave that improves the quality and time resolution of Fourier transforms while having the effect of focussing power extraction on the 0.5 s at its centre. The nominal 0.5 s segment after each event was then submitted to spectral analyses, followed by log transformation to normalise the distribution. Theta power over 4–7 Hz was then averaged to deliver single power indices. Note that data from 30 participants were sampled at 256 Hz instead of 512 Hz; these were analysed as sampled.

1.1.3. Neural measure of BIS tendencies

Previous work found that BIS-specific goal-conflict activity could range from 4 to 12 Hz at the right frontal sensor F8 (Neo & McNaughton, Reference Neo and McNaughton2011; Neo et al., Reference Neo, Thurlow and McNaughton2011; Neo et al., Reference Neo, Tinker and McNaughton2020; Shadli et al., Reference Shadli, Ando, McIntosh, Lodhia, Russell, Kirk and McNaughton2021; Shadli et al., Reference Shadli, Glue, McIntosh and McNaughton2015). However, conflict activity has also been observed at the midline sensor Fz (McNaughton et al., Reference McNaughton, Swart, Neo, Bates and Glue2013). F8 activity has been the focus of previous work because it showed consistency across studies. However, a preliminary analysis including the current sample showed that theta activity (4–7 Hz) at the midfrontal site Fz was a stronger indicator of externalising measured by MMPI-3. So here, we extracted goal-conflict-specific rhythmicity (GCSR4) using the same methods as (Shadli et al., Reference Shadli, Glue, McIntosh and McNaughton2015) but from Fz instead of F8. This Fz response, like the F8 one, has been shown to be sensitive to anxiolytics drugs (McNaughton et al., Reference McNaughton, Swart, Neo, Bates and Glue2013). A STOP trial differed from a GO trial only from the onset of the tone. To extract processing specific to response inhibition (signaled by the tone), GO trial theta activity matched to the onset of the stop-signal in the adjacent STOP trial was first subtracted from the STOP trial. (Matching GO trial activity data were extracted by inserting event marks into GO trials at SSDs matched to the paired STOP trials.) On the basis that goal conflict would be high in medium SSD trials (where stopping and going were equally likely) and low in early- and late-SSDs trials, the average of early and late STOP-GO trials was subtracted from the medium trials. This subtraction controlled for non-conflict-related activity. Participants with fewer than 5 trials in any of the averages were excluded.

1.1.4. Neural measure of FFFS tendencies

An error occurred when participants realised proprioceptively that they failed to inhibit a mouse-click in a STOP trial. We considered this a pure aversive signal that reflects receipt of a negative outcome, thereby eliciting avoidance tendencies controlled by the FFFS. We indexed this error-related activity with midfrontal (Fz) theta activity, which has previously been associated with errors (Cavanagh & Shackman, Reference Cavanagh and Shackman2015; Cavanagh et al., Reference Cavanagh, Zambrano-Vazquez and Allen2012; McLoughlin et al., Reference McLoughlin, Gyurkovics, Palmer and Makeig2021; Yeung et al., Reference Yeung, Bogacz, Holroyd, Nieuwenhuis and Cohen2007), and a broad range of psychopathology (McLoughlin et al., Reference McLoughlin, Gyurkovics, Palmer and Makeig2021). To extract processing specific to error, GO trial theta activity time-locked to a mouse click in a GO trial was subtracted from that in a STOP trial. The STOP-GO theta activity was then averaged. Individual average activity was calculated from a minimum of 20 trials.

Finally, note that the relationships between conflict and error theta activity and psychopathology (both externalising and internalising measured by MMPI-3) in the current sample were reported in Neo et al. (Reference Neo, Shadli, McNaughton and Sellbom2023). Briefly, we found that conflict was negatively associated with externalising, and error was positively associated with internalising.

1.5. Gold bar-lemon task

1.1.5. Task description

Designed to test for the human homologue of dopaminergic reward-prediction-error (RPE) signalling in animal models (Potts et al., Reference Potts, Martin, Burton and Montague2006), we used a version described in Neo et al. (Reference Neo, McNaughton and Sellbom2021). Briefly, as shown in Figure 2, two gold-bars/lemons were presented sequentially (80% of trials) and signalled expected reward ($0.50)/non-reward ($0.00). In 20% of the trials, the second cue (S2 in Figure 2) was a mismatch to the first cue (S1) and signalled unexpected reward/non-reward. 480 trials were spread over 8 blocks. Each trial was initiated by the program, and participants were told to attend to the trial closely. Although participants were informed that they would receive cash equal to the amount that they had made from the highest winnings block, the amount received was fixed at $6.

Figure 2. Sequence of events in an unpredicted reward trial. S1 indicates the first stimulus onset, and S2 indicates the second stimulus onset. S1 and S2 were always either a lemon or a bar. The period of interest for the current study is indicated by the vertical arrow.

1.1.6. EEG processing

All recorded EEG was filtered (eegfiltnew, lowpass at 100Hz), re-referenced to the average of all sensors other than the mastoids (M1/M2), and then segmented into epochs from S2 − 1100 through to S2 + 1400 ms. After linear drift was removed (eeglindetrend), each epoch was reduced to S2 − 100 ms through S2 + 400 ms and baseline corrected using the average of S2 − 100 ms to S2. Epochs with EEG > 70 µV in any recording sensors were rejected (pop_artextval). The epochs were than averaged according to the experimental conditions (unpredicted reward versus unpredicted non-reward trials). Participants with fewer than 20 trials in each average were excluded.

1.1.7. Measure of BAS activity

At midfrontal scalp sites, in the 200–300 ms after the onset of S2, unexpected reward trials showed a less negative potential than that of unexpected non-reward trials (Cooper et al., Reference Cooper, Duke, Pickering and Smillie2014; Smillie et al., Reference Smillie, Cooper and Pickering2011; Smillie et al., Reference Smillie, Jach, Hughes, Wacker, Cooper and Pickering2019; Walsh & Anderson, Reference Walsh and Anderson2012). The difference in the potential, i.e., RPE, is thought to reflect sensitivity to dopaminergic unexpected reward processing (Neo et al., Reference Neo, McNaughton and Sellbom2021). As mentioned in the introduction, dysfunctional dopamine reactivity has been theoretically linked to BAS via elevated reward sensitivity (Gray, Reference Gray, Stahl, Iversen and Goodman1987; Smillie & Jackson, Reference Smillie and Jackson2006). So here, we use RPE signalling as a proxy of BAS activity. Specifically, we subtracted mean amplitude of Fz event-related potentials in the 200–300 ms period post S2 in the unexpected non-reward trials from those in the unexpected reward trials. Previously, we reported the relationships of the current RPE and approach motivational traits measured by MMPI-3 (Neo et al., Reference Neo, McNaughton and Sellbom2021). We found that RPE was positively associated with impulsivity. The variation of the RPE signal across time, 100 ms before and 400 ms after the onset of S2, can be seen in Figure 3 of the prior study (Neo et al., Reference Neo, McNaughton and Sellbom2021).

Figure 3. Variation of ERP across time from 100 ms before to 400 ms after the onset of S2 in the Gold bar-lemon task. The waveform shown is an average of signals recorded from P7, P3, Pz, P4, and P8. The P300 peaks at about 270 ms.

1.1.8. P300

Weak “P300” reactivity is a well-established neural indicator of externalising (Bowyer et al., Reference Bowyer, Joyner, Latzman, Venables, Foell and Patrick2020; Pasion et al., Reference Pasion, Fernandes, Pereira and Barbosa2018; Venables et al., Reference Venables, Foell, Yancey, Kane, Engle and Patrick2018). The P300 is a positive brain wave observed ∼300 ms after the onset of an infrequent target stimulus. The P300 becomes larger when the target stimulus is rare. Weak P300 indicates poor attention allocation associated with disinhibition (Iacono et al., Reference Iacono, Malone and McGue2003). The P300 is well studied, with known neural circuitry (Polich, Reference Polich2007). EEG processing of P300 here followed the same steps as those of RPE. In contrast to RPE, P300 was averaged across the unexpected non-reward and unexpected reward trials. There are two known subcomponents of P300 (Polich, Reference Polich2007). Here, we averaged P300 across parietal sites (P7, P3, Pz, P4, and P8) on which the majority of the work on P300 and psychopathology has focused (Bernat et al., Reference Bernat, Ellis, Bachman and Hicks2020). The averaged ERP across time is shown from 100 ms before to 400 ms after the onset of S2 in Figure 3 and shows a P300 that peaks at ∼270 ms and was assessed as the average from 200 to 300 ms to capture its upward slope.

1.6. General procedures

Participants first completed a battery of questionnaires for a separate study while being prepared for EEG recording. Next, they completed the stop-signal task (∼25 min, Shadli et al., Reference Shadli, Glue, McIntosh and McNaughton2015), followed by the gold bar/lemon task (∼30 min, Neo et al., Reference Neo, McNaughton and Sellbom2021). Note that the latter task was introduced at a later stage of the study so some participants were only administered the SST. After completing the EEG recording, the participants were interviewed (see above) in a private room in the senior author’s research lab by a trained research assistant under the supervision of a registered clinical psychologist. The research assistants were clinical psychology students, with extensive practical experience, who had extensive training in these structured interviews by the senior author. They received cash winnings ($6) for the gold bar/lemon task and $50 of petrol or supermarket vouchers, as a thank you for volunteering with us. The experimental session lasted 3.5–4 h with breaks as needed.

1.7. Data analysis

First, we specified and estimated a one-factor EXT latent variable with the EXT disorder symptom counts as indicators. We used antisocial PD, conduct disorder, ADHD, intermittent explosive disorder symptoms, alcohol use disorder, and cannabis use disorderFootnote ¹ . We used the maximum likelihood estimator with robust scaling (MLR) and treated the indicators as count data using Mplus 8.8. Global model fit indices (confirmatory fit index, Tucker-Lewis index, root mean squared error of approximation) are not available in Mplus with count data. Therefore, for the sole purpose of approximating model fit, we estimated a standard MLR model with symptom counts as continuous indicators.

Second, we tested a series of multiple indicators and multiple causes (MIMIC) models in which each of the EEG-based RST proxies and P300 served as predictors to determine each bivariate association with overall EXT.

Finally, we examined an MIMIC model in which all of the RST and P300 predictors were entered simultaneously to evaluate the relative incremental predictions. Full information maximum likelihood estimation was used to handle missing EEG data.

2. Results

Correlations among the externalising symptom counts and the neural measures used can be found in the online supplemental materials (Tables S1 and S2, respectively).

The one-factor EXT model converged appropriately. The model fit was adequate (Comparative Fit Index [CFI] = .94, Tucker-Lewis Index [TLI] = .90, Root Mean Squared Error of Approximation [RMSEA] = .098, Standardized Root Mean Square Residual [SRMR] = .055) with RMSEA likely slightly higher than expected due the small model and sample size. All factor loadings were statistically significant (p < .001) – unstandardised coefficients appear in Table 1.Footnote ² Overall, we deemed the EXT model acceptable for the primary analyses.

Table 1. EXT symptom count factor loadings

We next tested our series of MIMIC models. Each of the RST constructs was added as covariates predicting the latent EXT factors. As shown in Table 2, our GCSR (BIS) and P300 variables were both significantly related to EXT in the bivariate analyses, whereas the RPE (BAS) and error (FFFS) variables were not related to EXT. In the multiple MIMIC model, all four predictors were entered simultaneously. Again, as shown in Table 2, both GCSR and P300 emerged as significant predictors, showing incremental predictive utility of EXT.

Table 2. Standardised coefficients from simple and multiple MIMIC models