Threat learning encompasses processes that rapidly generate associations between neutral stimuli and aversive outcomes. Influential theories implicate variations in the mechanisms underlying such conserved associative-learning processes in anxiety symptoms (Mineka and Oehlberg, 2008; Duits et al., 2015). Computational learning theory offers tools to quantify associative learning dynamics as they relate to variations in these mechanisms. Here, we study physiological data recorded during a threat learning paradigm in a large sample (N = 251) featuring a wide range of normative to pathological anxiety severity across childhood to adulthood. We use these data to uncover links among threat learning processes, their neuroanatomical substrates, and anxiety severity via latent learning parameters estimated by reinforcement learning models.

Neural circuits have evolved to allow rapid learning of threat associations following encounters with danger (Fanselow, 2018; LeDoux, 2014; LeDoux, 2000). Through threat conditioning, a neutral stimulus rapidly comes to elicit fear responding in anticipation of danger (Fanselow, 2018). Through extinction, such conditioned anticipatory responding is attenuated if the stimulus no longer predicts the occurrence of threat. Extensive research across species implicates conserved neural circuitry in these processes, highlighting their adaptive value (LeDoux, 2000).

At the same time, aspects of these processes could prove maladaptive. Thus, influential theories link individual differences in threat learning to the emergence and persistence of excessive threat-anticipatory fear responses which are central in expression of anxiety symptoms (Mineka and Oehlberg, 2008; Duits et al., 2015; Lissek, 2005; Barlow, 2002; Abend et al., 2021; Corchs and Schiller, 2019). Moreover, aberrant maturational processes in the circuitry supporting such learning may underlie the emergence of anxiety symptoms in childhood and adolescence (Beesdo et al., 2009; Kessler et al., 2012; Casey et al., 2015; Pattwell et al., 2012; Lau et al., 2011; Craske et al., 2018; Craske et al., 2012). However, studies in youth and adults attempting to link threat learning and anxiety symptoms yield inconsistent findings (Duits et al., 2015; Lissek, 2005; Dvir et al., 2019), hindering neuroscience research on normal and abnormal threat learning mechanisms.

Failure to detect robust and replicable associations between threat learning and anxiety symptoms could reflect limitations of standard analytic methods because these are not designed to capture associative learning dynamics (Lonsdorf et al., 2017; Ney et al., 2018). These limitations could potentially be overcome through applications of computational learning theory, which provides a mathematical framework for quantifying the temporal dynamics of associative learning processes, as indexed with latent variables (Garrison et al., 2013; Li and McNally, 2014; Keiflin and Janak, 2015; Rescorla and Wagner, 1972). Based on the measurement of prediction errors between expected and experienced outcomes, associative learning models were initially applied to reward learning (Keiflin and Janak, 2015; Schultz, 2013; Lee et al., 2012), including examining reward learning perturbation in anxiety (Huang et al., 2017; Aylward et al., 2019). Recent studies extend this approach to model associations between threat cues and aversive outcomes. These include studies in healthy participants (Wise et al., 2019; Tzovara et al., 2018; Zhang et al., 2016; Li et al., 2011; Schiller et al., 2008; Atlas et al., 2016), individuals with elevated anxiety symptoms (Browning et al., 2015; Michalska et al., 2019), and in adult patients with post-traumatic stress disorder and transdiagnostic symptom dimensions (Homan et al., 2019; Gagne et al., 2020; Wise and Dolan, 2020). These studies utilize variations of threat learning paradigms, such as reversal or instrumental conditioning procedures, to identify parameters that determine important aspects of learning, including some that relate to psychopathology. Here, we complement this body of work by studying the mechanisms that link individual differences in rapid threat conditioning and extinction to anxiety severity across the lifespan (Mineka and Oehlberg, 2008; Duits et al., 2015; Lissek, 2005).

In the current report, we address this gap by applying reinforcement learning models to skin conductance response (SCR) data recorded during a threat learning paradigm, in conjunction with structural brain imaging, to quantify age-dependent associations among threat learning processes, neuroanatomy, and anxiety severity. We complement and extend prior work in several ways. First, we study a sample that includes individuals with and without anxiety disorders falling along a wide age span, from age 8 to 50 years. This sample contains wide ranges of anxiety symptoms and age, which increase statistical power to detect associations among age, anxiety, and threat learning indices. Second, previous work modeled cue responding over many reinforcement trials (e.g. > 60 trials) (Homan et al., 2019; Tzovara et al., 2018; Zhang et al., 2016), which may more closely model the expression of conditioned responses. Here, we focus instead on the rapid learning of threat associations that takes place during a shorter schedule (Fanselow, 2018). Third, we model four learning effects: threat conditioning, threat generalization, safety learning, and threat extinction. This modeling approach provides tests of different theories that link mechanistic variations in these specific processes to the severity of anxiety (Mineka and Oehlberg, 2008; Lonsdorf et al., 2017; Craske et al., 2012; Abend et al., 2020; Michalska et al., 2017; Curzon et al., 2009). Finally, we use data from a relatively large sample (N = 215) of medication-free subjects, enabling parameter estimation with greater accuracy and over a large model space. We hypothesize that anxiety severity is associated with latent learning parameters (Mineka and Oehlberg, 2008; Duits et al., 2015); that age moderates specifically the association between extinction learning and anxiety (Casey et al., 2015); and that morphometry features in subcortical structures and prefrontal cortex moderate associations between learning parameters and anxiety (Fullana et al., 2018; Maren, 2001; Fullana et al., 2016).

This study modeled trial-by-trial psychophysiological data to quantify threat learning processes and their associations with anxiety symptoms and neuroanatomy (Mineka and Oehlberg, 2008; Duits et al., 2015; Fanselow, 2018; Casey et al., 2015). Several findings emerged. First, threat conditioning was optimally modeled with a RW model that featured Bayesian learning rate decay and a habituation rate. Within the conditioning process, slower safety learning rate was found to relate to more severe anxiety, and nucleus accumbens volume moderated this association. Finally, extinction learning was best modeled using a Bayesian (diminishing) learning rate decay and extinction of response to the safety, but not threat, stimulus was associated with greater anxiety. This study identifies specific latent threat-learning parameters that relate to anxiety symptom severity and neuroanatomical features.

The optimal fit for threat conditioning was a RW model that assumes Bayesian learning decay and habituation terms. A RW learning rule has long been proposed to underlie reinforcement learning, whereby point estimates of future outcomes are calculated by modifying current predictions according to prediction error, as weighted by a fixed learning rate (Rescorla and Wagner, 1972). Models based on a RW rule have been shown to describe the acquisition of conditioned fear responses in animals and humans (Herry and Johansen, 2014). Here, we show that an incrementally-diminishing threat-learning rate provides a better fit to physiology data (i.e., SCR) in humans, suggesting that conditioning of threat-anticipatory physiological responses is governed by a set learning parameter that is weighted less in proportion to additional reinforcement trials. It therefore places greater emphasis on learning from initial encounters with danger, supporting an adaptive role for this learning process in environments that contain threats (Fanselow, 2018).

The inclusion of a habituation term could indicate that conditioned responding decreases as the threat cue becomes more predictable, beyond what is explained by the basic RW model itself (Rescorla and Wagner, 1972). This could be a product of the 80% reinforcement schedule used here; a different reinforcement ratio could have affected this habituation term. Additionally, the habituation term could reflect our use of trial-level UCS value when modeling, rather than 0/1 as is often done, such that the magnitude of response to the UCS contributes to learning, rather than just its categorical absence/presence. Indeed, response to UCS in conditioning paradigms is often shown to naturally diminish as learning is achieved, potentially reflecting effective and reliable anticipation of it (Britton et al., 2013; Goodman et al., 2018; Johansen et al., 2010). Further, in other work (Abend, 2021) we show that increasing UCS potency results in increasing physiological response to the UCS. The response habituation term in the model may therefore account for individual differences in dynamic changes in the strength of response that both CS+ and UCS elicit.

This study extends prior work by modeling multiple learning effects that were evident in the data during threat conditioning and showed associations with anxiety severity and neuroanatomy. The introduction of an aversive UCS led to generalization of conditioned fear response to the CS- occurring most strongly early in conditioning, and followed by safety learning as this generalized response was inhibited by the CS- not predicting the aversive outcome. Learning theories of anxiety posit that each of these processes might play a role in the emergence and maintenance of pathological anxiety (Mineka and Oehlberg, 2008; Duits et al., 2015; Dymond et al., 2015; Vervliet et al., 2013; Lissek et al., 2014). Our modeling approach allowed us to test these theories. First, threat generalization was noted in the data, indicating that this process can be usefully modeled with this approach. However, the rate of generalization was not significantly associated with anxiety severity. Second, our findings provide support for propositions (Mineka and Oehlberg, 2008; Lissek et al., 2009) that link greater anxiety with slower safety learning rates (Mineka and Oehlberg, 2008; Craske et al., 2018). This association remained significant when controlling for CS+ habituation during conditioning, indicating that it was not a general habituation process. This finding suggests that fear responses to aversive events may initially generalize to neutral proximal stimuli, as a potentially adaptive mechanism, but that continued responding to these stimuli may contribute to the persistent and generalized fears associated with anxiety (Lissek et al., 2014). Importantly, the use of computational modeling enabled us to extend prior work by directly examining associations between learning rates and anxiety.

Our data suggest that the association between anxiety and safety learning was moderated by nucleus accumbens volume. Recent work indicates a role for this structure in regulating fear responses and threat-safety discrimination, including in the context of extinction learning (Ray et al., 2020; Dutta et al., 2021; Abraham et al., 2014). Our findings extend such work by suggesting that variation in accumbens structure relates to fear regulation following fear generalization to safety cues. Importantly, this effect is associated with anxiety severity, indicating potential pathophysiological relevance. These findings encourage more research on accumbens structure in predicting future anxiety as well as functional imaging work linking its structure and function in the context of safety learning.

Our results show that threat extinction learning also follows an incrementally-diminishing learning rate. As in threat generalization and safety learning, this is the first study to directly quantify the temporal dynamics of this learning process via modeling. We show an association between slower extinction of the safety stimulus and greater anxiety severity, providing some support for major theories of anxiety and its treatment that highlight extinction processes (Mineka and Oehlberg, 2008; Duits et al., 2015; Milad and Quirk, 2012; Barlow, 2002; Craske et al., 2018; Pittig et al., 2016; Vervliet et al., 2013; Papalini et al., 2020). Thus, in the context of exposure to conditioned stimuli during threat extinction, individuals with higher relative to lower anxiety symptoms diminish fear responses to safe stimuli more slowly. Importantly, this association was maintained when controlling for threat extinction rate as well as safety learning rate during conditioning, indicative of specificity. As in safety learning, accumbens structure moderated anxiety-learning associations. Thus, associations with anxiety severity, and moderation by accumbens anatomy, emerged for two processes in which the safe value of the CS- is to be learned, highlighting the role of response regulation to such stimuli in anxiety symptoms. It should be emphasized, however, that effects for extinction were weaker than those observed during conditioning and were evident only at trend-level, calling for caution in its interpretation and for replication.

Of note, prior work identified the RW-PH “hybrid” model as providing optimal fit to data for the expression of threat contingencies among healthy adults as well as adults with PTSD (Homan et al., 2019; Tzovara et al., 2018; Zhang et al., 2016). Here, we found that this model did not provide the best fit for rapid threat conditioning or extinction among the models tested. This discrepancy could potentially be attributed to the nature of threat learning processes studied. While we focused on rapid acquisition/extinction of threat contingencies which takes place over a short training schedule (LeDoux, 2000; Fanselow, 2018), prior modeling studies examined threat contingencies over much longer durations (over 80 CS+ trials). The hybrid model includes a cumulative associability term which could be more sensitive to tracking contingency values or the expression of conditioned fear as it is optimized over longer durations. Thus, differences in rapid, crude acquisition vs optimized expression of threat contingencies processes could account for model differences. As such, the current and prior work could be seen as complementary in terms of threat learning processes studied, and both sets of findings could usefully inform study design for future work.

While the design of this study offers a unique opportunity to examine age effects on threat learning as these relate to anxiety severity, an inherent challenge that arises in research on anxiety along the lifespan is how to combine anxiety data from youth and adult participants. Although the SCARED and STAI are each considered ‘gold standard’ measures in their respective target populations, they nevertheless are not identical. Under the assumption that these capture similar constructs, we uncovered several anxiety effects across the full age range. Alternatively, one may consider these measures to be incomparable, in which case analyses should be restricted to specific age groups. This alternative approach indicated an attenuation of anxiety-learning associations with age, as well as age group-specific moderation of these associations by brain structure. Given that age-dependent effects appear to emerge more strongly when age groups were considered separately, future research could emphasize this approach. When hypotheses call for analyses across the lifespan, consideration of optimal harmonization of clinical data is needed.

Our prior report used this dataset to examine links among associative threat learning, anxiety, and neuroanatomy using a ‘standard’ analytical approach whereby responses to CSs are averaged across trials (Abend et al., 2020). Here, we utilized a reinforcement modeling framework to test such links; this approach importantly extends our prior work in several ways. First, the approach applied here enabled us to directly quantify the temporal dynamics of CS and UCS associations that are at the core of associative learning (as opposed to measuring only responses to CSs and treating all trials as equivalent), thereby providing a more sensitive measure of threat learning processes, and how these relate to anxiety. Second, the current approach allowed us to quantify multiple characteristics evident in the learning processes (e.g. acquisition and habituation rates). As such, the modeling approach may be more sensitive to detect different, specific aspects of learning; this, in turn, enabled us to simultaneously test different theories linking these specific threat learning aspects and anxiety (e.g., slower safety learning) with greater sensitivity. Indeed, whereas our previous report failed to identify such associations, the current report provides novel insight on the pathophysiology of anxiety by linking variations in distinct threat learning processes to anxiety symptoms and subcortical structures. This distinction therefore highlights the utility of using a reinforcement learning framework to test questions on threat learning and anxiety.

Along these lines, the findings generated from this application of a computational approach to link biological, clinical, and imaging data could initiate continued research along several lines. In terms of clinical research, the identified associations between threat learning processes and anxiety severity could inform theories on the etiology of anxiety symptoms and guide studies that aim to further qualify the learning conditions and mechanisms that promote anxiety (Li and McNally, 2014; Vervliet et al., 2013). Further, influential theories that place threat extinction processes at the center of exposure-based therapy (Milad and Quirk, 2012; Craske et al., 2018; Pittig et al., 2016; Papalini et al., 2020) could be evaluated more sensitively using modeling-derived indices, while such treatment approaches could incorporate insight on safety learning (Craske et al., 2018). Translational research on mechanisms of threat learning could also benefit from these findings. Thus, neuroscience research in humans could focus on the specific processes identified here and extend our structural imaging findings to insight on function within this circuitry. Research in animals could complement such work by further delineating, via invasive manipulations (e.g., Likhtik and Paz, 2015), the roles in threat learning of the subcortical structures identified here.

Several important limitations should be acknowledged. First, modeling was based on relatively few task trials; while we focused on the specific processes of rapid learning of threat associations that take place over very few pairings (Fanselow, 2018), this could have led to noisier parameter estimates that could diminish accuracy and statistical power. Second, this was a cross-sectional study; a longitudinal design examining whether learning rates predict later emergence of symptoms would allow stronger inferences about developmental processes (Lonsdorf and Merz, 2017). Third, only structural MRI data were examined; functional imaging during threat learning could reveal additional correlates of learning circuitry (Homan et al., 2019), although the delivery of aversive stimuli to participants with anxiety in the MRI scanner presents a challenge (Thorpe et al., 2008). Fourth, SCR data were analyzed using a single method, which, while established, relies only on directly-observable effects; future studies may consider using novel, computational analysis methods which could potentially reveal effects not observed using the current method (Bach et al., 2018; Bach and Friston, 2013; Bach et al., 2020; Ojala and Bach, 2020). Along these lines, a multiverse approach may be used in future work to comprehensively compare multiple methods of quantifying threat learning. Fifth, while we used strict exclusion criteria to reduce effects of psychopathology other than anxiety, this does not eliminate all such potential effects (particularly when using the STAI Knowles and Olatunji, 2020); future research may wish to utilize computational approaches (e.g. bifactor models) to estimate symptom variability that is unique to anxiety (Tseng et al., 2021). Sixth, future research may consider alternative psychophysiology measures to SCR which may potentially be more sensitive to threat learning effects (Ojala and Bach, 2020). Finally, there were differences in sex and age between individuals included and excluded (non-responders) from analyses, as reported previously (Boucsein et al., 2012; Bari et al., 2020); while these differences did not influence our findings, they could still limit generalizability and thus future studies should consider such potential differences. Several strengths partially mitigate these limitations and address general shortcomings in threat learning research (Lonsdorf et al., 2017; Ney et al., 2018). First, the large sample size increases precision of estimated parameters (Asendorpf et al., 2020), offsetting, to some extent, the small number of trials used for modeling. Second, participants were carefully assessed and free of medications known to impact threat learning and psychophysiology (Lonsdorf et al., 2017). Third, wide anxiety-symptom and age ranges generate inferences with reasonable statistical power. Fourth, task and setting were identical for all participants, reducing measurement confounds and noise.

As a final technical comment, reinforcement learning model parameters were estimated using maximum likelihood techniques on individual subjects followed by model comparison. Future work could expand on this by using hierarchical Bayesian parameter estimation to reduce the variance around parameter estimates (Piray et al., 2019; van Geen and Gerraty, 2021; Lee and Newell, 2011). However, choosing prior distributions within the hierarchical Bayesian approach is not trivial and may not work for all of the models tested in this study. As such, future work could focus on fewer models, such as those that survived parameter recovery in this study, test a range of priors, and determine whether parameter estimation could be improved.

In conclusion, in this study we used computational modeling to index dynamic learning processes associated with threat learning. Through this modeling approach, we quantified these learning processes, revealed specific associations with anxiety symptoms, and identified neuroanatomical substrates. These findings extend our knowledge of how these learning processes manifest in humans and how variations in these could potentially contribute to anxiety symptomatology.

We cannot share the full dataset due to the NIH IRB requirements, which require participants to explicitly consent to their data being shared publicly. An important element in that is to protect patients who agree to participate in studies that relate to their psychopathology. Such consent was not acquired from most participants; as such, we cannot upload our complete dataset in its raw or deidentified form, or derivatives of the data, since we will be violating IRB protocols. Still, a subset of participants did consent to data sharing and we have uploaded their data as noted in the revised manuscript (https://github.com/rany-abend/threat_learning_eLife, copy archived at swh:1:rev:221919cc02c4b5ae00fc21764095b8b4d2a14201). Researchers interested in potentially acquiring access to the data could contact Dr. Daniel Pine (https://pined@mail.nih.gov/), Chief of the Emotion and Development Branch at NIH, with a research proposal; as per IRB rules, the IRB may approve adding such researchers as Associate Investigators if a formal collaboration is initiated. No commercial use of the data is alowed. The modeling and imaging analyses have now been uploaded in full as source code files.

1. 1. Abbott A

   (2008) Psychiatric genetics: The brains of the family

   Nature 454:154–157.

   https://doi.org/10.1038/454154a

   • PubMed
   • Google Scholar
2. 1. Abend R
   2. Naim R
   3. Pergamin-Hight L
   4. Fox NA
   5. Pine DS
   6. Bar-Haim Y

   (2019) Age Moderates Link Between Training Effects and Treatment Response to Attention Bias Modification Treatment for Social Anxiety Disorder

   Journal of Abnormal Child Psychology 47:881–894.

   https://doi.org/10.1007/s10802-018-0494-7

   • PubMed
   • Google Scholar
3. 1. Abend R
   2. Gold AL
   3. Britton JC
   4. Michalska KJ
   5. Shechner T
   6. Sachs JF
   7. Winkler AM
   8. Leibenluft E
   9. Averbeck BB
   10. Pine DS

   (2020) Anticipatory Threat Responding: Associations With Anxiety, Development, and Brain Structure

   Biological Psychiatry 87:916–925.

   https://doi.org/10.1016/j.biopsych.2019.11.006

   • PubMed
   • Google Scholar
4. 1. Abend R

   (2021) Threat-anticipatory psychophysiological response is enhanced in pediatric anxiety and correlates with prefrontal cortex neuroanatomy

   Journal of Psychiary and Neuroscience 46:e10.

   https://doi.org/10.1503/jpn.200110

   • Google Scholar
5. 1. Abend R
   2. Bajaj MA
   3. Coppersmith DDL
   4. Kircanski K
   5. Haller SP
   6. Cardinale EM
   7. Salum GA
   8. Wiers RW
   9. Salemink E
   10. Pettit JW
   11. Pérez-Edgar K
   12. Lebowitz ER
   13. Silverman WK
   14. Bar-Haim Y
   15. Brotman MA
   16. Leibenluft E
   17. Fried EI
   18. Pine DS

   (2021) A computational network perspective on pediatric anxiety symptoms

   Psychological Medicine 51:1752–1762.

   https://doi.org/10.1017/S0033291720000501

   • PubMed
   • Google Scholar
6. 1. Abraham AD
   2. Neve KA
   3. Lattal KM

   (2014) Dopamine and extinction: A convergence of theory with fear and reward circuitry

   Neurobiology of Learning and Memory 108:65–77.

   https://doi.org/10.1016/j.nlm.2013.11.007

   • PubMed
   • Google Scholar
7. 1. Asendorpf JB
   2. Conner M
   3. De Fruyt F
   4. De Houwer J
   5. Denissen JJA
   6. Fiedler K
   7. Fiedler S
   8. Funder DC
   9. Kliegl R
   10. Nosek BA
   11. Perugini M
   12. Roberts BW
   13. Schmitt M
   14. Van Aken MAG
   15. Weber H
   16. Wicherts JM

   (2020) Recommendations for Increasing Replicability in Psychology

   European Journal of Personality 27:108–119.

   https://doi.org/10.1002/per.1919

   • Google Scholar
8. 1. Atlas LY
   2. Doll BB
   3. Li J
   4. Daw ND
   5. Phelps EA

   (2016) Instructed knowledge shapes feedback-driven aversive learning in striatum and orbitofrontal cortex, but not the amygdala

   eLife 5:e15192.

   https://doi.org/10.7554/eLife.15192

   • PubMed
   • Google Scholar
9. 1. Aylward J
   2. Valton V
   3. Ahn W-Y
   4. Bond RL
   5. Dayan P
   6. Roiser JP
   7. Robinson OJ

   (2019) Altered learning under uncertainty in unmedicated mood and anxiety disorders

   Nature Human Behaviour 3:1116–1123.

   https://doi.org/10.1038/s41562-019-0628-0

   • PubMed
   • Google Scholar
10. 1. Bach DR
    2. Friston KJ

    (2013) Model-based analysis of skin conductance responses: Towards causal models in psychophysiology

    Psychophysiology 50:15–22.

    https://doi.org/10.1111/j.1469-8986.2012.01483.x

    • PubMed
    • Google Scholar
11. 1. Bach DR
    2. Castegnetti G
    3. Korn CW
    4. Gerster S
    5. Melinscak F
    6. Moser T

    (2018) Psychophysiological modeling: Current state and future directions

    Psychophysiology 55:e13214.

    https://doi.org/10.1111/psyp.13209

    • PubMed
    • Google Scholar
12. 1. Bach DR
    2. Melinščak F
    3. Fleming SM
    4. Voelkle MC

    (2020) Calibrating the experimental measurement of psychological attributes

    Nature Human Behaviour 4:1229–1235.

    https://doi.org/10.1038/s41562-020-00976-8

    • PubMed
    • Google Scholar
13. 1. Bari DS
    2. Yacoob Aldosky HY
    3. Martinsen ØG

    (2020) Simultaneous measurement of electrodermal activity components correlated with age-related differences

    Journal of Biological Physics 46:177–188.

    https://doi.org/10.1007/s10867-020-09547-4

    • PubMed
    • Google Scholar
14. Book

    1. Barlow DHA

    (2002)

    Its Disorders: The Nature and Treatment of Anxiety and Panic

    The Guilford.

    • Google Scholar
15. 1. Beesdo K
    2. Knappe S
    3. Pine DS

    (2009) Anxiety and anxiety disorders in children and adolescents: developmental issues and implications for DSM-V

    The Psychiatric Clinics of North America 32:483–524.

    https://doi.org/10.1016/j.psc.2009.06.002

    • PubMed
    • Google Scholar
16. 1. Behrens B
    2. Swetlitz C
    3. Pine DS
    4. Pagliaccio D

    (2019) The Screen for Child Anxiety Related Emotional Disorders (SCARED): Informant Discrepancy, Measurement Invariance, and Test-Retest Reliability

    Child Psychiatry and Human Development 50:473–482.

    https://doi.org/10.1007/s10578-018-0854-0

    • PubMed
    • Google Scholar
17. 1. Birmaher B
    2. Khetarpal S
    3. Brent D
    4. Cully M
    5. Balach L
    6. Kaufman J
    7. Neer SM

    (1997) The Screen for Child Anxiety Related Emotional Disorders (SCARED): scale construction and psychometric characteristics

    Journal of the American Academy of Child and Adolescent Psychiatry 36:545–553.

    https://doi.org/10.1097/00004583-199704000-00018

    • PubMed
    • Google Scholar
18. 1. Birmaher B
    2. Brent DA
    3. Chiappetta L
    4. Bridge J
    5. Monga S
    6. Baugher M

    (1999) Psychometric properties of the Screen for Child Anxiety Related Emotional Disorders (SCARED): A replication study

    Journal of the American Academy of Child and Adolescent Psychiatry 38:1230–1236.

    https://doi.org/10.1097/00004583-199910000-00011

    • PubMed
    • Google Scholar
19. 1. Boucsein W
    2. Fowles DC
    3. Grimnes S
    4. Ben-Shakhar G
    5. roth WT
    6. Dawson ME
    7. Filion DL
    8. Society for Psychophysiological Research Ad Hoc Committee on Electrodermal Measures

    (2012) Publication recommendations for electrodermal measurements

    Psychophysiology 49:1017–1034.

    https://doi.org/10.1111/j.1469-8986.2012.01384.x

    • PubMed
    • Google Scholar
20. 1. Britton JC
    2. Grillon C
    3. Lissek S
    4. Norcross MA
    5. Szuhany KL
    6. Chen G
    7. Ernst M
    8. Nelson EE
    9. Leibenluft E
    10. Shechner T
    11. Pine DS

    (2013) Response to learned threat: An FMRI study in adolescent and adult anxiety

    The American Journal of Psychiatry 170:1195–1204.

    https://doi.org/10.1176/appi.ajp.2013.12050651

    • PubMed
    • Google Scholar
21. 1. Browning M
    2. Behrens TE
    3. Jocham G
    4. O’Reilly JX
    5. Bishop SJ

    (2015) Anxious individuals have difficulty learning the causal statistics of aversive environments

    Nature Neuroscience 18:590–596.

    https://doi.org/10.1038/nn.3961

    • Google Scholar
22. 1. Casey BJ
    2. Glatt CE
    3. Lee FS

    (2015) Treating the Developing versus Developed Brain: Translating Preclinical Mouse and Human Studies

    Neuron 86:1358–1368.

    https://doi.org/10.1016/j.neuron.2015.05.020

    • PubMed
    • Google Scholar
23. 1. Corchs F
    2. Schiller D

    (2019) Threat-related disorders as persistent motivational states of defense

    Current Opinion in Behavioral Sciences 26:62–68.

    https://doi.org/10.1016/j.cobeha.2018.10.007

    • PubMed
    • Google Scholar
24. 1. Craske MG
    2. Wolitzky-Taylor KB
    3. Mineka S
    4. Zinbarg R
    5. Waters AM
    6. Vrshek-Schallhorn S
    7. Epstein A
    8. Naliboff B
    9. Ornitz E

    (2012) Elevated responding to safe conditions as a specific risk factor for anxiety versus depressive disorders: evidence from a longitudinal investigation

    Journal of Abnormal Psychology 121:315–324.

    https://doi.org/10.1037/a0025738

    • PubMed
    • Google Scholar
25. 1. Craske MG
    2. Hermans D
    3. Vervliet B

    (2018) State-of-the-art and future directions for extinction as a translational model for fear and anxiety

    Philos T R Soc B 373:e25.

    https://doi.org/10.1098/rstb.2017.0025

    • Google Scholar
26. Book

    1. Curzon P
    2. Rustay NR
    3. Browman KE

    (2009)

    Methods of Behavior Analysis in Neuroscience

    CRC Press/Taylor & Francis.

    • Google Scholar
27. 1. Dale AM
    2. Sereno MI

    (1993) Improved Localizadon of Cortical Activity by Combining EEG and MEG with MRI Cortical Surface Reconstruction: A Linear Approach

    Journal of Cognitive Neuroscience 5:162–176.

    https://doi.org/10.1162/jocn.1993.5.2.162

    • Google Scholar
28. 1. Dale AM
    2. Fischl B
    3. Sereno MI

    (1999) Cortical surface-based analysis. I. Segmentation and surface reconstruction

    NeuroImage 9:179–194.

    https://doi.org/10.1006/nimg.1998.0395

    • PubMed
    • Google Scholar
29. 1. Den ML
    2. Graham BM
    3. Newall C
    4. Richardson R

    (2015) Teens that fear screams: A comparison of fear conditioning, extinction, and reinstatement in adolescents and adults

    Developmental Psychobiology 57:818–832.

    https://doi.org/10.1002/dev.21330

    • PubMed
    • Google Scholar
30. 1. Desikan RS
    2. Ségonne F
    3. Fischl B
    4. Quinn BT
    5. Dickerson BC
    6. Blacker D
    7. Buckner RL
    8. Dale AM
    9. Maguire RP
    10. Hyman BT
    11. Albert MS
    12. Killiany RJ

    (2006) An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest

    NeuroImage 31:968–980.

    https://doi.org/10.1016/j.neuroimage.2006.01.021

    • PubMed
    • Google Scholar
31. 1. Duits P
    2. Cath DC
    3. Lissek S
    4. Hox JJ
    5. Hamm AO
    6. Engelhard IM
    7. van den Hout MA
    8. Baas JMP

    (2015) Updated meta-analysis of classical fear conditioning in the anxiety disorders

    Depression and Anxiety 32:239–253.

    https://doi.org/10.1002/da.22353

    • PubMed
    • Google Scholar
32. 1. Dutta S
    2. Beaver J
    3. Halcomb CJ
    4. Jasnow AM

    (2021) Dissociable roles of the nucleus accumbens core and shell subregions in the expression and extinction of conditioned fear

    Neurobiology of Stress 15:100365.

    https://doi.org/10.1016/j.ynstr.2021.100365

    • PubMed
    • Google Scholar
33. 1. Dvir M
    2. Horovitz O
    3. Aderka IM
    4. Shechner T

    (2019) Fear conditioning and extinction in anxious and non-anxious youth: A meta-analysis

    Behaviour Research and Therapy 120:103431.

    https://doi.org/10.1016/j.brat.2019.103431

    • PubMed
    • Google Scholar
34. 1. Dymond S
    2. Dunsmoor JE
    3. Vervliet B
    4. Roche B
    5. Hermans D

    (2015) Fear Generalization in Humans: Systematic Review and Implications for Anxiety Disorder Research

    Behavior Therapy 46:561–582.

    https://doi.org/10.1016/j.beth.2014.10.001

    • PubMed
    • Google Scholar
35. 1. Elwood LS
    2. Wolitzky-Taylor K
    3. Olatunji BO

    (2012) Measurement of anxious traits: a contemporary review and synthesis

    Anxiety, Stress, and Coping 25:647–666.

    https://doi.org/10.1080/10615806.2011.582949

    • PubMed
    • Google Scholar
36. 1. Fanselow MS

    (2018) The Role of Learning in Threat Imminence and Defensive Behaviors

    Current Opinion in Behavioral Sciences 24:44–49.

    https://doi.org/10.1016/j.cobeha.2018.03.003

    • PubMed
    • Google Scholar
37. 1. Fiesler E
    2. Beale R

    (1995) Neural Networks for Pattern Recognition

    Semantic Scholar 1:646.

    https://doi.org/10.1201/9781420050646

    • Google Scholar
38. 1. First MB
    2. Spitzer RL
    3. Gibbon M
    4. Williams JBW

    (2002) Structured Clinical Interview for DSM-IV-TR Axis I Disorders, Research Version, Patient Edition (SCID-I/P

    Biometrics Research, New York State Psychiatric Institute 1:e66.

    https://doi.org/10.1007/978-1-4419-1005-9_66

    • Google Scholar
39. 1. Fischl B
    2. Dale AM

    (2000) Measuring the thickness of the human cerebral cortex from magnetic resonance images

    PNAS 97:11050–11055.

    https://doi.org/10.1073/pnas.200033797

    • Google Scholar
40. 1. Fischl B
    2. Salat DH
    3. Busa E
    4. Albert M
    5. Dieterich M
    6. Haselgrove C
    7. van der Kouwe A
    8. Killiany R
    9. Kennedy D
    10. Klaveness S
    11. Montillo A
    12. Makris N
    13. Rosen B
    14. Dale AM

    (2002) Whole brain segmentation: automated labeling of neuroanatomical structures in the human brain

    Neuron 33:341–355.

    https://doi.org/10.1016/s0896-6273(02)00569-x

    • PubMed
    • Google Scholar
41. 1. Fischl B
    2. Salat DH
    3. van der Kouwe AJW
    4. Makris N
    5. Ségonne F
    6. Quinn BT
    7. Dale AM

    (2004) Sequence-independent segmentation of magnetic resonance images

    NeuroImage 23 Suppl 1:S69–S84.

    https://doi.org/10.1016/j.neuroimage.2004.07.016

    • PubMed
    • Google Scholar
42. 1. Fullana MA
    2. Harrison BJ
    3. Soriano-Mas C
    4. Vervliet B
    5. Cardoner N
    6. Àvila-Parcet A
    7. Radua J

    (2016) Neural signatures of human fear conditioning: an updated and extended meta-analysis of fMRI studies

    Molecular Psychiatry 21:500–508.

    https://doi.org/10.1038/mp.2015.88

    • PubMed
    • Google Scholar
43. 1. Fullana MA
    2. Albajes-Eizagirre A
    3. Soriano-Mas C
    4. Vervliet B
    5. Cardoner N
    6. Benet O
    7. Radua J
    8. Harrison BJ

    (2018) Fear extinction in the human brain: A meta-analysis of fMRI studies in healthy participants

    Neuroscience and Biobehavioral Reviews 88:16–25.

    https://doi.org/10.1016/j.neubiorev.2018.03.002

    • PubMed
    • Google Scholar
44. 1. Gagne C
    2. Zika O
    3. Dayan P
    4. Bishop SJ

    (2020) Impaired adaptation of learning to contingency volatility in internalizing psychopathology

    eLife 9:e61387.

    https://doi.org/10.7554/eLife.61387

    • PubMed
    • Google Scholar
45. 1. Garrison J
    2. Erdeniz B
    3. Done J

    (2013) Prediction error in reinforcement learning: a meta-analysis of neuroimaging studies

    Neuroscience and Biobehavioral Reviews 37:1297–1310.

    https://doi.org/10.1016/j.neubiorev.2013.03.023

    • PubMed
    • Google Scholar
46. 1. Gold AL
    2. Brotman MA
    3. Adleman NE
    4. Lever SN
    5. Steuber ER
    6. Fromm SJ
    7. Mueller SC
    8. Pine DS
    9. Leibenluft E

    (2016) Comparing Brain Morphometry Across Multiple Childhood Psychiatric Disorders

    Journal of the American Academy of Child and Adolescent Psychiatry 55:1027–1037.

    https://doi.org/10.1016/j.jaac.2016.08.008

    • PubMed
    • Google Scholar
47. 1. Gold AL
    2. Steuber ER
    3. White LK
    4. Pacheco J
    5. Sachs JF
    6. Pagliaccio D
    7. Berman E
    8. Leibenluft E
    9. Pine DS

    (2017) Cortical Thickness and Subcortical Gray Matter Volume in Pediatric Anxiety Disorders

    Neuropsychopharmacology : Official Publication of the American College of Neuropsychopharmacology 42:2423–2433.

    https://doi.org/10.1038/npp.2017.83

    • PubMed
    • Google Scholar
48. 1. Gold AL
    2. Abend R
    3. Britton JC
    4. Behrens B
    5. Farber M
    6. Ronkin E
    7. Chen G
    8. Leibenluft E
    9. Pine DS

    (2020) Age Differences in the Neural Correlates of Anxiety Disorders: An fMRI Study of Response to Learned Threat

    The American Journal of Psychiatry 177:454–463.

    https://doi.org/10.1176/appi.ajp.2019.19060650

    • PubMed
    • Google Scholar
49. 1. Goodman AM
    2. Harnett NG
    3. Knight DC

    (2018) Pavlovian conditioned diminution of the neurobehavioral response to threat

    Neuroscience and Biobehavioral Reviews 84:218–224.

    https://doi.org/10.1016/j.neubiorev.2017.11.021

    • PubMed
    • Google Scholar
50. 1. Herry C
    2. Johansen JP

    (2014) Encoding of fear learning and memory in distributed neuronal circuits

    Nature Neuroscience 17:1644–1654.

    https://doi.org/10.1038/nn.3869

    • PubMed
    • Google Scholar
51. 1. Homan P
    2. Levy I
    3. Feltham E
    4. Gordon C
    5. Hu J
    6. Li J
    7. Pietrzak RH
    8. Southwick S
    9. Krystal JH
    10. Harpaz-Rotem I
    11. Schiller D

    (2019) Neural computations of threat in the aftermath of combat trauma

    Nature Neuroscience 22:470–476.

    https://doi.org/10.1038/s41593-018-0315-x

    • PubMed
    • Google Scholar
52. 1. Huang H
    2. Thompson W
    3. Paulus MP

    (2017) Computational Dysfunctions in Anxiety: Failure to Differentiate Signal From Noise

    Biological Psychiatry 82:440–446.

    https://doi.org/10.1016/j.biopsych.2017.07.007

    • PubMed
    • Google Scholar
53. 1. Johansen JP
    2. Tarpley JW
    3. LeDoux JE
    4. Blair HT

    (2010) Neural substrates for expectation-modulated fear learning in the amygdala and periaqueductal gray

    Nature Neuroscience 13:979–986.

    https://doi.org/10.1038/nn.2594

    • PubMed
    • Google Scholar
54. 1. Kaufman J
    2. Birmaher B
    3. Brent D
    4. Rao U
    5. Flynn C
    6. Moreci P
    7. Williamson D
    8. Ryan N

    (1997) Schedule for Affective Disorders and Schizophrenia for School-Age Children-Present and Lifetime Version (K-SADS-PL): initial reliability and validity data

    Journal of the American Academy of Child and Adolescent Psychiatry 36:980–988.

    https://doi.org/10.1097/00004583-199707000-00021

    • PubMed
    • Google Scholar
55. 1. Keiflin R
    2. Janak PH

    (2015) Dopamine Prediction Errors in Reward Learning and Addiction: From Theory to Neural Circuitry

    Neuron 88:247–263.

    https://doi.org/10.1016/j.neuron.2015.08.037

    • PubMed
    • Google Scholar
56. Book

    1. Kent BR

    (2015)

    CA

    San Rafael: Morgan & Claypool Publishers.

    • Google Scholar
57. 1. Kessler RC
    2. Avenevoli S
    3. McLaughlin KA
    4. Green JG
    5. Lakoma MD
    6. Petukhova M
    7. Pine DS
    8. Sampson NA
    9. Zaslavsky AM
    10. Merikangas KR

    (2012) Lifetime co-morbidity of DSM-IV disorders in the US National Comorbidity Survey Replication Adolescent Supplement (NCS-A

    Psychological Medicine 42:1997–2010.

    https://doi.org/10.1017/S0033291712000025

    • PubMed
    • Google Scholar
58. 1. Knowles KA
    2. Olatunji BO

    (2020) Specificity of trait anxiety in anxiety and depression: Meta-analysis of the State-Trait Anxiety Inventory

    Clinical Psychology Review 82:101928.

    https://doi.org/10.1016/j.cpr.2020.101928

    • Google Scholar
59. 1. Lau JYF
    2. Lissek S
    3. Nelson EE
    4. Lee Y
    5. Roberson-Nay R
    6. Poeth K
    7. Jenness J
    8. Ernst M
    9. Grillon C
    10. Pine DS

    (2008) Fear conditioning in adolescents with anxiety disorders: results from a novel experimental paradigm

    Journal of the American Academy of Child and Adolescent Psychiatry 47:94–102.

    https://doi.org/10.1097/chi.0b01e31815a5f01

    • PubMed
    • Google Scholar
60. 1. Lau JY
    2. Britton JC
    3. Nelson EE
    4. Angold A
    5. Ernst M
    6. Goldwin M
    7. Grillon C
    8. Leibenluft E
    9. Lissek S
    10. Norcross M
    11. Shiffrin N
    12. Pine DS

    (2011) Distinct neural signatures of threat learning in adolescents and adults

    PNAS 108:4500–4505.

    https://doi.org/10.1073/pnas.1005494108

    • Google Scholar
61. 1. LeDoux JE

    (2000) Emotion circuits in the brain

    Annual Review of Neuroscience 23:155–184.

    https://doi.org/10.1146/annurev.neuro.23.1.155

    • PubMed
    • Google Scholar
62. 1. LeDoux JE

    (2014) Coming to terms with fear

    PNAS 111:2871–2878.

    https://doi.org/10.1073/pnas.1400335111

    • Google Scholar
63. 1. Lee MD
    2. Newell BR

    (2011)

    Using hierarchical Bayesian methods to examine the tools of decision-making

    Judgment and Decision Making 6:832–842.

    • Google Scholar
64. 1. Lee D
    2. Seo H
    3. Jung MW

    (2012) Neural basis of reinforcement learning and decision making

    Annual Review of Neuroscience 35:287–308.

    https://doi.org/10.1146/annurev-neuro-062111-150512

    • PubMed
    • Google Scholar
65. 1. Li J
    2. Schiller D
    3. Schoenbaum G
    4. Phelps EA
    5. Daw ND

    (2011) Differential roles of human striatum and amygdala in associative learning

    Nature Neuroscience 14:1250–1252.

    https://doi.org/10.1038/nn.2904

    • PubMed
    • Google Scholar
66. 1. Li SSY
    2. McNally GP

    (2014) The conditions that promote fear learning: prediction error and Pavlovian fear conditioning

    Neurobiology of Learning and Memory 108:14–21.

    https://doi.org/10.1016/j.nlm.2013.05.002

    • PubMed
    • Google Scholar
67. 1. Likhtik E
    2. Paz R

    (2015) Amygdala-prefrontal interactions in (mal)adaptive learning

    Trends in Neurosciences 38:158–166.

    https://doi.org/10.1016/j.tins.2014.12.007

    • PubMed
    • Google Scholar
68. 1. Lissek S

    (2005) Classical fear conditioning in the anxiety disorders: a meta-analysis

    Behaviour Research and Therapy 43:e7.

    https://doi.org/10.1016/j.brat.2004.10.007

    • Google Scholar
69. 1. Lissek S
    2. Rabin SJ
    3. McDowell DJ
    4. Dvir S
    5. Bradford DE
    6. Geraci M
    7. Pine DS
    8. Grillon C

    (2009) Impaired discriminative fear-conditioning resulting from elevated fear responding to learned safety cues among individuals with panic disorder

    Behaviour Research and Therapy 47:111–118.

    https://doi.org/10.1016/j.brat.2008.10.017

    • PubMed
    • Google Scholar
70. 1. Lissek S
    2. Kaczkurkin AN
    3. Rabin S
    4. Geraci M
    5. Pine DS
    6. Grillon C

    (2014) Generalized anxiety disorder is associated with overgeneralization of classically conditioned fear

    Biological Psychiatry 75:909–915.

    https://doi.org/10.1016/j.biopsych.2013.07.025

    • PubMed
    • Google Scholar
71. 1. Lonsdorf TB
    2. Menz MM
    3. Andreatta M
    4. Fullana MA
    5. Golkar A
    6. Haaker J
    7. Heitland I
    8. Hermann A
    9. Kuhn M
    10. Kruse O
    11. Meir Drexler S
    12. Meulders A
    13. Nees F
    14. Pittig A
    15. Richter J
    16. Römer S
    17. Shiban Y
    18. Schmitz A
    19. Straube B
    20. Vervliet B
    21. Wendt J
    22. Baas JMP
    23. Merz CJ

    (2017) Don’t fear “fear conditioning”: Methodological considerations for the design and analysis of studies on human fear acquisition, extinction, and return of fear

    Neuroscience and Biobehavioral Reviews 77:247–285.

    https://doi.org/10.1016/j.neubiorev.2017.02.026

    • PubMed
    • Google Scholar
72. 1. Lonsdorf TB
    2. Merz CJ

    (2017) More than just noise: Inter-individual differences in fear acquisition, extinction and return of fear in humans - Biological, experiential, temperamental factors, and methodological pitfalls

    Neuroscience and Biobehavioral Reviews 80:703–728.

    https://doi.org/10.1016/j.neubiorev.2017.07.007

    • PubMed
    • Google Scholar
73. 1. Lonsdorf TB
    2. Klingelhöfer-Jens M
    3. Andreatta M
    4. Beckers T
    5. Chalkia A
    6. Gerlicher A
    7. Jentsch VL
    8. Meir Drexler S
    9. Mertens G
    10. Richter J
    11. Sjouwerman R
    12. Wendt J
    13. Merz CJ

    (2019) Navigating the garden of forking paths for data exclusions in fear conditioning research

    eLife 8:e52465.

    https://doi.org/10.7554/eLife.52465

    • PubMed
    • Google Scholar
74. 1. Maren S

    (2001) Neurobiology of Pavlovian fear conditioning

    Annual Review of Neuroscience 24:897–931.

    https://doi.org/10.1146/annurev.neuro.24.1.897

    • PubMed
    • Google Scholar
75. 1. Marin M-F
    2. Zsido RG
    3. Song H
    4. Lasko NB
    5. Killgore WDS
    6. Rauch SL
    7. Simon NM
    8. Milad MR

    (2017) Skin Conductance Responses and Neural Activations During Fear Conditioning and Extinction Recall Across Anxiety Disorders

    JAMA Psychiatry 74:622–631.

    https://doi.org/10.1001/jamapsychiatry.2017.0329

    • PubMed
    • Google Scholar
76. 1. Marin M-F
    2. Barbey F
    3. Rosenbaum BL
    4. Hammoud MZ
    5. Orr SP
    6. Milad MR

    (2020) Absence of conditioned responding in humans: A bad measure or individual differences?

    Psychophysiology 57:e13350.

    https://doi.org/10.1111/psyp.13350

    • PubMed
    • Google Scholar
77. 1. Mechias M-L
    2. Etkin A
    3. Kalisch R

    (2010) A meta-analysis of instructed fear studies: implications for conscious appraisal of threat

    NeuroImage 49:1760–1768.

    https://doi.org/10.1016/j.neuroimage.2009.09.040

    • PubMed
    • Google Scholar
78. 1. Michalska KJ
    2. Machlin L
    3. Moroney E
    4. Lowet DS
    5. Hettema JM
    6. Roberson-Nay R
    7. Averbeck BB
    8. Brotman MA
    9. Nelson EE
    10. Leibenluft E
    11. Pine DS

    (2017) Anxiety symptoms and children’s eye gaze during fear learning

    Journal of Child Psychology and Psychiatry, and Allied Disciplines 58:1276–1286.

    https://doi.org/10.1111/jcpp.12749

    • PubMed
    • Google Scholar
79. 1. Michalska KJ
    2. Feldman JS
    3. Ivie EJ
    4. Shechner T
    5. Sequeira S
    6. Averbeck B
    7. Degnan KA
    8. Chronis-Tuscano A
    9. Leibenluft E
    10. Fox NA
    11. Pine DS

    (2019) Early-childhood social reticence predicts SCR-BOLD coupling during fear extinction recall in preadolescent youth

    Developmental Cognitive Neuroscience 36:100605.

    https://doi.org/10.1016/j.dcn.2018.12.003

    • PubMed
    • Google Scholar
80. 1. Milad MR
    2. Quirk GJ

    (2012) Fear extinction as a model for translational neuroscience: ten years of progress

    Annual Review of Psychology 63:129–151.

    https://doi.org/10.1146/annurev.psych.121208.131631

    • PubMed
    • Google Scholar
81. 1. Mineka S
    2. Oehlberg K

    (2008) The relevance of recent developments in classical conditioning to understanding the etiology and maintenance of anxiety disorders

    Acta Psychologica 127:567–580.

    https://doi.org/10.1016/j.actpsy.2007.11.007

    • PubMed
    • Google Scholar
82. 1. Ney LJ
    2. Wade M
    3. Reynolds A
    4. Zuj DV
    5. Dymond S
    6. Matthews A
    7. Felmingham KL

    (2018) Critical evaluation of current data analysis strategies for psychophysiological measures of fear conditioning and extinction in humans

    International Journal of Psychophysiology 134:95–107.

    https://doi.org/10.1016/j.ijpsycho.2018.10.010

    • PubMed
    • Google Scholar
83. 1. Nordenskjöld R
    2. Malmberg F
    3. Larsson E-M
    4. Simmons A
    5. Ahlström H
    6. Johansson L
    7. Kullberg J

    (2015) Intracranial volume normalization methods: Considerations when investigating gender differences in regional brain volume

    Psychiatry Research 231:227–235.

    https://doi.org/10.1016/j.pscychresns.2014.11.011

    • Google Scholar
84. 1. Ojala KE
    2. Bach DR

    (2020) Measuring learning in human classical threat conditioning: Translational, cognitive and methodological considerations

    Neuroscience and Biobehavioral Reviews 114:96–112.

    https://doi.org/10.1016/j.neubiorev.2020.04.019

    • PubMed
    • Google Scholar
85. 1. Orr SP
    2. Metzger LJ
    3. Lasko NB
    4. Macklin ML
    5. Peri T
    6. Pitman RK

    (2000) De novo conditioning in trauma-exposed individuals with and without posttraumatic stress disorder

    Journal of Abnormal Psychology 109:290–298.

    https://doi.org/10.1037/0021-843x.109.2.290

    • PubMed
    • Google Scholar
86. 1. Papalini S
    2. Beckers T
    3. Vervliet B

    (2020) Dopamine: from prediction error to psychotherapy

    Translational Psychiatry 10:164.

    https://doi.org/10.1038/s41398-020-0814-x

    • PubMed
    • Google Scholar
87. 1. Pattwell SS
    2. Duhoux S
    3. Hartley CA
    4. Johnson DC
    5. Jing D
    6. Elliott MD
    7. Ruberry EJ
    8. Powers A
    9. Mehta N
    10. Yang RR
    11. Soliman F
    12. Glatt CE
    13. Casey BJ
    14. Ninan I
    15. Lee FS

    (2012) Altered fear learning across development in both mouse and human

    PNAS 109:16318–16323.

    https://doi.org/10.1073/pnas.1206834109

    • PubMed
    • Google Scholar
88. 1. Piray P
    2. Dezfouli A
    3. Heskes T
    4. Frank MJ
    5. Daw ND

    (2019) Hierarchical Bayesian inference for concurrent model fitting and comparison for group studies

    PLOS Computational Biology 15:e1007043.

    https://doi.org/10.1371/journal.pcbi.1007043

    • PubMed
    • Google Scholar
89. 1. Pittig A
    2. van den Berg L
    3. Vervliet B

    (2016) The key role of extinction learning in anxiety disorders: behavioral strategies to enhance exposure-based treatments

    Current Opinion in Psychiatry 29:39–47.

    https://doi.org/10.1097/YCO.0000000000000220

    • PubMed
    • Google Scholar
90. 1. Priestley MB

    (1981) The Advanced Theory of Statistics

    Journal of the Royal Statistical Society 375:e814.

    https://doi.org/10.2307/2981814

    • Google Scholar
91. 1. Ray MH
    2. Russ AN
    3. Walker RA
    4. McDannald MA

    (2020) The Nucleus Accumbens Core is Necessary to Scale Fear to Degree of Threat

    The Journal of Neuroscience 40:4750–4760.

    https://doi.org/10.1523/JNEUROSCI.0299-20.2020

    • PubMed
    • Google Scholar
92. Book

    1. Rescorla RA
    2. Wagner AR

    (1972)

    In Classical Conditioning II (Ed A. H. Black & W. F. Prokasy) 64-99

    New York: Appleton-Century-Crofts.

    • Google Scholar
93. 1. Ruigrok ANV

    (2014) A meta-analysis of sex differences in human brain structure

    Neuroscience and Biobehavioral Reviews 39:34.

    https://doi.org/10.1016/j.neubiorev.2013.12.004

    • Google Scholar
94. 1. Ryan KM
    2. Zimmer-Gembeck MJ
    3. Neumann DL
    4. Waters AM

    (2019) The need for standards in the design of differential fear conditioning and extinction experiments in youth: A systematic review and recommendations for research on anxiety

    Behaviour Research and Therapy 112:42–62.

    https://doi.org/10.1016/j.brat.2018.11.009

    • PubMed
    • Google Scholar
95. 1. Schiller D
    2. Levy I
    3. Niv Y
    4. LeDoux JE
    5. Phelps EA

    (2008) From fear to safety and back: reversal of fear in the human brain

    The Journal of Neuroscience 28:11517–11525.

    https://doi.org/10.1523/JNEUROSCI.2265-08.2008

    • PubMed
    • Google Scholar
96. 1. Schultz W

    (2013) Updating dopamine reward signals

    Current Opinion in Neurobiology 23:229–238.

    https://doi.org/10.1016/j.conb.2012.11.012

    • PubMed
    • Google Scholar
97. 1. Schwarz GE

    (1978) Estimating the Dimension of a Model

    The Annals of Statistics 6:461–464.

    https://doi.org/10.1214/aos/1176344136

    • Google Scholar
98. 1. Shechner T
    2. Britton JC
    3. Ronkin EG
    4. Jarcho JM
    5. Mash JA
    6. Michalska KJ
    7. Leibenluft E
    8. Pine DS

    (2015) Fear conditioning and extinction in anxious and nonanxious youth and adults: examining a novel developmentally appropriate fear-conditioning task

    Depression and Anxiety 32:277–288.

    https://doi.org/10.1002/da.22318

    • PubMed
    • Google Scholar
99. 1. Smith SM
    2. Nichols TE

    (2009) Threshold-free cluster enhancement: addressing problems of smoothing, threshold dependence and localisation in cluster inference

    NeuroImage 44:83–98.

    https://doi.org/10.1016/j.neuroimage.2008.03.061

    • PubMed
    • Google Scholar
100. 1. Spielberger CD
     2. Gorsuch RL
     3. Lushene RE

     (1970) Manual for the State-Trait Anxiety Inventory

     Consulting Psychologists Press 1:e7.

     https://doi.org/10.1007/s11414-016-9526-7

     • Google Scholar
101. Book

     1. Talairach J
     2. Tournoux P

     (1988)

     Co-Planar Stereotaxic Atlas of the Human Brain: 3-Dimensional Proportional System: An Approach to Cerebral Imaging

     Georg Thieme.

     • PubMed
     • Google Scholar
102. 1. Thorpe S
     2. Salkovskis PM
     3. Dittner A

     (2008) Claustrophobia in MRI: the role of cognitions

     Magnetic Resonance Imaging 26:1081–1088.

     https://doi.org/10.1016/j.mri.2008.01.022

     • PubMed
     • Google Scholar
103. 1. Tottenham N
     2. Tanaka JW
     3. Leon AC
     4. McCarry T
     5. Nurse M
     6. Hare TA
     7. Marcus DJ
     8. Westerlund A
     9. Casey BJ
     10. Nelson C

     (2009) The NimStim set of facial expressions: judgments from untrained research participants

     Psychiatry Research 168:242–249.

     https://doi.org/10.1016/j.psychres.2008.05.006

     • PubMed
     • Google Scholar
104. 1. Tseng W-L
     2. Abend R
     3. Gold AL
     4. Brotman MA

     (2021) Neural correlates of extinguished threat recall underlying the commonality between pediatric anxiety and irritability

     Journal of Affective Disorders 295:920–929.

     https://doi.org/10.1016/j.jad.2021.08.117

     • PubMed
     • Google Scholar
105. 1. Tzovara A
     2. Korn CW
     3. Bach DR

     (2018) Human Pavlovian fear conditioning conforms to probabilistic learning

     PLOS Computational Biology 14:e1006243.

     https://doi.org/10.1371/journal.pcbi.1006243

     • Google Scholar
106. 1. van Geen C
     2. Gerraty RT

     (2021) Hierarchical Bayesian models of reinforcement learning: Introduction and comparison to alternative methods

     Journal of Mathematical Psychology 105:102602.

     https://doi.org/10.1016/j.jmp.2021.102602

     • Google Scholar
107. 1. Vervliet B
     2. Baeyens F
     3. Van den Bergh O
     4. Hermans D

     (2013) Extinction, generalization, and return of fear: A critical review of renewal research in humans

     Biological Psychology 92:51–58.

     https://doi.org/10.1016/j.biopsycho.2012.01.006

     • PubMed
     • Google Scholar
108. Book

     1. Wechsler D

     (1999)

     Wechsler Abbreviated Scale of Intelligence

     The Psychological Corporation.

     • Google Scholar
109. 1. Winkler AM
     2. Ridgway GR
     3. Webster MA
     4. Smith SM
     5. Nichols TE

     (2014) Permutation inference for the general linear model

     NeuroImage 92:381–397.

     https://doi.org/10.1016/j.neuroimage.2014.01.060

     • PubMed
     • Google Scholar
110. 1. Winkler AM
     2. Ridgway GR
     3. Douaud G
     4. Nichols TE
     5. Smith SM

     (2016) Faster permutation inference in brain imaging

     NeuroImage 141:502–516.

     https://doi.org/10.1016/j.neuroimage.2016.05.068

     • Google Scholar
111. 1. Wise T
     2. Michely J
     3. Dayan P
     4. Dolan RJ

     (2019) A computational account of threat-related attentional bias

     PLOS Computational Biology 15:e1007341.

     https://doi.org/10.1371/journal.pcbi.1007341

     • PubMed
     • Google Scholar
112. 1. Wise T
     2. Dolan RJ

     (2020) Associations between aversive learning processes and transdiagnostic psychiatric symptoms in a general population sample

     Nature Communications 11:4179.

     https://doi.org/10.1038/s41467-020-17977-w

     • PubMed
     • Google Scholar
113. 1. Zeileis A
     2. Grothendieck G

     (2005) zoo: S3 infrastructure for regular and irregular time series

     Journal of Statistical Software 14:i06.

     https://doi.org/10.18637/jss.v014.i06

     • Google Scholar
114. 1. Zhang Y
     2. Brady M
     3. Smith S

     (2001) Segmentation of brain MR images through a hidden Markov random field model and the expectation-maximization algorithm

     IEEE Transactions on Medical Imaging 20:45–57.

     https://doi.org/10.1109/42.906424

     • PubMed
     • Google Scholar
115. 1. Zhang SY
     2. Mano H
     3. Ganesh G
     4. Robbins T
     5. Seymour B

     (2016) Dissociable Learning Processes Underlie Human Pain Conditioning

     Current Biology 26:52–58.

     https://doi.org/10.1016/j.cub.2015.10.066

     • PubMed
     • Google Scholar

Decision letter after peer review:

Thank you for submitting your article "Computational Modeling of Threat Learning Reveals Links with Anxiety and Neuroanatomy" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, and the evaluation has been overseen by Drs. Shackman (Reviewing Editor) and Michael Frank (Senior Editor).

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

Summary:

The authors present an investigation into the role of threat learning processes in symptoms of anxiety across a broad sample of subjects with and without clinical anxiety, across multiple age groups. The study uses a simple fear conditioning paradigm in which subjects first undergo an acquisition procedure to learn associations between visual CSs and loud noise USs, before undergoing an extinction procedure where the CS+ is presented without the US. Reinforcement learning models are fit to skin conductance responses, and parameters from these models are used to predict symptoms of anxiety. This demonstrated weaker safety learning (i.e. learning that the CS– did not predict threat) in those who were more anxious, and slower extinction in more anxious individuals, a result that was moderated by gray matter volume in a number of subcortical brain regions.

The reviewers expressed enthusiasm for the manuscript:

– This is an important area, and the results will be of interest to readers across the field of computational psychiatry, and mental health research more widely.

– This work has the potential to be an important study in the field of computational psychiatry. Investigations of threat learning processes in individuals with clinical anxiety are lacking, despite their clear relevance as a mechanism underlying symptoms of these disorders.

– [noted by all reviewers] The sample is large and includes subjects with clinically significant levels of anxiety who were not receiving medication.

– The modelling approach is carefully considered, and the authors have clearly been thorough in testing different model variants.

– The inclusion of brain volume data provides some interesting insights into how gray matter volume relates to these core learning processes.

– The authors explain the added value brought by the computational modelling approach well, and it clearly provides some interesting insights into the role of threat learning that would not be seen in more traditional analyses (e.g. prior work by this team).

– All published work on computational models of threat learning in SCR used comparably smaller samples with more limited age range (Li et al. 2011, Zhang et al. 2016, Tzovara et al. 2018, Homan et al. 2019). Hence, this is an important data set that could offer substantial insights into threat learning.

– This paper is of potential interest to researchers within the fields of clinical psychology, clinical neuroscience and computational psychiatry.

– Over recent years, both the use of dimensional measures of symptoms and adoption of computational modelling has advanced our understanding of psychopathology–related differences in cognitive and brain function. This has encompassed both original empirical studies and application of computational modelling to previously published datasets. Both these approaches have potential value and different advantages and disadvantages. The current study is an example of the latter approach.

– The consideration of nine alternate models in the modelling of SCR responses to the CS+ during acquisition is a strength of the paper.

– The application of computational modelling techniques to such a large differential fear conditioning dataset, with data from both healthy adult and child participants and unmedicated patients seeking treatment for anxiety is highly appealing.

Nevertheless, all 3 reviewers expressed some significant concerns, mostly centered on the analytic framework and other aspects of the approach. There was a consensus among us that the necessary revisions will entail considerable effort on the part of the authors.

– Modeling Approach

Concern: The modelling approach could be validated more robustly. For example, model and parameter recovery analyses are not presented, and so it is unclear how much confidence can be placed in the winning model and the parameter values derived from this model.

Recommendation: Include a more thorough validation of the modelling approach, adding model recovery and parameter recovery analyses. This would provide confidence in the conclusions drawn from the modelling by demonstrating that it is possible to recover the winning model and its parameters accurately.

Recommendation: It would be helpful to also show the BICs for different models to give the reader a clearer idea of how model fit differed across the different models.

Recommendation: It would be helpful if the authors could conduct Bayesian model comparison (see Stephan et al., Neuroimage, 2009) and report the population–level estimate of the proportion of participants best fit by each model. Given the age by anxiety analyses, it would also be helpful if they could report the best–fitting model for high and low anxious children (as assessed using the SCARED) and high and low anxious adults (as assessed using the STAI). Does the same model 'win' for both child and adult participants divided into low and high anxious subgroups?

Concern: Description of model estimation and comparison lacks detail. Model parameters are fit with fminsearch (using which starting values? random? grid search? just one starting value would hardly be sufficient. any parameter constraints?) using OLS minimisation. How is BIC then computed? How are the models compared – the methods mention "per participant and across the sample"; do you assume a fixed or random model structure? What is the purpose of a t–test (p. 13) in the model comparison?

Recommendation: Provide modeling details.

Concern: The model space of the acquisition session does not include any of the "best" threat learning models from previously published work (see below for some specific examples). This precludes a direct comparison; more importantly it may also miss out on the best model in this sample. First, previous work has simultaneously modelled CS– and CS+ responses. This is appropriate (for all sessions), since a priori the agent does not know which is the "safety" cue, and will presumably use the same mechanisms to update value for both cues. The authors state that standard models cannot account for the large increase from first CS– trial (before first reinforced CS+) to next trial; while this is correct, the authors have a model to account for that, and more trivially one may just leave out the first CS– response. Secondly, the observation function (which is only specified in the legend of table 1 – this should appear in the text) maps SCR onto the value terms of all models. All previous SCR modelling work has found SCR is more closely related to associability or uncertainty, or a mixture of these with value.

Concern: The models considered are broadly suitable but comparison of model fit across both SCRs to the CS+ and CS– during acquisition and SCRs to the CS+ and CS– in extinction is missing, which is a limitation.

Recommendation: Include prior models in the model space. Models should simultaneously include CS+ and CS– responses. To this end, they may leave out responses before the first reinforced CS+ trial. Model space should include additional variants of models 9 that account for a mapping onto associability (as in Li et al. 2011, Zhang et al. 2016) or combined associability and value (as in Homan et al. 2019) and a variant of model 3 that accounts for a mixture of uncertainty and value (as in Tzovara et al. 2018).

Related Concern from Another Reviewer: The authors do not consider alternate models for the SCRs to the CS– during acquisition or for the SCRs to the CS+ during extinction and they do not model the SCRs to the CS– during extinction at all.

Recommendations: It would be helpful to know which model best fits all the acquisition data (to both the CS+ and CS–). Does the hybrid model still perform less well than the non–hybrid models? And does model 8 still outperform model 7? (in models 8 and 9, as in model 7, the CS– value could also be modelled as updating after each UCS delivery with model 8 now having 6 free parameters and model 9 4 free parameters.) Similarly, it would be useful to have both CS– and CS+ trials during extinction modelled. Given the raw data indicates the response to both the CS+ and CS– drop off during extinction, understanding if decreased responding to the CS+ during extinction represents active specific updating of the CS+ UCS contingency or a non–specific decreases in expectancy of the UCS after both the CS+ and CS– is clearly important.

– Multiple Comparisons [Multiple Reviewers]

Concern – There are many post–hoc tests on association of model parameters with brain results and participant traits. It is unclear how this was planned and performed – how did the authors correct for multiple comparison across parameters, traits, and experiment phases?

Recommendation: provide appropriate, principled correction for multiple comparisons to mitigate familywise α inflation

Recommendation: Please state how many tests were conducted and what correction was applied for example how many brain regions were investigated, was grey matter volume in each of these related to each of the five learning rate parameters (across acquisition, generalisation and extinction), and were the moderating effects of anxiety, age and anxiety by age considered in each case?

Recommendation: please be clear in defining what "counts" as a family of analyses

– Specificity of key results

Concern: The relationship between anxiety and safety learning (as indexed by the habituation parameter to the CS– during acquisition) is one of the main findings.

Concern: The novelty of this study depends both on the strength of the computational analysis and on the extent to which the results provide new insights into the relationship between child and adult anxiety and differential fear conditioning relative to the non–computational analyses reported in the authors' prior publication on this dataset (Abdend et al., 2020). The relationship between anxiety and safety learning (as indexed by the habituation parameter to the CS– during acquisition) could be one such novel insight – however here a direct comparison of the relationship between anxiety and this parameter versus the habituation parameter to the CS+ is needed to ensure this isn't simply a non–specific habituation effect. The same issue holds for the finding of a relationship between anxiety and rate of extinction to the CS+ – the authors need to show this holds over and above any relationship between anxiety and drop off in responsivity to the CS–.

Recommendation: Here, a comparison of the relative strength of relationship between anxiety and participants' habituation parameter values for the CS– versus that for the CS+ is needed to ensure this isn't simply a non–specific habituation effect. Similarly, the authors need to directly compare the relationship between anxiety and drop off in SCRs to the CS+ during extinction against the relationship between anxiety and drop off in SCRs to the CS– during extinction to be able to make any claim specific to extinction to the CS+. This also applies to the moderating effects of brain structure on this relationship.

– SCR Approach

Concern: SCR data are likely not optimal for model fitting at the subject level; SCR is noisy, and the task involves very few trials. This may result in poorly estimated parameters; however it is not clear from the presented analyses how much of a problem this is.

Recommendation – Given the noisiness of SCR data and the limited number of trials available, parameter estimation could likely be improved substantially by using hierarchical Bayesian parameter estimation methods.

Concern: The quantification of the conditioned response is in line with some previous work, but in fact there is a range of possible analysis choices, summarised in Lonsdorf et al. 2017.

Recommendation: Given this heterogeneity, it would make sense to use the most sensitive method. Bach et al. 2020 Behaviour Research and Therapy and Ojala and Bach 2020 Neurosci Biobeh Rev discuss the most sensitive method to quantify threat conditioning from skin conductance data, and Bach et al. 2020 Nat Hum Behav provide background.

– Missing SCR Data

[Multiple reviewers] Concern: More than 1/3 of the sample are excluded due to "excessive missing data", defined as >50% missing trials.

Recommendations: It is unclear why these data were missing. Please address the explanation and please discuss the potential consequences for inference. Do excluded subjects systematically differ from those included? It would be helpful if the authors could include a table showing a break–down of the characteristics of included and excluded participants by age, gender, diagnostic status and anxiety levels and if they could address in the discussion any associated limitations in particular in relation to the age by anxiety analyses reported.

Concern: It is unclear why missing data are inter/extrapolated?

Recommendation: Use MSE. The model fitting procedure should be happy with missing data points if MSE (rather than SSE) is used as objective function. This would be more appropriate than giving the model a data value that does not exist. Censor, rather than interpolate.

– Disparate Dimensional Measures of Anxiety [Multiple Reviewers]

Concern: The combination of multiple anxiety measures (SCARED, STAI) into a single anxiety measure may not be valid, as it's reasonable to assume these are measuring subtly different constructs (aside from anything else, the SCARED measures recent symptoms while the STAI is a trait measure).

Concern: The main analyses appear to combine measures of anxiety severity across age groups as the primary variable of interest (using the SCARED for youth and the STAI for adults), and in one analysis find an interaction between age and anxiety that is driven by the younger age group. This raises some concerns as these measures are not necessarily measuring the same construct, and may limit the interpretability of these results.

Concern: Children and adults show opposing direction relationships between anxiety and fear generalisation. It is difficult to interpret the anxiety by age results given different measures were used to assess anxiety in children and in adults. This could potentially account for the different relationship between anxiety and fear generalisation in under and over 18s (see Figure 3B). I am not convinced that the authors can simply merge scores from the SCARED (which averages symptom scores based on child and adult reports) and STAI (an adult self–report questionnaire) as if they came from the same measure.

Recommendation: Absent some sort of direct quantitative harmonization (as in the GWAS literature for N/NE) or additional evidence to motivate the Z–score approach, it may not be appropriate to combine or directly compare the younger (SCARED) and older (STAI) samples (i.e. more appropriate to keep them distinct and examine them separately)

– Potential nonlinear effects of age

Concern: Only linear effects of age are modelled.

Recommendation: explore potential nonlinear effects (given that many developmental effects are nonlinear).

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

Thank you for submitting your revised manuscript – "Computational Modeling of Threat Learning Reveals Links with Anxiety and Neuroanatomy in Humans" – for consideration by eLife. Your revised article has been evaluated by Drs. Frank (Senior Editor) and Shackman (Reviewing/Handling Editor) with guidance and consultation from 2 of the 3 original Reviewers.

The manuscript has improved, with one Reviewer emphasizing that "the authors have done a good job of addressing many of the suggestions and the manuscript is definitely substantially improved as a result."

Nevertheless, there was agreement that the paper would benefit from some additional modeling.

One Reviewer wrote that, Abend et al. have addressed some, but not all of my previous concerns. The model space of the acquisition session does not include any of the "best" threat learning models from previous publications. This precludes a direct comparison; more importantly it may also miss out on the best model in this dataset.

The authors now model CS– and CS+ responses simultaneously. Yet they still do not compare their models to any of the those used in previous studies of conditioned SCR (Li et al. 2011, Zhang et al. 2016, Tzovara et al. 2018, Homan et al. 2019). They claim that their model 9 is the hybrid model used in three of these studies, but actually it is not, because the update works differently (presently they update only upon US delivery, and both CS+ and CS– from PE in CS+ trials, where as previous RW and hybrid models used the standard update on every trial, on signed PE, and with no cross–update). They qualitatively argue why the associability term in previous work does not fit their data, but in fact they should just quantitatively compare models that map value and/or associability onto SCR. Finally, they argue that Tzovara's model can only update CS+ or CS– expectation with one parameter, and modelling both CS types within the same model would lead to parameter inflaton, but in fact Tzovara's winning learning model is entirely parameter–free (only has parameters for the observation function), so this is inaccurate.

The authors need to implement these pre–existing models and show quantitatively that they do not fit the data, rather than on rely on qualitative and methodological arguments for why this is unnecessary. Based on their rebuttal, it seems that they had already implemented the models, so it should be relatively straightforward to make the necessary direct ("head–to–head") comparisons.

Along related lines, the other Reviewer noted that, [page 11] The authors argue that to adapt the Tzovara et al. model, a 5th parameter would need to be added, and that this would make it too complex. It would still be worth testing this properly [per the other reviewer's comments], and it is possible (although probably unlikely based on the provided evidence) that the model could fit well enough that it still wins despite the additional complexity–related penalisation.

The Reviewers and Editors also discussed hierarchical Bayesian approaches, with one reviewer noting that, [page 19] Reviewer suggestions related to hierarchical Bayesian parameter estimation do not seem to have been addressed in the revision. For clarity, this refers to the method of estimating reinforcement learning model parameters (see van Geen and Gerraty, 2021), rather than the SCR quantification, and there's a strong possibility it would improve parameter estimation and hence sensitivity. I appreciate that this could be represent a substantial re–analysis.

During the consultation session, the 2 Reviewers agreed that the hierarchical Bayesian approach would be nice, but is not strictly necessary for the revision, with one noting that, "I wouldn't want to force them to redo all the model fitting with a hierarchical Bayesian approach, it would likely provide more accurate estimates, but it would be a fairly big undertaking. Some acknowledgement that other model–fitting procedures exist and could be more appropriate would be nice, however (this probably also applies to their SCR quantification methods)".

And the other writing that, "I agree. This model fitting approach is a major step and may not work for some of the models."

In sum, the Bayesian approach should –– at minimum –– be noted in the Discussion.

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled "Computational Modeling of Threat Learning Reveals Links with Anxiety and Neuroanatomy in Humans" for further consideration by eLife. Your revised article has been evaluated by Drs. Shackman (Reviewing Editor) and Frank (Senior Editor).

We are very pleased to accept your paper for publication pending receipt of a revision that adequately addresses the remaining suggestions from Reviewer #2. Congratulations in advance!

Reviewer #2:

This revision of the paper by Abend et al. is significantly stronger by including a direct comparison of the current model space with previous "best" models. There are two substantive and a few presentation points remaining.

1. The Tzovara et al. model is implemented differently from the original paper. The authors use separate regression parameters for CS+ and CS– trials for the Tzovara et al. model (i.e. 4 parameters) and it is not clear to me why – Tzovara et al. themselves had used only 2 regression parameters across these two trials types. With 2 instead of 4 parameters, the model might fare better (even though the fit/RSS will of course become worse). Could the authors include the original 2–parameter Tzovara et al. model?

2. Perhaps relatedly, I've been wondering about the initial values of the various terms in the update equations. Can they be presented in the tables? I'm wondering, for example, whether they used the same values for α_0 and β_0 in the β distribution of the Tzovara et al. model, but other readers might wonder about other models. Of course, it can make a big difference how a model is "initialised".

3. The authors make an effort to explain the difference between models 1–9, and models 10–14, by using different symbols on the LHS of the equations. Perhaps it could be made more prominent that models 1–9 predict CS–related SCR from US–related SCR and thus need no regression to scale the predictions to the data, whereas models 10–14 predict CS–related SCR from an arbitrary reinforcement variable.

4. I'm still unsure about using t–tests to compare BICs across a group. I have never seen this. What is the statistical foundation for this approach? I would suggest adding the RFX–model comparison by Stephan et al. as had been suggested by two reviewers in the context of the initial manuscript version. It is implemented in a matlab function (spm_BMS), requires only the matrix of BIC values, and takes a few seconds to run.

Summary:

[…]

– Modeling Approach

Concern: The modelling approach could be validated more robustly. For example, model and parameter recovery analyses are not presented, and so it is unclear how much confidence can be placed in the winning model and the parameter values derived from this model.

Recommendation: Include a more thorough validation of the modelling approach, adding model recovery and parameter recovery analyses. This would provide confidence in the conclusions drawn from the modelling by demonstrating that it is possible to recover the winning model and its parameters accurately.

Thank you for this suggestion. We have now carried out parameter recovery and model recovery, as specified below.

Parameter recovery:

To carry out parameter recovery, we pulled the parameters for each model from the subjects to whom it was best fit and used those parameters to generate synthetic data from the model. We then fit the original model to the synthetic data and compared the corresponding actual and “recovered” parameters using a correlation coefficient. In general, the parameters were well recovered, with the exception of model 9 (RW-Pearce-Hall hybrid). We believe that our task design, probing our specific processes of interest, is not well-suited to drive the flexible learning component of the RW-PH model. Indeed, most prior work that finds this model to work well has included substantially more trials and contingency reversals; here, we focus only on initial, rapid acquisition and extinction as opposed to change in threat value over many trials. The main text now directs the reader (p. 15) to the full description of this rationale in Appendix 1 (pp. 57-58): “We carried out parameter recovery by using the parameters for each model from the participants to whom it was best fit and used those parameters to generate synthetic data from the model. For each model, we then fit the original model to the synthetic data and compared the corresponding actual and “recovered” parameters using correlation coefficients; see Appendix 1-table 2. In general, parameters were recovered well, with the exception of model 9 (RW-PH hybrid). This may suggest that this model is not well-suited to capture rapid conditioning. Indeed, most prior work that finds this model to work well has included substantially more trials and contingency reversals; here, we focus only on initial, rapid acquisition as opposed to change in threat value over many trials. Thus, when modeling contingencies and their change over longer durations, the hybrid model may provide an optimal fit, whereas initial acquisition of contingencies might reflect a specific case in which other models perform better.”

The correlation between the actual parameters and recovered parameters for each model was calculated and placed in a table (Appendix 1-table 2):

Model recovery:

We now note in Appendix 1 (p. 59): “Model recovery was performed by simulating data from all the models using the parameters from the participants best fit by each model. However, due to the unequal distribution of participants best fit by each model, we calculated the mean and standard deviation of the parameters from the best fit participants, and then sampled from the corresponding Gaussian distributions to create a total of 100 sets of parameters for each model. These parameters were then used to generate synthetic datasets. That simulated data was then fit with all the models to determine the extent to which the model selection approach would identify the model used to generate the data. Note that in some cases, for example in habituation models, we could sample very small parameter values for the habituation term, which would lead to selection of a simpler model.”

We summarized this with a confusion matrix (added to Appendix 1 as Figure 2—figure supplement 2):

The models were recovered reasonably well, with the exception of model 6 and model 9. It is important to note that models 1-8 are related variants of a model that captures the process of interest, i.e., they are rather similar (as we now note in p. 59). As such, the models are generally comparable and the accuracy of parameter and model recovery lower and more similar relative to model comparison between very different models. Note that we did not intentionally bias our selection of parameters or data to get full coverage of the parameter or data space best represented by these models, which would separate the differences and seemingly improve model recovery. To do this, one would need a process that exploits what each model is capturing differently than all other models. Thus, we chose the subjects best fit by the models as a reasonable starting point for parameter and model recovery procedures. In this context, models 3 and 7 recovered relatively better, while model 9 (RW-PH hybrid) and model 6 did not recover as well. This is now described in full in Appendix 1 (p. 59).

Recommendation: It would be helpful to also show the BICs for different models to give the reader a clearer idea of how model fit differed across the different models.

Thank you for this recommendation. In Author response image 1 we present a bar plot of the average BIC values for each of the models fit to the acquisition data to show similarity. (We performed similar analyses for extinction; those results are included in a later reply to a reviewer requesting more modeling of extinction). (Author response image 1) shows that the average BIC values for the acquisition models all fall within a range of ~10 values, reflecting the similarity of the models to each other. This suggests that the different variants capture the acquisition process comparably well. Our response to the next recommendation provides more detail on BIC comparison.

Author response image 1

Download asset Open asset

Recommendation: It would be helpful if the authors could conduct Bayesian model comparison (see Stephan et al., Neuroimage, 2009) and report the population–level estimate of the proportion of participants best fit by each model. Given the age by anxiety analyses, it would also be helpful if they could report the best–fitting model for high and low anxious children (as assessed using the SCARED) and high and low anxious adults (as assessed using the STAI). Does the same model 'win' for both child and adult participants divided into low and high anxious subgroups?

Please see our response to the previous recommendation for a figure showing mean BIC by model. In Author response image 2 we present the fraction of participants best fit by each model, as suggested.

Author response image 2

Download asset Open asset

We now note in p. 14 how we compared the models: “We then compared the distributions of BIC values generated for each model across all subjects using t-tests. First, the BIC value for each model for each participant was computed. The differences between sets of BIC values (by model, for all participants) were used to run t-tests to determine whether there were statistically significant differences in the distributions of values for each of the models. We also examined the best model for each subject. In this case the BIC values were compared, and the model with the lowest BIC value for each subject was selected (without t-test).” We then note in pp. 18-19: “To characterize model fits at a population level, we first performed a repeated-measures ANOVA on BIC values (model as fixed effect, participant as random effect, BIC value as dependent variable). We found that BICs varied across models, F(1,8)=38.98, p<0.001. We further found that models 7 and 8 performed better than all other models based on post-hoc comparisons (ps<0.05, Bonferroni-corrected). These models did not, however, statistically differ (p=0.20). Since model 7 is a simpler model, we chose it as the optimal model for our data…”With regard to proportion of models providing best fit, we do not see large stand-outs. Rather, we see that subjects are spread across models, reflecting natural variability among the subtle variants. Our goal with model fitting was to find models which captured the skin conductance response dynamics well, rather than support assertions about which model captured a particular observed behavior (as is often common in studies doing model comparison). In our original analyses, we used only simple models and found that without a habituation term in the model, the models fit poorly, and the learning rates tended to absorb aspects of the learning curves that were not due to differences in learning, because these models had large bias. The main thing that comes from this figure is that the simplest model (model 1) does not seem to capture the skin conductance data well, while the models with additional parameters can provide a more reliable estimate of skin conductance response values. Because the BIC distributions did offer a reasonable way to identify an accurate model, we based our model selection most strongly on that.

We want to emphasize that for this manuscript, we did not bin the participants into categories of anxiety (high vs. low anxiety) or age (youth vs. adults). This is because both anxiety and age are naturally continuous variables, and it is not clear that breaking them down by categories offers an advantage, either empirically or conceptually. Nevertheless, when subjects are binned categorically, we do not see a clear relationship between which model is picked and anxiety. Author response image 3 shows the fractions of participants best fit by each model, broken down by group. We do report in Appendix 1 results for the primary analyses (associations between learning parameters and anxiety, and moderation by brain structure). This way, the reader has results for both the full sample and by age group.

Author response image 3

Download asset Open asset

Concern: Description of model estimation and comparison lacks detail. Model parameters are fit with fminsearch (using which starting values? random? grid search? just one starting value would hardly be sufficient. any parameter constraints?) using OLS minimisation. How is BIC then computed? How are the models compared – the methods mention "per participant and across the sample"; do you assume a fixed or random model structure? What is the purpose of a t–test (p. 13) in the model comparison?

Recommendation: Provide modeling details.

We apologize for not clarifying. We added more information to the text to clarify this. The model parameters were fit with fminsearch. We used starting values between 0.1 and 0.8 for learning rate. The habituation parameters started at 0. This is now noted in pp. 13-14. In pilot exploration, we found that this approach gave good fits, with a reasonable amount of computation. There were no parameter constraints for the acquisition data fits, as estimates naturally fell in the range -1:+1. This is noted in pp. 13-14. For extinction, unconstrained fits with all 9 models (described in more detail later) led to extreme estimates for some participants in order to accommodate the data. When learning rates fall outside 1/-1, the RW model is unstable. As such, learning rates were constrained to -1:+1, which improved model fit. Still, ~14% of participants had estimated pinned at -1 or 1, and for four participants, models did not converge. This is now all noted in p. 15. Data for these were not used in analyses.

The parameter set associated with the maximum log-likelihood (same as the minimum negative log-likelihood) was selected out of 10 iterations across parameter starting values. We calculated the BIC using a standard Gaussian assumption on the residuals:

${\displaystyle BIC=kln\left(n\right)+nln(variance)}$ Where k is the number of parameters in the model (2 or 4), and n is the number of data points in the dataset that was fit (which corresponds to the number of trials in CS+ and CS- data for the given phase being fit). This is the BIC calculation under the assumption that model errors are i.i.d. and normally distributed (Priestley, M.B. (1981). Spectral Analysis and Time Series. Academic Press. ISBN 978-0-12-564922-3. (p 375)). This is now noted in p.14.

As noted in our response to the previous recommendation, we also added information on how models were statistically compared (p. 14; pp. 18-19).

Concern: The model space of the acquisition session does not include any of the "best" threat learning models from previously published work (see below for some specific examples). This precludes a direct comparison; more importantly it may also miss out on the best model in this sample. First, previous work has simultaneously modelled CS– and CS+ responses. This is appropriate (for all sessions), since a priori the agent does not know which is the "safety" cue, and will presumably use the same mechanisms to update value for both cues. The authors state that standard models cannot account for the large increase from first CS– trial (before first reinforced CS+) to next trial; while this is correct, the authors have a model to account for that, and more trivially one may just leave out the first CS– response. Secondly, the observation function (which is only specified in the legend of table 1 – this should appear in the text) maps SCR onto the value terms of all models. All previous SCR modelling work has found SCR is more closely related to associability or uncertainty, or a mixture of these with value.

Concern: The models considered are broadly suitable but comparison of model fit across both SCRs to the CS+ and CS– during acquisition and SCRs to the CS+ and CS– in extinction is missing, which is a limitation.

Recommendation: Include prior models in the model space. Models should simultaneously include CS+ and CS– responses. To this end, they may leave out responses before the first reinforced CS+ trial. Model space should include additional variants of models 9 that account for a mapping onto associability (as in Li et al. 2011, Zhang et al. 2016) or combined associability and value (as in Homan et al. 2019) and a variant of model 3 that accounts for a mixture of uncertainty and value (as in Tzovara et al. 2018).

As requested by the reviewers, the 9 models were entirely refit to the combined CS+ and CS- trials during conditioning (see Figure 2—figure supplement 1), although it should be noted that it is not clear that threat and safety learning rely on the same circuitry and thus the same learning processes. We now clarify in the text in multiple places that both CSs were modeled in the same model (p. 10-15).

This was also done for the extinction process (discussed later; Figure 3—figure supplement 1). The models referenced in Li et al., Zhang et al., and Homan et al., are all the same version of the Pearce-Hall model which has already been used in our study (model 9). We apologize for not describing this model clearly. We include two associability terms in the model that drive changes in learning rate (for fitting CS+ and CS-), as now noted in p. 14.

We stress that the family of models chosen for this study was intentionally similar, and we are using them to improve the quality of the metrics (parameters) used to link the SCR data to the symptom and neural data. We note this in pp. 9-10.

We reviewed the variants of the PH model mentioned by reviewers (Li et al. 2011, Zhang et al. 2016, Homan et al. 2019) and examined the associability terms for model 9. We plotted the median values of the two associability parameters (α positive and α negative for CS+ and CS-), resulting from the model 9 fit (see Author response image 4) and compared the profiles to the acquisition data. However, the time course of associability for acquisition lacked the learning dynamics we observed in the SCR data, and for extinction, associability and value were highly similar.

Author response image 4

Download asset Open asset

As suggested, we also attempted to adapt the Tzovara et al. model, which uses uncertainty and value to co-determine the model’s predictions, to our reinforcement learning data. As Tzovara et al. conceived the model, it contains two parameters, and can only account for one type of reinforced data (i.e., CS+ or CS-, but not both). To model the response to both the CS+ and CS- simultaneously, we would need to add a third parameter (in addition to another two for the CS-) to the model, for a total of 5 parameters (more than any of our models, which have a maximum of 4 parameters). While possible, the BIC calculation would deduct points for this additional parameter, thus making it not an ideal model for these data. Author response image 5 comparing the Tzovara et al. uncertainty model to the CS+ data shows that the model fit across subjects does not pick up on the initial acquisition or subsequent habituation to the CS+:

Author response image 5

Download asset Open asset

Thus, it is possible that the process captured by Tzovara et al., which is modeled over many trials, is different than the rapid, initial acquisition process targeted in our study. Given this result and the difference in processes of interest, we did not pursue this model further. We provide this rationale in full in Appendix 1 (p. 60).Of note, we also discuss the absence of findings regarding the hybrid model in the context of this specific task (Discussion, p. 25): “Of note, prior work identified the RW-PH “hybrid” model as providing optimal fit to data for the expression of threat contingencies among healthy adults as well as adults with PTSD^12,33,34. Here, we found that this model did not provide the best fit for rapid threat conditioning or extinction among the models tested. This discrepancy could potentially be attributed to the nature of threat learning processes studied. While we focused on rapid acquisition/extinction of threat contingencies which takes place over a short training schedule, prior modeling studies examined threat contingencies over much longer durations (over 80 CS+ trials). The hybrid model includes a cumulative associability term which could be more sensitive to tracking contingency values or the expression of conditioned fear as it is optimized over longer durations. Thus, differences in rapid, crude acquisition vs optimized expression of threat contingencies processes could account for model differences. As such, the current and prior work could be seen as complementary in terms of threat learning processes studied, and both sets of findings could usefully inform study design for future work.”

Related Concern from Another Reviewer: The authors do not consider alternate models for the SCRs to the CS– during acquisition or for the SCRs to the CS+ during extinction and they do not model the SCRs to the CS– during extinction at all.

Recommendations: It would be helpful to know which model best fits all the acquisition data (to both the CS+ and CS–). Does the hybrid model still perform less well than the non–hybrid models? And does model 8 still outperform model 7? (in models 8 and 9, as in model 7, the CS– value could also be modelled as updating after each UCS delivery with model 8 now having 6 free parameters and model 9 4 free parameters.) Similarly, it would be useful to have both CS– and CS+ trials during extinction modelled. Given the raw data indicates the response to both the CS+ and CS– drop off during extinction, understanding if decreased responding to the CS+ during extinction represents active specific updating of the CS+ UCS contingency or a non–specific decreases in expectancy of the UCS after both the CS+ and CS– is clearly important.

Joint fitting of CS+ and CS- during acquisition:

As noted, we re-fit our nine models to both the CS+ and CS- during acquisition and performed model comparison as suggested. This is detailed in the previous responses and in the text. The hybrid model (model 9) did not perform particularly well, as can be seen in the bar chart of average BIC values across the models or fraction of best fit across the sample (Figure 2A). Notably, model 9 is also one of the two models (the other being model 6) that did not do well in model recovery. This model is designed to allow flexibility in learning rate, which is important in situations where there are changes in uncertainty over time. Because our task is a straightforward rapid acquisition/extinction task, we believe it might not necessarily provide the right statistics to lead to it fitting best. We note in this in Appendix 1 (pp. 59-60).

We performed a repeated-measures ANOVA on BIC values (model as fixed effect, subject as random effect), yielding a significant effect of model, F(1,8)=38.98, p<0.001, as noted in pp.18-19. Pairwise t-tests were conducted for model 9 versus all other models (Bonferroni corrected). Models 3, 4, 7, and 8 were significantly better than model 9, model 2 performed worse. Thus, we conclude that fitting the data from this task with this model does not produce parameter values that are optimal representations for this type of learning (i.e., rapid acquisition/extinction). This is now reported in Appendix 1 (pp. 59-60). Models 7 and 8, on the other hand, still perform best compared to all the other models. When directly compared to each other, model 8 and model 7 do not statistically differ (p=0.20). As before, since model 7 is a simpler model and they do not statistically differ, we chose model 7 as the optimal model for our data and used these parameters in subsequent analyses (pp.19-20). Thus, we have completely redone the analyses to account for these changes (pp. 19-21).

Extended modeling of extinction:

All 9 models were next fit to the CS+ and CS- extinction data; see Figure 3—figure supplement 1.

As noted in our response to previous recommendations, we provide information on joint CS-/CS+ fitting in p. 10-11. Several models had relatively low BIC values across the population, although overall there were no large differences between models, as can be seen Author response image 6.

Author response image 6

Download asset Open asset

We now note in p. 20 that we conducted a mixed-effects ANOVA on the BIC values (subject as random effect, model as fixed effect) to determine whether any of the models were significantly different than other models, and found a difference between models, F(1,8)=26.89, p<0.001. As can be seen, this effect is driven primarily by worse fits of models 4 and 9. We then ran follow-up pairwise t-tests for model 3, which was picked most often as the best at the individual subject level, and found that model 3 was not statistically different than models 1,2,5,6,7, or 8. Since model 3 was the simplest among the better models, and picked most often for individual subject fits, we used its parameters to model extinction. This is all now described fully in the text (p. 20).

We have accordingly redone the entire relevant section of the results concerning the relationship between the extinction data, anxiety symptoms, and neural data using model 3 parameters (pp. 20-21).

– Multiple Comparisons [Multiple Reviewers]

Concern – There are many post–hoc tests on association of model parameters with brain results and participant traits. It is unclear how this was planned and performed – how did the authors correct for multiple comparison across parameters, traits, and experiment phases?

Recommendation: provide appropriate, principled correction for multiple comparisons to mitigate familywise α inflation

We apologize for not making this part sufficiently clear, and we appreciate the opportunity to clarify our analysis plan as well as test level (α) adjustments. We now specify in the Methods (p. 16) that for each of the two phases (conditioning, extinction), we first identified the model best accounting for the phase data (which are now modeled using both CS- and CS+). For each estimated parameter in the winning model, we then examined whether it was associated with anxiety severity, and the moderation of this association by age using a regression model. Conditioning effects were best modeled using a model with four parameters (2 for CS+ and 2 for CS- responses, as we show in the Results section), and thus the test level was determined via Bonferroni as α=0.05/(4 parameters)=0.0125. Extinction effects were best modeled using two parameters (one for each CS, see Results), and thus significance level for each tested effect was determined as α=0.05/(2 parameters)=0.025. In terms of imaging analyses, we examined whether brain vertex (cortical thickness)/subcortical structure (GMV) was associated with each learning parameter as moderated by age. All imaging analyses used FWE rate correction for multiple comparisons of α<0.05, whereby the family of tests for each analysis consisted of all vertices/subcortical structures across all tested effects. This is now all detailed in the Methods section, p. 16.

Recommendation: Please state how many tests were conducted and what correction was applied for example how many brain regions were investigated, was grey matter volume in each of these related to each of the five learning rate parameters (across acquisition, generalisation and extinction), and were the moderating effects of anxiety, age and anxiety by age considered in each case?

Recommendation: please be clear in defining what "counts" as a family of analyses

We hope that our previous response clarifies the types of multiple-comparison correction used in analyses and what the tested effects were. Specifically, we note (p. 16) what counts as a family for behavioral analyses (all tested effects) and for imaging analyses (all vertices/subcortical structures across tested effects). We also now specify the number of vertices tested in the whole-brain cortical thickness analysis and moved the listing of subcortical structures tested for GMV from Appendix 1 to the main text (p. 15).

– Specificity of key results

Concern: The relationship between anxiety and safety learning (as indexed by the habituation parameter to the CS– during acquisition) is one of the main findings.

Concern: The novelty of this study depends both on the strength of the computational analysis and on the extent to which the results provide new insights into the relationship between child and adult anxiety and differential fear conditioning relative to the non–computational analyses reported in the authors' prior publication on this dataset (Abdend et al., 2020). The relationship between anxiety and safety learning (as indexed by the habituation parameter to the CS– during acquisition) could be one such novel insight – however here a direct comparison of the relationship between anxiety and this parameter versus the habituation parameter to the CS+ is needed to ensure this isn't simply a non–specific habituation effect. The same issue holds for the finding of a relationship between anxiety and rate of extinction to the CS+ – the authors need to show this holds over and above any relationship between anxiety and drop off in responsivity to the CS–.

Recommendation: Here, a comparison of the relative strength of relationship between anxiety and participants' habituation parameter values for the CS– versus that for the CS+ is needed to ensure this isn't simply a non–specific habituation effect. Similarly, the authors need to directly compare the relationship between anxiety and drop off in SCRs to the CS+ during extinction against the relationship between anxiety and drop off in SCRs to the CS– during extinction to be able to make any claim specific to extinction to the CS+. This also applies to the moderating effects of brain structure on this relationship.

Thank you for this recommendation, which allowed us to examine stronger inferences regarding the specificity of effects. The association between anxiety and CS- habituation rate during conditioning remained significant when controlling for CS+ habituation rate, β=-0.239, p=0.001, indicating the specificity of anxiety effects to safety learning. This is now noted in p. 19-20, and in the Discussion, pp. 24-25. Similarly, the association between anxiety and CS- extinction rate maintained its magnitude when controlling for CS+ extinction rate, p=0.031, and also when controlling for CS- learning and habituation rates during conditioning, p=0.030, providing support for the specificity of the anxiety-extinction association (reported in p. 21). Likewise, the significant brain structure finding for CS- habituation and anxiety moderation by accumbens GMV remained significant when controlling for CS+ habituation rate p=0.040 (p. 20). The effect for extinction was not re-assessed as it was only at trend level.

– SCR Approach

Concern: SCR data are likely not optimal for model fitting at the subject level; SCR is noisy, and the task involves very few trials. This may result in poorly estimated parameters; however it is not clear from the presented analyses how much of a problem this is.

Recommendation – Given the noisiness of SCR data and the limited number of trials available, parameter estimation could likely be improved substantially by using hierarchical Bayesian parameter estimation methods.

We agree that SCR data are indeed inherently noisy. This was a major motivation to accumulate a large sample of participants, especially in the studied populations. Accordingly, data collection took several years, as recruitment of the target populations (medication-free, clinically-diagnosed anxiety patients, including children, who are willing to take part in an aversive experiment) is a difficult, expensive, and time-consuming process. We opted to retain our original equipment and software during this period. Since it has been several years since data were collected and analyzed, and as our group has moved on to newer equipment for data collection and analysis since then, we, unfortunately, no longer have access to the original raw data and cannot reanalyze them. While model fit to the SCR data was good (as was model recovery) and the sample size was large, we are aware that not using newer methods could possibly decrease statistical power, rendering our approach rather conservative. It should also be noted, however, that recent work does not clearly indicate whether newer processing methods provide stronger or more valid estimates in the context of the paradigm used here (Kuhn, M., Gerlicher, A., and Lonsdorf, T. B. (2021, April 30). Navigating the manifold of skin conductance response quantification approaches – a direct comparison of Trough-to-Peak, Baseline-correction and model-based approaches in Ledalab and PsPM. https://doi.org/10.31234/osf.io/9h2kd). Nevertheless, we now note this as a limitation of the study in p. 28, citing papers by Bach et al. which provides background and examples for such methods: “Fourth, SCR data were analyzed using conservative methods (base to peak amplitude); future studies should consider utilizing more novel analysis methods which may increase sensitivity and statistical power^96-99”.

Concern: The quantification of the conditioned response is in line with some previous work, but in fact there is a range of possible analysis choices, summarised in Lonsdorf et al. 2017.

Recommendation: Given this heterogeneity, it would make sense to use the most sensitive method. Bach et al. 2020 Behaviour Research and Therapy and Ojala and Bach 2020 Neurosci Biobeh Rev discuss the most sensitive method to quantify threat conditioning from skin conductance data, and Bach et al. 2020 Nat Hum Behav provide background.

We appreciate the suggestion to re-analyze the dataset using alternative methods of analysis as indeed multiple different choices do exist. It is quite possible that changing the way that we analyze data in our group to a different method might increase the effects sizes, although, as noted above, it is not clear that this effect would be achieved (see reference to Kuhn et al.; newer methods might not necessarily show more sensitivity than the method used here). In any event, re-analyzing the data will likely change the nature of findings, from learning models to their associations with symptoms/brain structure, since different methods might not necessarily capture the same physiological processes (e.g., assuming a canonical phasic response vs possible changes in SC that include phasic but also tonic changes – our own recent work, under review elsewhere, supports the latter notion). As such, it is difficult for us to conclusively say that all alternative methods measure the same exact physiological process, and we cannot conclusively say (given recent evidence) which method is the ideal one in this context. In this case, the dataset should be re-analyzed using multiple methods, and we worry that having to analyze the data using multiple methods and indices might make interpretation difficult and the paper more challenging to read. Importantly, as we note in our response to the previous comments, we were also unable to reanalyze the data in this specific case, for practical reasons. Nevertheless, we agree that using the current method, while standard, might not necessarily be the most powerful method and could therefore be considered more conservative. As such, as noted above, we added an important limitation that pertains to the analysis method that also encourages future work to explore newer analytic methods (p. 28; see above). We cite the papers suggested by the Editor, as well as additional work by Bach et al., to provide context.

– Missing SCR Data

[Multiple reviewers] Concern: More than 1/3 of the sample are excluded due to "excessive missing data", defined as >50% missing trials.

Recommendations: It is unclear why these data were missing. Please address the explanation and please discuss the potential consequences for inference. Do excluded subjects systematically differ from those included? It would be helpful if the authors could include a table showing a break–down of the characteristics of included and excluded participants by age, gender, diagnostic status and anxiety levels and if they could address in the discussion any associated limitations in particular in relation to the age by anxiety analyses reported.

Indeed, a substantial amount of SCR was missing. We agree that this is an important point to clarify and discuss. We report the criteria and exclusion process, as recommended by Lonsdorf et al. (2019, eLife), in Appendix 1 (pp. 52-53). Exclusion was due to individuals with no response or with outlier values. A comparably large proportion of non-responders has previously been reported (e.g., Homan et al., 2019; Lonsdorf et al., 2019). To assess whether the SCR processing method led to this dropout level, we analyzed using newer approaches (as suggested above) a different, ongoing dataset of adolescents who completed this task. We observed ~25% missing data using similar criteria, suggesting that it is not the processing method. Instead, based on our extensive experience, the noisiness and difficulty in collecting good signal reflects the challenges of physiological data collection particularly across a wide age range and in those diagnosed with anxiety disorders, in the context of uninstructed aversive tasks. Due to these challenges, we chose conservative criteria, resulting in a large proportion of exclusion, since modeling was based on trial-by-trial data, and we needed sufficient usable trial-level data. Additionally, data collection for this study took place over several years, due to the need to acquire a large sample of patients across age; newer systems used today may be more sensitive, signal-wise.

As recommended by the Editor, we added Appendix 1-table 1 which details differences between included and excluded participants in terms of age, sex, anxiety (both diagnostic status and continuous severity), and IQ. We noted significant differences in age and sex (but not anxiety or IQ), such that excluded (non-responder) individuals tended to be older (M=23.4 vs. 18.8 years) and with a greater proportion of females (69.1% vs. 54.0%). This is now noted in the main text (p. 8) and Appendix 1. Indeed, age and sex differences in responsivity have been previously noted, but inconsistently (e.g., Boucsein, et al., 2012, Psychophysiology; Bari, et al., 2020, J Biol Phys; Ganella, et al., 2017, Front Hum Neuro).

It should be noted that primary analyses included age in the models, and as such, anxiety effects are independent of age effects. We further examined the primary analyses using sex as a covariate (p. 53).

Finally, we added a paragraph discussing these important issues to Appendix 1 (p. 53): “This proportion of exclusion is similar to excluded data in prior work^12,67. See Appendix 1-table 1 for demographic and clinical differences as a function of inclusion/exclusion. No significant differences in exclusion were noted for anxiety severity (by diagnosis or continuously) or IQ, but excluded relative to included participants had a higher mean age and a greater proportion of females. Note that age was included in tested models, and thus observed anxiety effects are independent of age effects; further, results did not change when using sex as a covariate. Based on our experience, we believe that this exclusion proportion reflects challenges of physiological data collection during uninstructed aversive tasks, particularly across a wide age range and in those diagnosed with anxiety disorders, in conjunction with conservative exclusion rules since modeling required sufficient trial-by-trial data. These considerations should be taken into account in future research, particularly in studies across a wide age range and in those that consider sex, and require sufficient trial-level data. Implementing newer analytic techniques and equipment, as well as use of more robust aversive stimuli, could potentially mitigate this issue.”

Given other comments, we no longer report on significant Age x Anxiety interactions (see response to later comment). As requested, to address this important issue, we added another limitation referring to age and sex differences (p. 28): “Finally, there were differences in sex and age between individuals included and excluded (non-responders) from analyses, as previously reported^102,103; while these differences did not influence our findings, they could still limit generalizability and thus future studies should consider such potential differences.”

We hope that this information could help researchers by alerting them to such potential differences in responsivity.

Concern: It is unclear why missing data are inter/extrapolated?

Recommendation: Use MSE. The model fitting procedure should be happy with missing data points if MSE (rather than SSE) is used as objective function. This would be more appropriate than giving the model a data value that does not exist. Censor, rather than interpolate.

We chose to impute missing time-series data points since model generation relied on trial-by-trial data for a limited number of trials, and learning effects were very rapid. We chose to assume that missing data points lie between the recorded data around them and impute missing data, rather than censor since we were worried that censoring would lead to spurious estimates. We now provide this rationale in Appendix 1, p. 52.

– Disparate Dimensional Measures of Anxiety [Multiple Reviewers]

Concern: The combination of multiple anxiety measures (SCARED, STAI) into a single anxiety measure may not be valid, as it's reasonable to assume these are measuring subtly different constructs (aside from anything else, the SCARED measures recent symptoms while the STAI is a trait measure).

Concern: The main analyses appear to combine measures of anxiety severity across age groups as the primary variable of interest (using the SCARED for youth and the STAI for adults), and in one analysis find an interaction between age and anxiety that is driven by the younger age group. This raises some concerns as these measures are not necessarily measuring the same construct, and may limit the interpretability of these results.

Concern: Children and adults show opposing direction relationships between anxiety and fear generalisation. It is difficult to interpret the anxiety by age results given different measures were used to assess anxiety in children and in adults. This could potentially account for the different relationship between anxiety and fear generalisation in under and over 18s (see Figure 3B). I am not convinced that the authors can simply merge scores from the SCARED (which averages symptom scores based on child and adult reports) and STAI (an adult self–report questionnaire) as if they came from the same measure.

Recommendation: Absent some sort of direct quantitative harmonization (as in the GWAS literature for N/NE) or additional evidence to motivate the Z–score approach, it may not be appropriate to combine or directly compare the younger (SCARED) and older (STAI) samples (i.e. more appropriate to keep them distinct and examine them separately)

Thank you for this interesting comment. We agree that the SCARED and STAI are not the same measure, despite each being considered a ‘gold standard’ for the overall assessment of anxiety severity for its age group, and thus may not be fully comparable. At the same time, given the developmental trajectory of anxiety (Beesdo et al., 2009, Psych Clin N Am) and the great interest in developmental effects on threat learning circuitry in the context of anxiety and its treatment (e.g., Casey, et al., 2015, Neuron; Shechner, et al., 2015, Dep and Anx; King, et al., 2014, Stress), which was the impetus for this study, we believe it is still important to report on age effects across the age continuum. Balancing the need to comprehensively report our results as intended with the potential incongruity of anxiety measures, we repeated the primary analyses separately in youth and adults while also keeping the original analyses.

To alert the reader to this important issue and to the fact that additional analyses were conducted, we provide this acknowledgement relatively early in the Methods section (p. 7): “Given that the SCARED and STAI might not capture the construct of anxiety in an identical manner, we also report on SCARED analyses in youth and STAI analyses in adult participants separately.”

We retained the analyses examining associations between anxiety and learning across the age continuum since these already control for age. Further, we now report in the Results section (p. 21) that additional analyses were performed, as well as the rationale, and provide a brief summary. We provide the results in full in Appendix 1. Specifically, the brief summary in the main text notes: “In addition to the primary analyses examining associations between anxiety and learning parameters and the moderation of these associations by brain structure across the full sample, we also examined these effects separately among youth and adult participants. These are reported in full in Appendix 1 (see below). Briefly, they suggest that observed anxiety effects on safety learning and extinction are attenuated with age. Further, these suggest moderation of anxiety-safety learning rate by several structures (bilateral nucleus accumbens, brain stem, and amygdala) in youth participants but not in adults. In contrast, extinction rates were associated with amygdala, accumbens, and brainstem volume moderation in adults, but not in youths; see Appendix 1.”

The full text in Appendix 1 (pp. 61-62) states: “In addition to the primary analyses reported in the main text, we also examined associations between learning parameters and anxiety separately among youth and adult participants, as described below.

Conditioning. Anxiety was not significantly associated with CS+ threat conditioning or habituation rates in the youth group, βs≤0.11, ps≥0.665, and in the adult group, βs≤0.53, ps≥0.289. Anxiety severity was not significantly associated with CS- generalization rate in the youth group, βs≤0.51, ps≥0.297, and in the adult group, βs≤0.45, ps≥0.297. CS- safety learning rate was negatively correlated with anxiety in the youth group, β=-0.27, p=0.002, as in the full sample, and marginally so in the adult group, β=-0.24, p=0.056, suggesting an attenuation of anxiety-safety learning association with age. No other effects were observed.

Effects of brain structure emerged in the youth group. These included a negative association between brain stem volume and CS- safety learning rate, as well as bilateral accumbens moderation effects similar to those reported in the main text, p_FWEs<0.05. Right amygdala showed a similar moderation effect at trend level, p_FWE=0.0628. In the youth group, left accumbens showed a moderation effect similar to that reported in the main text and in the youth group, but at trend level, p_FWE=0.073.

Extinction. In the youth group and the adult group, anxiety was not significantly associated with CS+ extinction rate, β=0.12, p=0.227, and β=-0.16, p=0.204, respectively. CS- extinction rate was negatively correlated with anxiety in the youth group at comparable magnitude to what was observed across the full sample, β=-0.20, p=0.032, but not in the adult group, β=-0.09, p=0.504, suggesting an attenuation of the association between anxiety and CS- extinction with age. No other effects were observed.

In the youth group, no effects of brain structure emerged. In contrast, in the adult group, a significant positive association between right amygdala volume and CS+ extinction rate was noted, p_FWE=0.008. Brain stem volume moderated the association between anxiety and CS+ extinction, p_FWE=0.012. Anxiety-CS- extinction rate was moderated by left accumbens and brain stem volume at trend level, p_FWEs<0.073.”

Importantly, we also added a full paragraph in the Discussion section in which the differences between analyses in full vs. age-specific samples are discussed (pp. 25-26). Specifically, the text reads: “While the design of this study offers a unique opportunity to examine age effects on threat learning as these relate to anxiety severity, an inherent challenge that arises in research on anxiety along the lifespan is how to combine anxiety data from youth and adult participants. Although the SCARED and STAI are each considered “gold standard” measures in their respective target populations, they nevertheless are not identical. Under the assumption that these capture similar constructs, we uncovered several anxiety effects across the full age range. Alternatively, one may consider these measures to be incomparable, in which case analyses are restricted to specific age groups. This alternative approach indicated an attenuation of anxiety- learning associations with age. As such, one interpretation of the discrepancies is that observed effects may in fact be instrument-specific. Another interpretation is that limited range and sample size reduced statistical power to detect associations in the different sub-groups. In either case, consideration of harmonizing clinical data is needed in future research that examines individuals across a wide age range.”

– Potential nonlinear effects of age

Concern: Only linear effects of age are modelled.

Recommendation: explore potential nonlinear effects (given that many developmental effects are nonlinear).

We previously reported quadratic effects (all with null results) in Appendix 1. These results are reported in p. 62.

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

The manuscript has improved, with one Reviewer emphasizing that "the authors have done a good job of addressing many of the suggestions and the manuscript is definitely substantially improved as a result."

Thank you for this feedback. We agree that subsequent recommendations in the second round of reviews have also significantly improved our manuscript and the clarity of our findings.

Nevertheless, there was agreement that the paper would benefit from some additional modeling.

One Reviewer wrote that, Abend et al. have addressed some, but not all of my previous concerns. The model space of the acquisition session does not include any of the "best" threat learning models from previous publications. This precludes a direct comparison; more importantly it may also miss out on the best model in this dataset.

The authors now model CS– and CS+ responses simultaneously. Yet they still do not compare their models to any of the those used in previous studies of conditioned SCR (Li et al. 2011, Zhang et al. 2016, Tzovara et al. 2018, Homan et al. 2019). They claim that their model 9 is the hybrid model used in three of these studies, but actually it is not, because the update works differently (presently they update only upon US delivery, and both CS+ and CS– from PE in CS+ trials, where as previous RW and hybrid models used the standard update on every trial, on signed PE, and with no cross–update). They qualitatively argue why the associability term in previous work does not fit their data, but in fact they should just quantitatively compare models that map value and/or associability onto SCR. Finally, they argue that Tzovara's model can only update CS+ or CS– expectation with one parameter, and modelling both CS types within the same model would lead to parameter inflaton, but in fact Tzovara's winning learning model is entirely parameter–free (only has parameters for the observation function), so this is inaccurate.

The authors need to implement these pre–existing models and show quantitatively that they do not fit the data, rather than on rely on qualitative and methodological arguments for why this is unnecessary. Based on their rebuttal, it seems that they had already implemented the models, so it should be relatively straightforward to make the necessary direct ("head–to–head") comparisons.

Along related lines, the other Reviewer noted that, [page 11] The authors argue that to adapt the Tzovara et al. model, a 5th parameter would need to be added, and that this would make it too complex. It would still be worth testing this properly [per the other reviewer's comments], and it is possible (although probably unlikely based on the provided evidence) that the model could fit well enough that it still wins despite the additional complexity–related penalisation.

These suggestions are valid. In response, we have fit 5 additional models to our data, including three from Li et al., 2011: Hybrid(V), Hybrid(α) and Hybrid(V+α) and the Tzovara et al., 2018 probabilistic uncertainty model with and without an additional habituation parameter. We fit these 5 models to our acquisition data and 3 of the models (Li et al., 2011) to the extinction data. It is not possible to fit the Tzovara model in a meaningful way to extinction because there are no UCSs. This led to a total of 14 models used to model acquisition data and 12 models used to model extinction data. We have updated the manuscript to reflect this (detailed below).

We found that on average, these models provided reasonable fits to subject data. Please see Figure 2—figure supplement 1 and Figure 3—figure supplement 1 in the Appendix for all model and data overlays that include the new models for acquisition and extinction. Although we were able to fit the models, the Li et al. models had a large number of parameters relative to the number of data points available to fit them. After fitting these models, we also carried out parameter recovery on all of the models to evaluate the quality of the fits. We found that the Li et al., models had poor parameter recovery. Therefore, to improve the model selection process, we imposed a minimal criterion on the correlation coefficients of the parameters to determine which models should be considered in final model comparison. We imposed a criterion that all parameter correlation coefficients for a given model should be ≥0.20, if we subsequently included that model in model comparison.

Many parameters could not be recovered for the new models (see Appendix 1- Table 2). The Li et al., models had considerable additional complexity, and thus it is likely that the regression step that adds 1-3 additional parameters was allowing for too much flexibility to fit noise in the data. Our original model 9, which we called the Pearce-Hall Hybrid model, lacked this regression step, but was otherwise a similar model. This model also did not survive parameter recovery, suggesting that this family of models was not the best set of models for this particular dataset and processes modeled, despite previous reports and use of these models for other datasets. Importantly, as we note in the text, this dataset features a limited number of trials per subject in order to capture rapid learning processes; while we compensate for that with a large number of subjects, perhaps those models are better suited for processes that use more trials for each subject to estimate the larger number of parameters.

The Tzovara models fit the data reasonably well, and because parameters could be recovered (Appendix table 1- table 2), were included in model comparison. However, these models did not produce an average BIC better than the previously chosen model (Figure 2A). Pairwise t-tests showed that models 7 and 8 remained the best models (ps<0.05, Bonferroni-corrected). As these two models did not differ from each other, we chose the simpler model (model 7, Bayesian learning rate and habituation) and the original conclusions from this section did not change.

Of note, we also tried another variant of the Tzovara model with two additional habituation parameters (habituation for CS+ and CS- for 6 total parameters) but the habituation parameters of this model were not recoverable during parameter recovery (p. 14; Appendix 1-table 2). Thus, we retained the model with a single habituation parameter.

Using the same methods, we evaluated the new models for extinction data, which included the 3 Li et al., models (Figure 3—figure supplement 1). After parameter recovery and imposing a criterion for parameter recovery, four models remained for model comparison: models 1, 2, 3 and 5 (Appendix 1- Table 3). These models were not statistically different from each other (ps>0.05, Bonferroni-corrected), and model 3 was the best fitting model for more subjects than the competing models (Figure 3B). Thus, we retained our original model and use of parameters from this model to correlate with anxiety severity and brain volume.

Please see updated figures and tables that account for the additional models:

– Figure 2A, 2B (main text): updated bar plots for models compared for acquisition

– Figure 2—figure supplement 1: model/data overlays for all 14 models fit to acquisition

– Figure 2—figure supplement 3: fractions of subjects by age/anxiety for acquisition

– Figure 3A, 3B (main text): updated bar plots for models compared for extinction Figure 3—figure supplement 1: model/data overlays for all 12 models fit to extinction

– Figure 3—figure supplement 1: model/data overlays for all 14 models fit to extinction

– Appendix 1-table 2: parameter recovery for acquisition

– Appendix 1-table 3: parameter recovery for extinction

The Reviewers and Editors also discussed hierarchical Bayesian approaches, with one reviewer noting that, [page 19] Reviewer suggestions related to hierarchical Bayesian parameter estimation do not seem to have been addressed in the revision. For clarity, this refers to the method of estimating reinforcement learning model parameters (see van Geen and Gerraty, 2021), rather than the SCR quantification, and there's a strong possibility it would improve parameter estimation and hence sensitivity. I appreciate that this could be represent a substantial re–analysis.

During the consultation session, the 2 Reviewers agreed that the hierarchical Bayesian approach would be nice, but is not strictly necessary for the revision, with one noting that, "I wouldn't want to force them to redo all the model fitting with a hierarchical Bayesian approach, it would likely provide more accurate estimates, but it would be a fairly big undertaking. Some acknowledgement that other model–fitting procedures exist and could be more appropriate would be nice, however (this probably also applies to their SCR quantification methods)".

And the other writing that, "I agree. This model fitting approach is a major step and may not work for some of the models."

In sum, the Bayesian approach should –– at minimum –– be noted in the Discussion.

Thank you for this suggestion and we apologize for not acknowledging other model fitting techniques more clearly. We have updated the Discussion section to include a section acknowledging Bayesian parameter estimation and potential future work in this direction. We acknowledge the suggested papers in addition to others describing hierarchical Bayesian methods. Please see Discussion, pages 29,30:

“As a final technical comment, reinforcement learning model parameters were estimated using maximum likelihood techniques on individual subjects followed by model comparison. Future work could expand on this by using hierarchical Bayesian parameter estimation to reduce the variance around parameter estimates^105-107. However, choosing prior distributions within the hierarchical Bayesian approach is not trivial and may not work for all of the models tested in this study. As such, future work could focus on fewer models, such as those that survived parameter recovery in this study, test a range of priors, and determine whether parameter estimation could be improved.”

We have also expanded our description of other SCR quantification methods. (Discussion, page 29):

“Fourth, SCR data were analyzed using a single method which, while established, relies only on directly-observable effects. Future studies may consider using novel, computational analysis methods which could potentially reveal effects not observed using the current method^96-99. Along these lines, a multiverse approach may be used in future work to comprehensively compare multiple methods of quantifying threat learning”.

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

We are very pleased to accept your paper for publication pending receipt of a revision that adequately addresses the remaining suggestions from Reviewer #2. Congratulations in advance!

Reviewer #2 (Recommendations for the authors):

This revision of the paper by Abend et al. is significantly stronger by including a direct comparison of the current model space with previous "best" models. There are two substantive and a few presentation points remaining.

1. The Tzovara et al. model is implemented differently from the original paper. The authors use separate regression parameters for CS+ and CS– trials for the Tzovara et al. model (i.e. 4 parameters) and it is not clear to me why – Tzovara et al. themselves had used only 2 regression parameters across these two trials types. With 2 instead of 4 parameters, the model might fare better (even though the fit/RSS will of course become worse). Could the authors include the original 2–parameter Tzovara et al. model?

Thank you for this detail and for the suggestion. We have repeated the fitting of the Tzovara model with only 2 regression parameters for CS+ and CS- trials. Previously, we had tried to give the model the best chance of success by using separate regressors for CS+ and CS-, as the CS+ data has different dynamics than the CS- data. The model and data overlays for the Tzovara model with 2 parameters are shown in the updated Figure 2—figure supplement 1. For completeness, we also repeated the modeling using additional habituation terms for a Tzovara model with 4 total parameters. The model and data overlays are also shown in updated Figure 2—figure supplement 1. We replaced the previous models 13 and 14 with the requested versions of the Tzovara models in all aspects of the manuscript. Parameter recovery values have been updated in Appendix1-Table 2. These are the new models 13 and 14 to replace those reported previously. Model 13 survived the parameter recovery criteria, whereas model 14 did not. Thus, model 13 was included in the revised versions of Figure 2A and Figure 2—figure supplement 3 and BIC analyses.

We have also updated the uploaded code (available on a GitHub page) to reflect the most recent versions of the Tzovara models.

2. Perhaps relatedly, I've been wondering about the initial values of the various terms in the update equations. Can they be presented in the tables? I'm wondering, for example, whether they used the same values for α_0 and β_0 in the β distribution of the Tzovara et al. model, but other readers might wonder about other models. Of course, it can make a big difference how a model is "initialised".

Thank you for this suggestion. We have added the initialization values, for both conditioning and extinction models, to the table of equations (Table 1) for completeness and clarity.

3. The authors make an effort to explain the difference between models 1–9, and models 10–14, by using different symbols on the LHS of the equations. Perhaps it could be made more prominent that models 1–9 predict CS–related SCR from US–related SCR and thus need no regression to scale the predictions to the data, whereas models 10–14 predict CS–related SCR from an arbitrary reinforcement variable.

Thank you for this suggestion to improve the clarity of our modeling procedures. We have added the following text to the manuscript in the Methods on page 14:

“In summary, for models 1-9, predictions were not scaled to the data by using regression. Models 1-9 predict CS-related SCR from US-related SCR. For models 10-14, CS-related SCR is predicted using a regression to associability, value, or uncertainty, as described for each model. For models 1-9, outcome was modeled as a continuous variable. For models 10-14, we modeled outcome as a binary value of 0 or 1, as described in past publications using these models.”

4. I'm still unsure about using t–tests to compare BICs across a group. I have never seen this. What is the statistical foundation for this approach? I would suggest adding the RFX–model comparison by Stephan et al. as had been suggested by two reviewers in the context of the initial manuscript version. It is implemented in a matlab function (spm_BMS), requires only the matrix of BIC values, and takes a few seconds to run.

We apologize that our model comparison procedures were unclear. The referenced paper by Stephan et al., (2009), “Bayesian model selection for group studies,” states a description of the random effects procedure as was used in our manuscript:

“A straightforward random effects procedure to evaluate the between-subject consistency of evidence for one model relative to others is to use the log-evidences across subjects as the basis for a classical log-likelihood ratio statistic, testing the null hypothesis that no single model is better (in terms of their log-evidences) than any other. This essentially involves performing an ANOVA, using the log evidence as a summary statistic of model adequacy for each subject.

This ANOVA then compares the differences among models to the differences among subjects with a classical F-statistic. If this statistic is significant one can then compare the best model with the second best using a post hoc t-test.”

“The most general implementation would be a repeated-measures ANOVA, where the log-evidences for the different models represent the repeated measure. At its simplest, the comparison of just two models over subjects reduces to a simple paired t-test on the log-evidences.”

This was the methodology used in our manuscript as described in the Methods (pp. 15-16)­.

In addition, per the Reviewer’s suggestion, we have run the spm_BMS function from the SPM imaging toolbox on our BIC values. Using a flat prior across the models, we reproduced the posterior probabilities of the six models compared for the acquisition data:

Model 1: 0.0351

Model 4: 0.1356

Model 5: 0.1950

Model 7: 0.1423

Model 8: 0.2369

Model 13: 0.2551

These values are nearly identical to the fraction of subjects best fit by each model, shown in Figure 2A. Given that we do not expect that prior probabilities of each model should be anything but uniform, we did not expect significant differences between the RFX model comparison method and the random effects procedure.

Author details

1. Rany Abend

   Emotion and Development Branch, National Institute of Mental Health, National Institutes of Health, Bethesda, United States

   Contribution

   Conceptualization, Formal analysis, Investigation, Methodology, Project administration, Software, Supervision, Validation, Visualization, Writing - original draft, Writing – review and editing

   Contributed equally with

   Diana Burk

   For correspondence

   rany.abend@nih.gov

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-0022-3418

2. Diana Burk

   Laboratory of Neuropsychology, National Institute of Mental Health, National Institutes of Health, Bethesda, United States

   Contribution

   Formal analysis, Methodology, Software, Visualization, Writing – review and editing

   Contributed equally with

   Rany Abend

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0001-7775-2989

3. Sonia G Ruiz

   Emotion and Development Branch, National Institute of Mental Health, National Institutes of Health, Bethesda, United States

   Contribution

   Data curation, Formal analysis, Methodology, Project administration, Visualization, Writing - original draft, Writing – review and editing

   Competing interests

   No competing interests declared

4. Andrea L Gold

   1. Emotion and Development Branch, National Institute of Mental Health, National Institutes of Health, Bethesda, United States
   2. Department of Psychiatry and Human Behavior, Brown University Warren Alpert Medical School, Providence, United States

   Contribution

   Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Supervision, Writing - original draft, Writing – review and editing

   Competing interests

   No competing interests declared

5. Julia L Napoli

   Laboratory of Neuropsychology, National Institute of Mental Health, National Institutes of Health, Bethesda, United States

   Contribution

   Formal analysis, Investigation, Methodology, Software, Visualization, Writing – review and editing

   Competing interests

   No competing interests declared

6. Jennifer C Britton

   1. Emotion and Development Branch, National Institute of Mental Health, National Institutes of Health, Bethesda, United States
   2. Department of Psychology, University of Miami, Coral Gables, United States

   Contribution

   Conceptualization, Data curation, Investigation, Methodology, Project administration, Supervision, Writing - original draft, Writing – review and editing

   Competing interests

   No competing interests declared

7. Kalina J Michalska

   1. Emotion and Development Branch, National Institute of Mental Health, National Institutes of Health, Bethesda, United States
   2. Department of Psychology, University of California, Riverside, Riverside, United States

   Contribution

   Data curation, Investigation, Methodology, Project administration, Supervision, Writing - original draft, Writing – review and editing

   Competing interests

   No competing interests declared

8. Tomer Shechner

   1. Emotion and Development Branch, National Institute of Mental Health, National Institutes of Health, Bethesda, United States
   2. Psychology Department, University of Haifa, Haifa, Israel

   Contribution

   Data curation, Investigation, Project administration, Supervision, Writing – review and editing

   Competing interests

   No competing interests declared

9. Anderson M Winkler

   Emotion and Development Branch, National Institute of Mental Health, National Institutes of Health, Bethesda, United States

   Contribution

   Formal analysis, Investigation, Methodology, Software, Visualization, Writing – review and editing

   Competing interests

   No competing interests declared

10. Ellen Leibenluft

    Emotion and Development Branch, National Institute of Mental Health, National Institutes of Health, Bethesda, United States

    Contribution

    Methodology, Supervision, Writing – review and editing

    Competing interests

    No competing interests declared

11. Daniel S Pine

    Emotion and Development Branch, National Institute of Mental Health, National Institutes of Health, Bethesda, United States

    Contribution

    Conceptualization, Funding acquisition, Investigation, Project administration, Resources, Supervision, Writing – review and editing

    Competing interests

    No competing interests declared

12. Bruno B Averbeck

    Laboratory of Neuropsychology, National Institute of Mental Health, National Institutes of Health, Bethesda, United States

    Contribution

    Conceptualization, Formal analysis, Investigation, Methodology, Project administration, Software, Supervision, Writing - original draft, Writing – review and editing

    Competing interests

    No competing interests declared

    "This ORCID iD identifies the author of this article:" 0000-0002-3976-8565

• 1,338

  Page views

• 275

  Downloads

• 1

  Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.