Neural representations of ambiguous affective stimuli and resilience to anxiety in emerging adults

The tendency to interpret ambiguous stimuli as threatening has been associated with a range of anxiety disorders. Responses to ambiguity may be particularly relevant to mental health during the transition from adolescence to adulthood ( “ emerging adulthood ” ), when individuals encounter unfamiliar challenges and navigate novel social situations. However, it remains unclear whether neural representations of ambiguity relate to risk for anxiety. The present study sought to examine whether multivariate representations of ambiguity — and their similarity to representations of threat — relate to appraisals of ambiguity or anxiety in a sample of emerging adults. Participants ( N = 41) viewed threatening (angry), nonthreatening (happy), and ambiguous (surprised) facial stimuli while undergoing fMRI. Outside of the scanner, participants were presented with the same stimuli and categorized the ambiguous faces as positive or negative. Using representational similarity analyses (RSA), we investigated whether the degree of pattern similarity in responses to ambiguous, nonthreatening, and threatening faces within the amygdala related to appraisals of ambiguous stimuli and anxiety symptomatology. We found that individuals who evidenced greater similarity (i.e., less differentiation) in neural representations of ambiguous and nonthreatening faces within the left amygdala reported lower concurrent anxiety. Additionally, trial-level pattern similarity predicted subsequent appraisals of ambiguous stimuli. These findings provide insight into how neural representations of ambiguity relate to risk or resilience for the development of anxiety.


Introduction
Anxiety disorders are estimated to affect 31.1% of adults over the course of their lifetime (Kessler, Berglund, et al., 2005;Merikangas et al., 2010). The prevalence of anxiety and its high comorbidity with a multitude of other illnesses (de Mathis et al., 2013;Groen et al., 2020;Kalin, 2020;Kessler, Chiu, et al., 2005;Spoorthy et al., 2019) presents a significant public health concern. It is therefore crucial to identify neurobehavioral markers of risk for anxiety that might inform intervention and prevention efforts. Additionally, careful study of resilience factors can facilitate discovery of protective mechanisms that buffer risk for anxiety, which can be integrated into treatment development and prevention efforts tailored to at-risk individuals (Kalisch et al., 2017).
One widely studied marker of risk for anxiety is the tendency to evaluate ambiguous stimuli or situations negatively (Chen et al., 2020;Hirsch et al., 2016;Stuijfzand et al., 2018;Yoon et al., 2020). Relative to individuals low in anxiety, anxious individuals are more likely to endorse threatening interpretations of ambiguous sentences and scenarios (Butler & Mathews, 1983;Eysenck et al., 1991;Taghavi et al., 2000), associate ambiguous auditory tones with undesirable outcomes (i.e., low reward) in a reward learning task (Aylward, Hales, Robinson, & Robinson, 2020), and perceive negative outcomes as more likely to occur in ambiguous situations (Smith et al., 2016). In addition to correlational evidence linking responses to ambiguity to anxiety symptom severity, there is some evidence that negative responses to ambiguity may play a causal role in the development and maintenance of anxiety: Experimental manipulations that bias individuals to appraise ambiguous stimuli as threatening have been found to increase stressor-induced anxiety (Wilson et al., 2006), whereas cognitive-behavioral interventions designed to reduce these biases have been found to lower anxiety symptomatology (Hallion & Ruscio, 2011;Jones & Sharpe, 2017). Examinations of responses to ambiguity thus provide a powerful lens through which to study risk and resilience to anxiety.
A number of studies have specifically demonstrated a link between anxiety and the tendency to interpret ambiguous facial expressions negatively, often referred to as a "negative valence" or "negative interpretation" bias (Gebhardt & Mitte, 2014;Heuer, Lange, Isaac, Rinck, & Becker, 2010;Maoz et al., 2016;Park, Vasey, Kim, Hu, & Thayer, 2016). Facial expressions can signal interpersonal conflict (e.g., anger toward another individual) as well as potential environmental threat (e.g., a reaction to something frightening in the environment (Gilboa-Schechtman & Shachar-Lavie, 2013;Heuer, Lange, Isaac, Rinck, & Becker, 2010). Rapid processing of facial expressions is thus integral to threat detection (Feldmann-Wüstefeld, Schmidt-Daffy, & Schubö, 2011;Haxby, Hoffman, & Gobbini, 2002). For this reason, facial expressions are considered to be ecologically valid stimuli (Gilboa-Schechtman & Shachar-Lavie, 2013) that are useful for probing threat sensitivity in relation to anxiety (Gilboa-Schechtman & Shachar-Lavie, 2013;Staugaard, 2010). Whereas certain facial expressions signal clear or unambiguous affective states (e.g., angry or happy expressions), surprised faces are ambiguous in that they can signal positive or negative affect. Given the notable between-person variability in interpretations of surprised expressions, surprised faces are well-suited for indexing individual differences in affective processing and biases in interpretations of ambiguity (Neta, Norris, & Whalen, 2009). In a previous study, Park et al. (2016) found that higher anxiety was associated with a greater tendency to interpret surprised faces negatively. Notably, ratings of unambiguous (i.e., angry and happy) images did not relate to anxiety, suggesting that anxiety-related negative valence biases are appreciable specifically under conditions of ambiguity. Crucially, the opposite tendencyhaving a more "positive valence bias" when appraising ambiguous facesmay be protective and buffer risk for anxiety. For instance, in children who were exposed to extreme early life adversity, a group at high risk for anxiety disorders (Silvers et al., 2017), positive evaluations of surprised faces predicted lower internalizing symptoms (VanTieghem et al., 2017).
The majority of studies linking amygdala reactivity to anxiety have relied on univariate signals to examine responses to affective cues (Dunsmoor & Paz, 2015;Etkin & Wager, 2007;Freitas-Ferrari et al., 2010;Kolesar et al., 2019;Shackman & Stockbridge, Tillman, et al., 2016). While useful, univariate analyses rely on averaging BOLD signals across voxels and distilling them into a single index (i.e., the overall magnitude of a response), thereby removing information contained in the distributed patterns of responses. Such approaches are thus unable to capture voxelwise response patterns to stimuli within a given region, or how similar these distributed patterns are across conditions Weaverdyck et al., 2020). Multivariate approaches, including multivariate pattern analysis (MVPA) and representational similarity analysis (RSA), are alternative techniques that preserve distributed spatial patterns of voxelwise BOLD responses and enable comparisons between neural representations of different types of stimuli Kriegeskorte, 2008;Weaverdyck et al., 2020). Intriguingly, in studies that did not investigate anxiety, researchers demonstrated that multivariate representations within the amygdala track perceived trustworthiness of ambiguous faces (FeldmanHall et al., 2018;Tashjian et al., 2019). Such techniques may provide insight into whether similarity in multivariate representations of ambiguity and threat can predict later appraisals of ambiguity, and whether similarity in these representations varies as a function of anxiety symptomatology.
Biases in responses to ambiguity may be particularly important during transitional stages of development, during which individuals experience profound changes and are challenged with navigating unfamiliar and ambiguous contexts (Silvers & Peris, 2023). One transitional stage that has been of particular interest to developmental scientists in recent years is the transition from adolescence to adulthood, often referred to as "emerging adulthood" (Arnett, Ž ukauskienė, & Sugimura, 2014;Swanson, 2016). Generally conceptualized as the time between late teenage years and twenties (Arnett et al., 2014), emerging adulthood presents a unique period of development characterized by exposure to novel and unfamiliar stressors (e.g., choosing a career, maintaining serious romantic relationships), instability and changes across several personal domains (e.g., place of residence, jobs, relationships), and feelings of being "in between" life stages (Arnett, 2000;Lanctot & Poulin, 2018). The challenges associated with emerging adulthood appear to be especially pronounced in 18-to 19-year-olds (Skulborstad & Hermann, 2016) and are associated with increased anxiety symptomatology (Arnett et al., 2014;Lanctot & Poulin, 2018;LeBlanc et al., 2020), highlighting the importance of identifying protective mechanisms during this early transitional window. Positive interpretations of ambiguity may promote resilience during the initial stages of emerging adulthood as individuals navigate the unfamiliar challenges and ambiguity associated with this developmental transition (Bardi et al., 2009;Silvers & Peris, 2023). In contrast, a tendency to perceive ambiguity as threatening may put early emerging adults at heightened risk for anxiety. However, it remains unclear how neurobehavioral responses to ambiguity relate to risk or resilience to anxiety during this formative developmental stage.
In a longitudinal, preregistered study (https://osf.io/9g5p2), we leveraged self-report, behavioral, and neuroimaging measures to investigate how neurobehavioral responses to ambiguity relate to concurrent and future anxiety in a group of emerging adults. We examined the transition to college, a stage that aligns with emerging adulthood for millions of youth in the United States (U.S. Census Bureau, 2018) and is marked by heightened risk for psychological distress (Bewick, Koutsopoulou, Miles, Slaa, & Barkham, 2010;Conley, Kirsch, Dickson, & Bryant, 2014;Sher, Wood, & Gotham, 1996). Our focus on this specific transition was partially motivated by prior work suggesting that responses to ambiguity predict wellbeing during the first year of college, but not later years (Bardi et al., 2009), perhaps due to the inherent uncertainty characteristic of this initial transitional period (e.g., moving away from home, starting new courses, navigating a new environment). In the current study, first-year college students (N = 101) completed a set of validated questionnaires assessing anxiety symptomatology and ambiguity tolerance. To evaluate trajectories of anxiety during the first year of college, we also examined anxiety levels across four additional timepoints (measured approximately one month apart). A subset of participants (N = 41) also participated in a neuroimaging portion of the study. During the neuroimaging session, participants passively attended to threatening (angry), nonthreatening (happy), and ambiguous (surprised) faces while undergoing functional magnetic resonance imaging (fMRI). After the scan, participants categorized each facial expression as positive or negative, thereby indicating their propensity to evaluate ambiguous faces more positively or negatively (i.e., valence biases). Using representational similarity analyses (RSA), we examined whether similarity in representations of threatening, nonthreatening, and ambiguous stimuli within the amygdala varied as a function of anxiety levels and whether trial-level similarity patterns predicted subsequent appraisals of individual stimuli.
We tested several preregistered hypotheses. We hypothesized that individuals who self-reported greater tolerance for ambiguity and exhibited more positive evaluations of ambiguity in the post-scan task would demonstrate lower baseline and future anxiety symptomatology (research question 1). We also predicted that, within subjects, trial-level similarity in representations of ambiguous and unambiguous images within the amygdala would predict subsequent appraisals of the ambiguous image in the post-scan task. Specifically, we hypothesized that greater representational similarity between ambiguous and threatening images would predict negative appraisals of the ambiguous image, whereas greater similarity between ambiguous and nonthreatening images would predict positive appraisals of the ambiguous image (research question 2). Lastly, we tested two distinct but complementary between-subject hypotheses (research question 3): (1) Individuals who on average demonstrate greater similarity in amygdala representations of ambiguous and threatening images would demonstrate higher baseline anxiety symptomatology and more pronounced within-subject increases in anxiety over time, whereas (2) those who demonstrated greater similarity in amygdala representations of ambiguous and nonthreatening images would exhibit lower levels of baseline and future anxiety.

Preregistration and data availability
The analysis plan and hypotheses for this study were preregistered on Open Science Framework (https://osf.io/9g5p2). All data and code (for tasks and analyses) are available on GitHub (https://github.com/ nsaragosaharris/anxiety_ambiguity_study).

Participants
We recruited first-year freshman college students (N = 101, 80 female; age range 18-19) via flyers and online recruitment for a longitudinal study with five timepoints of anxiety assessment and one (optional) fMRI scan session. Because we were interested in the transition from adolescence to adulthood during the first year of college, we only collected data from first year college students, which here does not include first year transfer students (who have previous college experience). Participants reported their race as 36.63% Asian, 28.71% Caucasian, 9.90% African American, 1.98% Native American Indian or Alaska Native, 10.89% multiracial, 7.92% other, and 3.96% declined to report (Supplemental Table 1a). Additionally, 24% of participants selfidentified as Hispanic or Latinx and 37.8% of participants reported being first-generation college students.
Based on prior neuroimaging studies that have included similar multivariate modeling techniques to the ones planned for this study (Connolly et al., 2012;FeldmanHall et al., 2018;Stolier & Freeman, 2016), the total planned sample size 100 participants for the lab session and 40 for the fMRI session. The final sample was 101 participants for the lab session and 41 (29 female) participants for the scanning session. Anxiety levels were measured via self-report at five timepoints (each measured approximately one month apart, with the exception of T3 and T4, which were spaced approximately two months apart due to pandemic-related changes to study protocols). Of the 101 participants who completed the baseline lab session, 77 completed all five time points (T1 to T5), four completed the first four timepoints (T1 to T4), 11 completed the first three timepoints (T1 to T3), one completed the first two timepoints (T1 and T2), and three completed only baseline timepoint (T1). Additionally, three participants completed T1, T2, T3, and T5 (i.e., missed T4), one participant completed T1, T4, and T5 (i.e., missed T2 and T3), and one participant only completed T1 and T3. The number of timepoints completed was not associated with any demographic variables (see Supplemental Table 1c for results from chi-square tests of independence), baseline anxiety levels (F(4,96) = 1.71, p = 0.16), or average anxiety levels (F(4,96) = 1.45, p = 0.22). For a summary and visualization of the number of observations at each timepoint, see Supplemental Table 1b and Supplemental Fig. 1.
Eligibility for inclusion in the study was assessed at two stages. First, for the 101 participants who completed the lab portion of the study, we determined eligibility based on the following criteria (assessed via brief in-person interview at the beginning of their session): (a) individuals ages 18-25 years in their freshman year of college; (b) no medical or psychiatric conditions contraindicating study participation (e.g., psychosis); (c) no current psychotropic medication; (d) no current treatment for anxiety or depression; and (e) willing to complete an fMRI session. Participants were then assessed for eligibility to participate in the fMRI session of the study. For the 41 participants who completed the fMRI portion of the study, we determined eligibility based on the following criteria (assessed via brief in-person interview and fMRI screening form): (a) no presence of metal in the body; (b) no current report of pregnancy; (c) no spontaneous report of pressing mental health concern requiring immediate follow up (e.g. psychosis) during the in person laboratory visit; and (d) no fear of enclosed spaces (claustrophobia).
Participants were required to complete the fMRI session within two weeks of their lab session. The participants selected to participate in the fMRI session were chosen based on whether their availability aligned with scanner availability in the two weeks following their session. The only exception is for several participants towards the end of recruitment, who were prioritized for scheduling based on their anxiety scores. In an effort to capture a wider range of anxiety scores in our fMRI subsample and make it more representative of our larger sample, we prioritized scheduling low anxiety participants for the fMRI session for a short period of time. Participants were compensated separately for the lab session, fMRI session, and at-home questionnaires. All participants received information about counseling services at the university at the beginning of their participation in the study. All recruitment, consenting, and data collection were completed in accordance with the requirements of the University of California Los Angeles Institutional Review Board (IRB# 19-001000).

Questionnaires
Self-reported anxiety. Anxiety was measured using the Screen for Adult Anxiety Related Disorders (SCAARED), a self-report questionnaire that has been validated in 18-to 27-year-olds and shows good internal consistency and discriminant validity between anxiety disorders and non-anxiety psychiatric disorders (Angulo et al., 2017). The SCAARED consists of 44 items assessing symptoms of panic disorder ("I get really frightened for no reason at all"), somatic symptoms ("When I get anxious, I feel dizzy"), separation anxiety ("I am afraid to be alone in the house"), social anxiety ("I feel nervous when I go to parties, dances, or any place where there will be people that I don't know well"), and generalized anxiety ("I worry about things working out for me"). Participants indicated their level of agreement on a three-point Likert scale (0 = very true or hardly ever true; 1 = somewhat true or sometimes true; 2 = very true or often true). Total scores, which can range from 0 to 88, reflect overall anxiety symptomatology, with greater scores indicating more anxiety symptoms (Angulo et al., 2017). Total scores on the SCAARED were used to assess anxiety symptoms at five time points, each assessed approximately one month apart. Participants completed the SCAARED in the lab at T1 (baseline) and at home for T2, T3, T4, and T5. Internal reliability for the SCAARED (measured from T1 responses) was high (Cronbach's α = 0.94, 95% CI [0.92, 0.95]). Descriptive statistics for baseline anxiety (SCAARED) scores are reported in Supplemental Table 1a.
Self-reported ambiguity tolerance. Self-reported ambiguity tolerance was measured using the Multiple Stimulus Types Ambiguity Tolerance Scale-II (MSTAT-II), a 13-item self-report questionnaire that has been validated in adults (McLain, 2009). The MSTAT-II assesses reactions to five types of ambiguous stimuli: generally ambiguous stimuli ("I prefer a situation in which there is some ambiguity"), complex stimuli ("I enjoy tackling problems that are complex enough to be ambiguous"), uncertain stimuli ("I find it hard to make a choice when the outcome is uncertain" (reverse scored)), new/unfamiliar/novel stimuli ("I generally prefer novelty over familiarity"), and insoluble/illogical/irreducible/internally inconsistent stimuli ("I try to avoid problems that don't seem to have only one 'best' solution" (reverse scored)). Participants rated their level of agreement with each statement on a five-point Likert scale (1 = strongly disagree, 2 = disagree, 3 = neutral, 4 = agree, 5 = strongly agree). Total scores can range from 13 to 65. Scores across all items were totaled, with greater scores indicating greater tolerance for ambiguity (McLain, 2009). This questionnaire was only administered at the baseline lab session (T1). Internal reliability for the MSTAT-II was good (Cronbach's α = 0.81; 95% CI [0.74, 0.86]). Descriptive statistics for self-reported ambiguity tolerance (MSTAT) scores are reported in Supplemental Table 1a.

fMRI task and preprocessing
fMRI paradigm. A subset of participants (N = 41) completed the fMRI portion of the study. The fMRI task was coded in PsychoPy2 (Peirce et al., 2019). Task stimuli were drawn from the racially diverse affective expression (RADIATE) dataset, a large, open-access dataset of facial expressions of adult models (Conley et al., 2018). The fMRI task included facial expressions from 99 unique actors (22 Asian actors, 32 Black/African American actors, 19 Hispanic or Latinx actors, and 26 White actors; 50 female, 49 male). Each of the 99 actors had three unique images, each with a different facial expression: angry, happy, and surprised (297 face images total). Angry, happy, and surprised faces were considered the threatening, nonthreatening, and ambiguous stimuli, respectively (Feldmann-Wüstefeld et al., 2011;Marsh et al., 2005;Neta et al., 2020;Rohner, 2002;VanTieghem et al., 2017). A blurred image was created by superimposing all of the 297 face images from the task and was presented on 27 trials as an attentional control.
The fMRI task (Fig. 1a) was an event-related design (no blocks) presented across three runs. Each trial required participants to view a single face image (threatening, nonthreatening, or ambiguous) for 500 ms. Each run consisted of 108 trials: 33 threatening, 33 nonthreatening, 33 ambiguous faces, and 9 blurred images (324 total trials). Every actor was presented once per run, for a total of three presentations during the experiment, each time with a different facial expression (threatening, nonthreatening, or ambiguous). The order of the runs was counterbalanced across participants and within-run trial order was randomized. A jittered fixation cross was presented between stimulus trials. Jitter times were created in OptSeq2 (https://surfer.nmr. mgh.harvard.edu/optseq/) to have a mean length of about 3 seconds and a range of 1.5-8 seconds.
Participants looked at the images while in the scanner. To help sustain attention, they were asked to press any button on the button box when they saw the blurred image (attentional control). To ensure they knew which picture to respond to, they were shown the blurred image during the instructions. This was the only response required of participants in the fMRI task. fMRI acquisition. Data were acquired on a 3 T Siemens Magnetom Prisma scanner. Functional data were acquired using the following parameters: voxel size = 2.0 × 2.0 × 2.0 mm, slices = 60 (interleaved), slice thickness = 2.0 mm, repetition time (TR) = 1000 ms, echo time (TE) = 37 ms, flip angle = 60 • , field of view = 208 mm, multiband acceleration = 6x. AutoAlign was used to position and align slices. Structural images were acquired using a high-resolution MPRAGE sequence (voxel size = 0.8 × 0.8 × 0.8 mm; TR = 2400 ms, echo time = 2.22 ms, field of view = 256 mm, slice thickness = 0.8 mm, 208 slices).
fMRI preprocessing. fMRI data were preprocessed with FEAT (FMRI Expert Analysis Tool) Version 6.00, part of FSL (FMRIB's Software Library, www.fmrib.ox.ac.uk/fsl). Functional data was registered to participants' high resolution structural images using boundary based registration (BBR) (Greve & Fischl, 2009). High resolution structural images were registered to standard space (MNI 2.0 ×2.0 ×2.0 mm stereotaxic space) with 12 degrees of freedom using FLIRT (FMRIB's Linear Image Registration Tool) (Jenkinson et al., 2002;Jenkinson & Smith, 2001) and FNIRT nonlinear registration (Andersson et al., 2007a(Andersson et al., , 2007b. Preprocessing included motion correction using MCFLIRT (Jenkinson et al., 2002;Jenkinson & Smith, 2001) using 24 standard and extended regressors, non-brain extraction using BET (Brain Extraction Tool) (Smith, 2002), grand-mean intensity normalization, and high-pass temporal filtering (Gaussian-weighted least-squares straight line fitting, with sigma = 50.0 s). To account for head motion, individual volumes with a framewise displacement greater than 0.9 mm were included as regressors (spike regressors created using 'fsl_motion_outliers'; Siegel  (happy) faces. Attention check trials included a blurred image and required a button box response. In the subsequent behavioral categorization task (B), participants categorized a subset of the previously seen threatening (not depicted), nonthreatening, and ambiguous as "feels good" or "feels bad" outside of the scanner. Trial order for both tasks was randomized (actors are shown in the same order between the two tasks for illustration only). et al., 2014). Our a priori inclusion criteria required that to be included in analyses, a given run could not have more than 10% of volumes exceed this framewise displacement. However, no run exceeded this threshold, and thus no participants were excluded due to motion. In line with similar MVPA and RSA work (Harry et al., 2013;Jin et al., 2015;Lee et al., 2020;Liang et al., 2017;Tashjian et al., 2019) and current recommendations Weaverdyck et al., 2020), in order to maintain fine-grained spatial details across voxels for RSA we did not apply spatial smoothing. Time-series statistical analysis was carried out using FILM (FMRIB's Improved Linear Model) prewhitening with local autocorrelation correction (Woolrich et al., 2001).
Our preregistration specified that runs in which a participant responded to fewer than 5 of the 9 of the attention check trials (responses to superimposed images) would be removed from analyses, and any participants with fewer than two usable runs (based on this response criterion) were to be excluded. However, no participant had more than one run with less than 5 responses to catch trials and all participants had at least 2 usable runs. Given recent work demonstrating attenuated signals in subcortical regions during multiband sequencing (Srirangarajan et al., 2021), univariate BOLD responses to ambiguous, threatening, and nonthreatening were examined to ensure the stimuli evoked detectable signal within the amygdala (i.e., beta values were significantly greater than zero) but were not analyzed further. One-sample t-tests confirmed that univariate responses within the right and left amygdala to all trial types (ambiguous, threatening, and nonthreatening) were significantly greater than zero (all ps < 0.001).
Amygdala masks. Binarized masks of the right and left amygdala were created based on FSL's Harvard-Oxford atlas and were thresholded in MNI space using Harvard-Oxford's probabilistic masks, which specify the probability that a given voxel falls within the specified brain region. Both masks were thresholded at p = 0.50 based on a similar RSA study (Tashjian et al., 2019) and visual inspection of anatomical alignment. Both masks were originally created in MNI space and transformed into native space in FSL for each participant prior to multivariate (RSA) analyses. Masks were applied to estimates within run (from first level models; see fMRI analysis) in order to keep the data in subject-specific native space. All analyses (for trial-level and condition-level analyses; see below) were conducted in subject-specific native space. Because masks were participant-specific and in native space, there was variability in amygdala size. In cases in which an amygdala-based statistical estimate was significant, we conducted sensitivity analyses in which we controlled for the number of voxels within the amygdala to ensure that differences in amygdala size across participants did not affect statistical estimates.

fMRI analysis
Single-trial (actor-level) representational similarity analyses. We were interested in whether single-trial, multivariate BOLD responses predicted subsequent appraisals of ambiguous stimuli. First level models in FSL were used to model single-trial activation patterns for every face trial (N = 297) within subjects using least squares-single (LSS) methods. In LSS models, a separate GLM estimates the voxel-wise activation pattern for each trial (Mumford et al., 2012(Mumford et al., , 2014. Each single-trial general linear model (GLM) included one regressor for a single trial of interest and three regressors for all other remaining trials (e.g., 'nuisance regressors') within that run, which were modeled in separate regressors by emotion category (i.e., threatening, nonthreatening, or ambiguous; Mumford et al., 2014). Blurred face trials were modeled but not further analyzed. Fixation screens between trials served as implicit baseline (i. e., were not explicitly modeled). Temporal derivatives for all regressors were included as covariates. Regressors were modeled using a double-gamma hemodynamic response function (HRF). Because every actor (N = 99) was shown across three trials (once per expression type: threatening, nonthreatening, and ambiguous; each presented in a different run), for a given participant, there were 3 first level models for each actor (297 total).
Using the Python package 'nilearn' (Abraham et al., 2014), for every participant we extracted vectors of voxel-level coefficients (z-transformed parameter estimates) from the first level models separately within the right and left amygdala masks. These vectors correspond to multivariate response patterns (or "representations") for a given trial. Within a given mask for a given participant, the number of vectors were thus equal to the number of first level models (3 trial-specific vectors for each of the 99 actors).
Within a given mask (right or left amygdala), pairwise Pearson correlations were computed between each of the three actor-matched vectors. This resulted in three correlations for each actor, corresponding to a participant's representational similarity (RS), over "overlap", in multivariate responses to a given actor's (1) ambiguous trial and threatening trial, (2) ambiguous trial and nonthreatening trial, and (3) threatening trial and nonthreatening trial ( Fig. 2A). Each participant had 297 Pearson correlation values (99 actors x 3 correlations) within both the right and left amygdala. Fisher's r-to-z transformation was then applied to the Pearson correlation values (Dimsdale-Zucker & Ranganath, 2018), with greater values indicating relatively greater RS in voxelwise patterns of activation between the two actor-matched trials. These actor-specific RS values served as the predictors in multilevel, single trial models (see Statistical modeling; Fig. 2a). The distributions of these actor-specific RS values within the right and left amygdala are provided in the supplement (Supplemental Figure 8).
Condition-level representational similarity analyses. In addition to the single-trial analyses (which corresponded to RS between conditions for a specific actor), we were interested in examining RS at the condition level as wellthat is, RS between a participant's responses to ambiguous, threatening, and nonthreatening trials, collapsed across all trials (i.e., actors) of a given condition. To index overall representations of ambiguous, threatening, and nonthreatening stimuli (collapsed across actors), we also examined multivariate responses by image type. For these estimates, we conducted separate first level (i.e., withinparticipant, within-run) models in FSL from the ones described previously. Using least squares all (LSA) methods, for each run of the task, first level models in FSL were used to extract BOLD response patterns separately by trial expression type (threatening, nonthreatening, or ambiguous). Each run-specific general linear model (GLM) included a regressor for each expression type. That is, all 33 trials of a given expression type (e.g., ambiguous) for a given run were estimated simultaneously in a single regressor per GLM. Blurred face trials were modeled but not further analyzed. Following the same procedures described in Single-trial (actor-level) representational similarity analyses, for every participant we extracted vectors of voxel-level coefficients (ztransformed parameter estimates) from the first level models separately within the right and left amygdala masks. Within each mask, each participant had three vectors per run, corresponding to their multivariate responses to ambiguous, threatening, and nonthreatening trials (Fig. 2B).
We then computed pairwise Pearson correlation values among the three vectors per run. These correlations were then averaged across runs and were Fisher-z transformed. This resulted in three Fisher-z transformed RS values (corresponding to ambiguous/threatening RS, ambiguous/nonthreatening RS, and threatening/nonthreatening RS) per participant per mask (Fig. 2B). A boxplot of these three Fisher-z values across participants is provided in the supplement (Supplemental Figure 7).

Post-scan behavioral categorization task
After the fMRI task, participants completed a behavioral task (Fig. 1B) outside of the scanner that was adapted from prior work investigating interpretive biases in youth (VanTieghem et al., 2017). In this task, participants viewed a subset of the images they previously saw in the fMRI task. To ensure an equal number of actors across 10 blocks, one additional actor was added to the fMRI set for a total of 100 actors in the behavioral set. Because surprised (ambiguous) faces were the primary interest for this study, participants viewed twice as many ambiguous faces. Of the 100 actors, 50 were only shown making a surprised face, and the other 50 were shown three times (once surprised, once happy, and once angry). In total, participants were shown 200 images (100 surprised) over ten blocks. Block order and within-block trial order was randomized.
Following the task design from VanTieghem et al. (2017), on each trial, participants were asked to indicate whether the person in the image "feels good" or "feels bad" by pressing a button on the keyboard (1 or 0, counterbalanced across participants). The stimulus was presented for 500 ms, and then text appeared on the screen for 1500 ms prompting their response (Fig. 2B). Early responses (during the initial 500 ms presentation screen) were accepted and included in analyses. The response screen lasted for 1500 ms regardless of their response time. After the response screen, there was a 200 ms fixation screen. In each of the ten blocks, participants were shown 10 surprised faces, 5 happy faces, and 5 angry faces. In between blocks, there was a ten second fixation screen.
Before the task, participants completed six practice trials (2 happy, 2 angry, and 2 surprised). The images in the practice trial were novel images from a separate dataset (NimStim; Tottenham et al., 2009). The practice included feedback to indicate if they responded quickly enough. The actual task included no feedback.

Behavioral data analysis
To index valence biases in the task, for each participant we computed the percent of ambiguous (surprised) faces categorized negatively (i.e., trials in which participants select the "feels bad" option). In this measure, greater scores indicate greater negative valence biases. Descriptive statistics for negative valence bias scores are reported in Supplemental  Table 1a. For single-trial analyses, positive appraisals ("feels good") were coded as 1 and negative appraisals ("feels bad") were coded as 0.

Fig. 2.
In the single-trial (actor-level) representational similarity analyses (A), we used single-trial estimates to extract representations (indicated by patterned boxes) within the amygdala for a given trial. Within subjects, we evaluated three representational similarity (RS) values (indicated by arrows) for every actor by correlating the representation patterns from the actor's ambiguous, threatening, and nonthreatening trials. We then tested whether these actor-specific RS values could predict how the participant categorized that actor's ambiguous face in the post-scan task (face shown on the same screen as rating for illustrative purposes only). In the condition-level representational similarity analyses (B), all trials of a given type (e.g., ambiguous) were modeled together per run to index a participant's overall response pattern to each stimulus type (ambiguous, threatening, and nonthreatening). Within subjects, we evaluated representational similarity (RS) between these condition-level patterns (indicated by arrows). We tested whether, between subjects, these condition-level RS values related to baseline or longitudinal anxiety.

Statistical modeling
Between-subject analyses. We examined the relationships among anxiety, self-reported ambiguity tolerance, task-based negativity biases, and condition-level representational similarity (RS) values within the amygdala in between-subject analyses. Between-subject analyses were conducted using linear regression ('lm' function in R). Self-report and task-based measures were z-scored between participants in order to obtain standardized beta coefficients (denoted β). The full output from all statistical models are available in table format in the supplement.
Longitudinal modeling of anxiety over time. To test whether tolerance for ambiguity (based on self-report and behavior) predicted resilience to anxiety during the transition to adulthood, we examined how anxiety varied over time and whether these trajectories varied as a function of ambiguity tolerance. Although not all participants completed all five timepoints, the number of timepoints completed was not associated with any demographic variables (see Supplemental Table 1c for results from chi-square tests of independence), baseline anxiety levels (F(4,96) = 1.71, p = 0.16), or average anxiety levels (F(4,96) = 1.45, p = 0.22), so all responses were included in growth curve models to maximize power. We first tested a simpler model in which we examined fluctuations in anxiety over time. Using the 'lme4′ R package (Bates et al., 2015), we assessed trajectories of anxiety via a longitudinal growth model with time as the predictor and anxiety level (at each time point) as the outcome. Time was modeled as months since the day on which participants' baseline measures were collected (i.e., participant-specific baseline = month 0) to adjust for between-subject differences in the first day of study participation. Because trajectories of anxiety over time may vary between individuals, we tested whether the addition of random slopes (in addition to random intercepts) improved the fit of the model. A likelihood ratio test run in 'varTestnlme' (Baey & Kuhn, 2019) indicated that inclusion of both random intercepts and random slopes (i.e., allowing the effect of time on anxiety to vary between participants) improved the fit of the model relative to a simpler model that included random intercepts but no random slopes (LRT = 28, p < 0.01). Therefore, in analyses that included longitudinal anxiety scores, both random intercepts and slopes (for the time variable) were included in the models.
The COVID-19 pandemic and mandatory quarantine orders began during the longitudinal portion of the study. Because participants did not all begin the study on the same day, the beginning of quarantine occurred at different points in their study timelines (e.g., for some this occurred between their T2 and T3, whereas for others this occurred between T3 and T4). To account for this unexpected confound, we conducted an additional sensitivity analysis in which we controlled for when in a participant's study timeline (i.e., between which two timepoints) the quarantine began (defined as March 13th, 2020, the day the university canceled in-person instruction). In this model, time of quarantine onset relative to a participant's timeline (a between-subjects factor, corresponding to whether the quarantine began between their T1 and T2 timepoints, or between their T2 and T3 timepoints, etc.) was added as a fixed effect to the longitudinal growth model.
Single-trial (actor-level) modeling. Single-trial analyses examined whether representational similarity (RS) between actor-matched trials of different types (ambiguous, threatening, or nonthreatening) predicted subsequent appraisals of that actor's ambiguous image in the post-scan task. The Fisher-z transformed amygdala RS values (ambiguous/ threatening, ambiguous/nonthreatening, threatening/nonthreatening), corresponding to a participant's "overlap" in representations between two actor-matched trials ( Fig. 2A), were modeled as predictors in separate multilevel logistic regressions in R using the 'lme4′ package (Bates et al., 2015). Within each region of the amygdala, one multilevel model was run for each overlap type (ambiguous/threatening, ambiguous/nonthreatening, threatening/nonthreatening; Fig. 2A). For instance, in multilevel models examining ambiguous/threatening similarity, the predictor was the ambiguous/threatening overlap value for a given actor and the outcome was the post-scan appraisal (positive or negative) of that actor's ambiguous trial. Actor-specific RS values and corresponding responses were nested within participants. The same process was followed for the ambiguous/nonthreatening RS values and the threatening/nonthreatening RS values (separately in right and left amygdala). In each of the models, post-scan appraisals of a given actor's ambiguous trial (positive or negative) was the outcome variable. These multilevel logistic models allowed for random intercepts for each participant. We hypothesized that, within participants, a greater ambiguous/threatening RS value for a given actor would be associated with greater likelihood of a negative appraisal of that actor's ambiguous image ( Fig. 2A). In the same reasoning, we predicted that a greater ambiguous/nonthreatening RS value for a given actor would be associated with greater likelihood of a positive appraisal of the ambiguous image.

Descriptive statistics
Descriptive statistics detailing participant demographics, baseline anxiety scores (T1 SCAARED scores), self-reported ambiguity tolerance scores (MSTAT-II scores), and negative valence biases (percent of ambiguous faces interpreted negatively in the post-scan task) are presented in Supplemental Table 1a. The distributions of negative valence biases across participants and the percent of negative interpretations of each actor are provided in Supplemental Fig. 4 and Supplemental Fig. 5. Analyses of average reaction time by condition type are also provided in the supplement. Visualizations of average reaction times by expression type and by actor in the post-scan task are provided in Supplemental Fig. 3b and Supplemental Figure 6. To ensure participants understood the post-scan task, we also examined accuracy on the threatening and non-threatening (i.e., unambiguous) trials. Results from a one-tailed ttest confirm that accuracy for both angry (mean accuracy = 0.93; t (40) = 31; p < 0.001) and happy (mean accuracy = 0.96; t(40) = 72; p < 0.001) trials were significantly above chance performance (50%), such that participants correctly rated angry facial expressions negatively and happy facial expressions positively (Supplemental Table 6a; Supplemental Fig. 3a). This confirms that these two types of facial expressions were clearly valenced (i.e., unambiguous) and that participants understood the task instructions. One participant had chance performance on the angry trials in the post-scan task, which appeared to be due to a decline in performance partway through the task. In a sensitivity analysis, this participant was excluded from all analyses involving data from the post-scan task. All findings reported here held when this participant was excluded from analyses. Results from the sensitivity analysis can be found in the supplement.

Longitudinal changes in anxiety over time
Across participants and timepoints, anxiety (SCAARED) scores ranged from 0 to 85 overall with an average of 32. Within participants, anxiety scores appeared to be relatively stable, with an average withinperson standard deviation in scores of 6.1 (range = 1.5-16.8). Results from longitudinal growth models demonstrated no significant fixed effect of time on anxiety scores (β = − 0.01, 95% CI [− 0.47,0.45], t (80.67) = − 0.03, p = 0.97), likely due to the relative within-participant stability in anxiety scores across the study period (Supplemental Fig. 2). There was also no significant effect of time on anxiety scores after controlling for the onset of quarantine relative to a participant's study timeline (β = 0.001, 95% CI [− 0.46,0.46], t(80.57) = 0.01, p = 0.995).
Research question 1:. Do self-report or behavioral responses to ambiguity predict anxiety?
Self-reported ambiguity tolerance and changes in anxiety over time. We hypothesized that individuals who self-reported lower ambiguity tolerance at baseline would exhibit greater increases in anxiety over time. We tested a longitudinal growth model with fixed effects of self-reported ambiguity tolerance, time, and an interactive term (self-reported ambiguity tolerance x time) to test whether the effect of time on anxiety differed as a function of baseline self-reported ambiguity tolerance. Contrary to our hypothesis, self-reported ambiguity tolerance and time did not interact to predict anxiety; that is, trajectories of anxiety over time did not significantly differ as a function of baseline self-reported ambiguity tolerance (β = 0.20, 95% CI [− 0.25,0.64], t(86.75) = 0.90, p = 0.37).
Negative valence biases in post-scan task and longitudinal anxiety. We hypothesized that individuals who exhibited greater negative valence biases in the post-scan task would exhibit greater increases in anxiety over time. We tested a longitudinal growth model with fixed effects of negative valence biases, time, and an interactive term (negative valence biases x time) to test whether the effect of time on anxiety differed as a function of negative valence biases. Although the relationship was in the expected direction, wherein greater negative valence biases predicted more positive slopes in anxiety, this effect was only trending (β = 0.72, 95% CI [− 0.02, 1.46], t(36.61) = 1.91, p = 0.06).
Research question 2:. Does representational similarity between single trials predict later appraisals of ambiguous stimuli?
Single-trial (actor-level) representational similarity in the amygdala and behavioral categorizations of ambiguous stimuli. Single-trial analyses examined whether representational similarity (RS) between actormatched trials of different types (ambiguous, threatening, or nonthreatening) predicted subsequent appraisals of that actor's ambiguous image in the post-scan task. Within the right amygdala, neither ambiguous/threatening, ambiguous/nonthreatening, nor threatening/ nonthreatening RS values predicted appraisals (Supplemental Tables 20  to 22.) Similarly, within the left amygdala, neither ambiguous/ threatening nor ambiguous/nonthreatening RS values predicted appraisals of the actor's ambiguous image (Supplemental Tables 20 and  21). However, we did observe a significant effect within the left amygdala: Greater overlap between actor-matched threatening and nonthreatening trials were associated with greater likelihood of a positive appraisal of that actor's ambiguous image (OR = 2.28,95% CI [1.18,4.41], z = 2.50, p = 0.01; Fig. 4). This association remained unchanged after controlling for participant-specific left amygdala size (OR = 2.30,95% CI [1.18,4.41], z = 2.45, p = 0.01). This suggests that, within participants, greater similarity in representations of an actor's unambiguous images predicted more positive appraisals of that actor's ambiguous image.
Condition-level representational similarity in the amygdala and negative valence biases. We next examined whether representational similarity (RS) among ambiguous, threatening, or nonthreatening images predicted negative valence biases (i.e., percent of surprised faces categorized negatively) in the post-scan task. Neither ambiguous/threatening, ambiguous/nonthreatening, nor threatening/nonthreatening RS values within the right or left amygdala were associated with negative valence biases (Supplemental Tables 17 to 19).
Research question 3:. Does representational similarity in patterns of amygdala responses to ambiguous, threatening, and nonthreatening images relate to anxiety?
Condition-level representational similarity in the amygdala and baseline anxiety. We hypothesized that individuals who exhibit greater condition-level RS between ambiguous and threatening images would report higher anxiety, whereas those who exhibit greater condition-level RS between ambiguous and nonthreatening images would report lower anxiety. Contrary to our hypothesis, RS between ambiguous and threatening stimuli within the right and left amygdala did not predict baseline anxiety (Supplemental Table 11). Similarity in representations of threatening and nonthreatening stimuli were also not significantly related to baseline anxiety (Supplemental Table 13 Condition-level representational similarity in the amygdala and longitudinal anxiety. We hypothesized that between-subject differences in condition-level RS values would relate to between-subject differences in anxiety trajectories over time. Specifically, we hypothesized that individuals who exhibited greater RS in multivariate amygdala responses to ambiguous and threatening images (at the condition level) would exhibit greater increases in anxiety over time, whereas those who exhibited RS between ambiguous and nonthreatening images would demonstrate attenuated anxiety over time. We tested longitudinal growth models with fixed effects of condition-level RS values (Fisher-z scored), time, and an interactive term (condition-level RS x time) to test whether the effect of time on anxiety differed as a function of pattern similarity values. Similarity in representations of ambiguous and threatening, ambiguous and nonthreatening, and threatening and nonthreatening stimuli within the right and left amygdala was not associated with anxiety slopes (i.e., did not significantly interact with time to predict anxiety levels; Supplemental Tables 14 to 16).

Discussion
In the current longitudinal study, we integrated self-report, behavioral, and neuroimaging measures to examine how responses to ambiguity relate to concurrent and future anxiety in a group of emerging adults. Leveraging multivariate pattern analysis to characterize representations of ambiguity, we found that individuals who evidenced greater similarity in representations of ambiguous and nonthreatening images within the left amygdala reported lower baseline anxiety. Moreover, individuals who self-reported greater tolerance for ambiguity also reported fewer anxiety symptoms at baseline. Our self-report and neuroimaging findings highlight the association between ambiguity processing and resilience to anxiety during the transition to adulthood. In doing so, this work provides support for longstanding theories of anxiety (Aikins & Craske, 2001;Eysenck et al., 1987;Mathews et al., 1997) while offering novel insights into how neural representations of ambiguity may promote resilience to psychopathology.

Self-reported ambiguity tolerance is associated with lower anxiety in emerging adults
Research in groups undergoing stressful transitions, including individuals transitioning to college (Bardi et al., 2009) and beginning medical internships (Kleim et al., 2014), demonstrates that greater tolerance for ambiguity predicts resilience to psychopathology during transitional periods. In line with these findings, we found that greater self-reported tolerance for ambiguity was associated with fewer baseline anxiety symptoms in 18-to 19-year-olds transitioning to college. These results both replicate prior work by linking ambiguity tolerance to wellbeing during a transitional stage of development and further support the clinical relevance of ambiguity tolerance in anxiety specifically (Eysenck et al., 1987;Mathews et al., 1997). Practically, greater ambiguity tolerance likely promotes healthy exploration and adaptive responses to the ambiguous challenges characteristic of this developmental stage (Romer et    examine whether ambiguity tolerance promoted resilience across the transition period, we assessed whether baseline ambiguity tolerance predicted differences in anxiety trajectories over time. However, we did not observe an association between baseline ambiguity tolerance and anxiety trajectories, likely due to the notable within-subject stability in anxiety scores over the five timepoints. Future work employing methods better suited to capture within-person fluctuations in anxietyperhaps by using ecological momentary assessment to capture day-to-day fluctuations in mood (Heller et al., 2019;Puccetti et al., 2020) -may elucidate the role of ambiguity tolerance in anxiety progression. Furthermore, future longitudinal studies may benefit from sampling individuals in "pre-transition" stages (i.e., graduating high school seniors), as it is possible that the greatest within-person increases in anxiety occurred in between the pre-transition and early-transition phases. It will also be important for future work to include a more gender diverse sample, as the imbalance towards female-identifying participants in the current sample limits generalizability of the findings.
Contrary to our hypothesis and prior work (Park et al., 2016;Van-Tieghem et al., 2017), we did not observe an association between anxiety levels and interpretations of ambiguity in the post-scan behavioral task. Although negative valence biases in anxiety are well-documented (Chen et al., 2020;Hirsch et al., 2016), it is possible that the responses elicited by the post-scan task were unable to fully capture these biases. The post-scan task was adapted from a child-friendly task that asks whether the person in the picture "feels good" or "feels bad" (VanTieghem et al., 2017), a categorization that may reflect emotion perception rather than the threat-related interpretations often assessed in studies of anxiety-related interpretive biases (Hirsch et al., 2016). Further work investigating how valence biases shape behavior and anxiety risk, particularly during the transition to adulthood, is merited.

Greater similarity in representations of ambiguous and nonthreatening affective cues relates to lower anxiety
Our primary research aim was to examine how neural representations of ambiguity within the amygdala relate to anxiety symptomatology. Leveraging representational similarity analysis (RSA) within the amygdala, we examined whether condition-level representational similarity (RS) among ambiguous, threatening, and nonthreatening cues varied as a function of anxiety. We tested two related but distinct possibilities: That (1) greater RS between ambiguous and threatening cues would be associated with greater anxiety, whereas (2) greater RS between ambiguous and nonthreatening cues would be associated with lower anxiety. Our results support the second hypothesis, such that greater RS between ambiguous and nonthreatening cues within the left amygdala correlated with lower baseline anxiety. Although we did not have a priori hypotheses regarding laterality, the observed specificity of this association to the left amygdala is consistent with prior work demonstrating that anxiety is linked to activity within the left, but not right, amygdala in response to ambiguous threat (Im et al., 2017).
The observed heightened RS between ambiguous and nonthreatening stimulispecifically, explicitly positive, happy facesin lowanxiety individuals could indicate that anxious and non-anxious individuals specifically differ in their tendency to represent ambiguity positively. Intriguingly, this dovetails with prior behavioral work demonstrating that interpretive biases in anxiety may stem from a decreased tendency to form positive interpretations of ambiguity, rather than an increased proclivity to form negative interpretations (Constans et al., 1999;Jopling et al., 2020). In fact, some work suggests that healthy controls demonstrate overly positive interpretations of ambiguous information (potentially seeing the world through "rose-colored glasses"), whereas clinical samples do not show these biases (Hirsch & Mathews, 2000). Moreover, given research suggesting that rapid, initial responses to ambiguity tend to be negative (Neta & Tong, 2016;Neta & Whalen, 2010), greater RS between ambiguous and nonthreatening, positive stimuli may be indicative of a less common phenotypepotentially one that promotes resilience during stressful periods of development. However, it is important to note that condition-level RS was not associated with negativity biases in the post-scan task. While our results cannot speak to the behavioral manifestations of these multivariate representations, this presents a promising avenue for future work. An important next step would be to examine multivariate representations of ambiguity during the interpretation stage in which participants make explicit judgments about the ambiguous stimuli. Given work suggesting that individuals engage in emotion regulation when evaluating ambiguous faces (Neta et al., 2022), this work could provide insight into how regulatory mechanisms may shape negative valence biases in anxiety.
In the current study, we chose to focus on representations within the amygdala based on its role in responding to motivationally salient cues (Phelps & LeDoux, 2005), its functional association with anxiety , and prior research using similar multivariate techniques to interrogate representations of ambiguity (FeldmanHall et al., 2018;Tashjian et al., 2019). However, the amygdala is only one structure within a network of regions that facilitate detection and appraisal of affective and threat-related cues (Grupe & Nitschke, 2013). Future work could benefit from leveraging paradigms capable of capturing how neurobehavioral responses unfold over time as individuals appraise ambiguous cues Neta & Tong, 2016). Such work could elucidate, for instance, whether multivariate representations of ambiguous cues within the amygdala can be shaped by prefrontal regulatory mechanisms, or whether the tendency to gradually shift toward more "positive" representations when formulating appraisals (Petro et al., 2018) varies as a function of anxiety. That said, it is notable that we observed anxiety-related patterns during rapid, uninstructed viewing of ambiguous images, particularly given the role of automatic, unintentional biased processing of threat-related information in anxiety (Teachman et al., 2012). By interrogating these processes in the absence of explicit regulation, these findings shed light on how initial representations of ambiguity relate to differential risk for anxiety.

Trial-level representational similarity predicts subsequent appraisals of ambiguous stimuli
To investigate whether trial-level representations predict subsequent appraisals of ambiguous cues, we examined whether RS between actormatched trials predicted post-scan interpretations of a given actor's ambiguous image. Surprisingly, we found that greater RS between an actor's two unambiguous trials (threatening and nonthreatening facial expressions) predicted more positive interpretations of that actor's ambiguous facial expression. Intriguingly, these findings align with a between-person observation in Tashjian et al. (2019), wherein adolescents who exhibited greater RS between trustworthy and untrustworthy facial expressions within the amygdala were more likely to rate ambiguous images as trustworthy. That said, given that our finding operates at the single-trial level, rather than capturing between-person averages, it is difficult to consolidate these similar patterns of results. Additionally, to minimize demand characteristics, participants were not told to make any judgments about the images or informed that they would be doing so following the scan. This enabled us to interrogate rapid, initial responses to ambiguous stimuli in the absence of prior learning or information, but precluded interrogation of neural representations during the stage in which participants explicitly appraised the stimuli. It thus remains unclear to what extent these initial multivariate representations of ambiguity determine the degree to which a given ambiguous cueor a given actoris perceived as threatening. Rather, these findings more generally suggest that the amygdala functionally encodes features of stimuli that are relevant for appraisals of those stimuli under conditions of ambiguity.

Conclusions
Over the past two decades, clinical psychology and developmental neuroscience research have highlighted the transition to adulthood as a key developmental period for understanding risk and resilience to psychopathology generally (Schulenberg & Zarrett, 2006) and anxiety specifically (Spielberg et al., 2019). Our results from a sample of first-year college students demonstrate a relationship between anxiety and responses to ambiguity during the initial transition to adulthood. Identification of protective factors that promote successful adaptation during this period has crucial implications for clinical work (Arnett et al., 2014;LeBlanc et al., 2020). While the long-term consequences of neural representations of ambiguity remain unclear, these results suggest that the amygdala rapidly and functionally encodes information about the valence of ambiguous affective stimuli, and that these initial representations vary with anxiety symptomatology. Together, these findings demonstrate that multivariate representations of affective cues can provide novel insights into resilience to psychopathology during transitional stages of development.

Declaration of Competing Interest
The authors report no conflicts of interest.

Data availability
The data and task and analysis code are available on GitHub (https:// github.com/nsaragosaharris/anxiety_ambiguity_study).