Introduction

Every day, multiple times a day, people must make decisions about what to eat. A growing body of research suggests that these decisions involve considerations of different attributes of the available food options (e.g., healthiness and tastiness), and that these attributes often present the decision-maker with challenging conflicts (e.g. “Should I choose the healthier, less delicious food or the better tasting, less healthy food?”). Difficulties resolving these conflicts can take on pathological forms in individuals with eating disorders, and some eating disorders are characterized by food choices that are biased toward one extreme (e.g., toward subjectively less tasty, but healthier, low-fat foods in anorexia nervosa; (Foerde et al., 2015)). However, food choices seem to episodically oscillate between extremes in the case of bulimia nervosa (BN), which is commonly marked by prolonged periods of rigid dietary restriction punctuated by out-of-control binge eating of highly palatable foods. Although subclinical forms of this oscillation are ubiquitous in examples of “yo-yo dieting” and “dietary lapses,” surprisingly little cognitive neuroscience research to date has attempted to explain it.

Self-report data suggest that rising negative affect is one reliable predictor of the switch from more to less restrictive food choices (Konttinen et al., 2010). Although some individuals reduce intake in response to stress (Hill et al., 2022), research using both animal models and healthy adult samples show that stress can precipitate the overconsumption of highly palatable foods (Dionysopoulou et al., 2021; Jacques et al., 2019; Tomiyama, 2019). This connection between negative emotion and food choice is even more pronounced in binge eating. Inducing negative affect in the laboratory environment disproportionately increases food intake among individuals with binge eating (Telch & Agras, 1996). Ecological momentary assessment data show that binge eating in the natural environment is also much more likely to occur in the context of increasing and unstable negative emotions (Alpers & Tuschen-Caffier, 2001; Berg et al., 2013; Haedt-Matt & Keel, 2011; Hilbert & Tuschen-Caffier, 2007; Smyth et al., 2007). In addition, reactivity to negative emotions, or negative urgency, has been shown to predict or correlate with more frequent binge eating in individuals with BN (Fischer et al., 2013; Schnepper et al., 2021).

Although momentary increases in negative emotions are closely tied to the overconsumption of highly palatable foods, the cognitive processes underlying this tight link, particularly in the case of binge eating, are not well understood. Studies to date have failed to find support for the notion that increased negative affect, specifically stress, impairs inhibitory control processes in healthy adults (Allen et al., 2022) or in BN (Dreyfuss et al., 2017; Westwater et al., 2021). While another study found that healthy adults and individuals with BN differed in how stress impacted visual processing-related neural activity, they failed to find support for the hypothesis that increased negative affect abnormally increased limbic neural responses to palatable food stimuli in BN (Collins et al., 2017).

We recently tested possibility: that negative affect abnormally impacts decision-making in BN by promoting food choices that include the high-fat, appetitive foods that are typically consumed during binge-eating episodes (Gianini et al., 2019). Individuals with BN and healthy controls completed a food choice task after a negative and neutral affect induction (Figure 1). Counter to expectations, negative affect did not influence ultimate food choices in either group, and women with BN showed restrictive food choices whether neutral or negative affect was induced: they chose low-fat foods more often (Figure 2A), and these choices were strongly predicted by healthiness ratings (Figure 2B). However, these analyses focused primarily on individuals’ ultimate choices and may have overlooked the possibility that even small changes in negative affect could influence more subtle aspects of the food decision-making process leading up to those choices.

Figure 1.

Figure 2.

These more subtle features can be quantified with computational cognitive modeling. Sequential sampling models leverage both choice and response time data to estimate parameters that capture individual differences in decision-making dynamics (Forstmann et al., 2016; Ratcliff & Smith, 2004). These models assume that decision-making involves the noisy accumulation of evidence over time until a predetermined threshold is reached, leading to a response. Therefore, both ultimate responses and response times are meaningful signals about the cognitive processes underlying decision-making. Though originally developed to describe perceptual, psychomotor, and memory tasks, these models can be applied to value-based decision-making tasks by incorporating participants’ ratings of different attributes for the stimuli presented in a trial. This allows researchers to examine how and when these attributes influence the decision-making process (Busemeyer et al., 2019; Clithero, 2018).

Research using sequential sampling models indicates that food choices are influenced not only by the weights assigned to attributes like healthiness and tastiness, but also by the time at which these attributes enter the decision-making process (Chen et al., 2022; HajiHosseini & Hutcherson, 2021; Hutcherson & Tusche, 2022; Maier et al., 2020; Sullivan et al., 2015; Sullivan & Huettel, 2021). Specifically, using time-varying sequential sampling models, studies have demonstrated that healthy individuals assign different subjective weights to healthiness and taste attribute ratings of foods (i.e., they have differing degrees to which they influence the evidence accumulation rate); that healthiness and taste attribute information can enter the decision process at different times; and that these relative weighting strengths and timing of attribute consideration have distinct influences on ultimate food choices (Maier et al., 2020). This notion is consistent with theories of both emotion- and food-craving regulation, which posit that there are separate attention deployment and stimulus valuation steps involved in these regulation processes (e.g., (Giuliani & Berkman, 2015; Gross, 2015; Han et al., 2018).

Notably, even if one of these decision parameters (e.g., weights on healthiness or tastiness, attribute onsets) biased individuals toward selecting a low-fat (or high-fat) food, the other parameter could compensate for it to ultimately generate the opposite response (Figure 3). For example, if taste information was considered before healthiness information, but healthiness information was weighted strongly enough, a decision-maker could choose a healthier, instead of a tastier, food (Figure 3B). These latent aspects of the decision-making process, either orthogonally or in parallel, could be sensitive to state changes like increases in negative affect and could be altered in BN. Specifically, these latent decision parameters could help explain why dysregulated consumption of tasty, high-fat foods in BN is more likely during periods of high negative affect, but restricted intake of lower-fat foods is more likely during periods of low negative affect. In one case, increasing negative affect could alter an individual’s attribute weights, either reducing their weight on healthiness (dashed blue line) or increasing their weight on tastiness (dashed red line), which would increase their likelihood of choosing binge-type foods that are considered to be highly palatable yet also unhealthy (Figure 3C). Alternatively, increasing negative affect could cause individuals with BN to consider healthiness information later in the decision-making process, giving taste information more time to bias choices towards binge-type foods (Figure 3D). Clarifying these mechanisms is critical for translational efforts because data suggest that the weighting of attributes and the time at which they are considered have different neural substrates (Maier et al., 2020), and the treatment offered to patients should depend on which scenario more accurately reflects their pathology.

Figure 3.

Here, we applied a diffusion decision model (DDM) with a time-varying drift rate to food choice and response time data from our controlled, randomized crossover study in women with BN and healthy controls (Figure 1A). This model allowed us to investigate the potential exaggerated influence of negative affect on the dynamics of food-specific decision-making in BN. Specifically, we tested how food type (low-fat/high-fat) and affective state (neutral/negative) influenced decision-makers’ attribute weights and onsets.

Our hypotheses were derived from the clinical phenomenology underlying BN. As outlined above, individuals with BN seem to vacillate between unstable decision dynamics: many individuals with the disorder maintain a relatively restrictive diet until negative emotional states induce binge-eating episodes and subsequent compensatory behaviors (i.e., purging). Given this instability, we predicted that our BN participants’ food choices would be influenced by components of the decision-making process whose opposing dynamics could be disrupted by increasing negative affect.

We first hypothesized that in states of neutral affect and for high-fat foods, one of these opposing dynamics would compensate for the other to promote ultimate choices that are restrictive. This could be achieved by either a) weighing taste information more strongly but considering it later than healthy controls, or b) weighing tastiness less strongly but considering it earlier than healthy controls. Second, we hypothesized that, in BN, increasing negative affect would abnormally impact at least one of these latent aspects of decision-making to increase the bias towards high-fat foods. Even though prior analyses indicate that high-fat foods were not ultimately chosen more often, we predicted that after a negative affect induction, individuals would either a) be slower to consider healthiness information, but still weight taste information just as strongly, or b) they would put less weight on healthiness information, but still consider taste information sooner.

Because these combinations require one feature of the decision process to compensate for another, they would result in a much less stable dynamic compared to one in which information weights and onset times are concordant. For example, when healthiness information is considered sooner and is more strongly weighted than taste information, foods low in healthiness would almost always be rejected. In contrast, when healthiness information is considered later but is more strongly weighed than taste information, food choice behavior can be more easily perturbed with state-based changes in weights or the delay in attribute onsets. Here, we show evidence for these unstable decision dynamics that can account for a bias in BN toward low-fat, healthier foods in states of relatively low negative affect, but toward high-fat, tastier foods in states of high negative affect. Although negative affect delayed the onset of the consideration of healthiness information among both HC and BN participants, this effect was more pronounced in the BN group. Ultimately, we found that although individuals with BN consistently weighed healthiness information more strongly, negative affect delayed its entry into the evidence accumulation process. With this longer delay, taste information has more time to bias decision-makers towards high-fat food choices before healthiness information can come online.

Results

Negative affect aberrantly delays consideration of healthiness information among women with BN

We analyzed data from individuals with BN (n = 25) and healthy controls (HC, n = 21) who completed a computerized Food Choice Task following negative and neutral affect inductions, across two visits (Figure 1). Mood was assessed before and after the affect inductions using the Profile of Mood States (POMS; (McNair et al., 1989)). Negative affect scores were calculated by summing the five negative subscale scores (e.g., anger, confusion, depression, fatigue, and tension) and subtracting the positive subscale score (e.g., vigor). Negative affect scores increased significantly for both groups following the negative affect induction (Gianini et al., 2019, reproduced in Table S1).

To quantify the specific computations performed during Food Choice Task decision-making, we used a time-varying diffusion decision model (DDM; (Ratcliff & McKoon, 2008)). This time-varying model introduces a starting time parameter that indicates the delay with which different attributes enter the evidence accumulation process (Chen et al., 2022; Lombardi & Hare, 2021; Maier et al., 2020). Using this starting time DDM (stDDM), we can determine how individuals weigh healthiness versus taste information and when those attributes influence food-related decision-making in BN and HC groups in each affective state. The stDDM was fitted to data that included individual-level response times, choices, z-scored healthiness and tastiness ratings, and affect induction.

We specified a stDDM where individual parameters were drawn from four separate hyperparameters that varied both by group (BN or HC) and by condition (neutral or negative affect induction). To test our hypotheses about the interactions between affect and food type (low-fat or high-fat) on food choice, we allowed each attribute weight parameter (ω_Taste and ω_Health) and the relative-starting time parameter (τ_S) to vary as a function of food type. The drift rate determining the evidence update can be written as follows:

where ω_Taste is the subjective weight given to tastiness, ω_Health. is the subjective weight given to healthiness, VD_Taste and VD_Health are the value differences in taste and healthiness attribute ratings, respectively, and τ_S is the time at which the taste and healthiness attributes come into the evidence accumulation process. If τ_S > 0, healthiness information is accumulated first, and evidence from taste comes into consideration at time τ_S, whereas ω_S < 0 means that taste information is accumulated first and evidence from healthiness information starts to come into the decision process at time |τ_S|. To test our hypotheses, medians of the posterior distributions for subject-level parameters were used as the dependent variables in our statistical analyses. We ran three separate, fully factorial (Group-by-Affect Condition-by Food Type), mixed-effects linear regressions, one for each parameter of interest. Each of these binary categorical variables were contrast coded using the values −1 and 1, allowing us to separately assess group differences across condition.

Results exploring additional parameters (boundary separation (α), non-decision time (τ_ND), and starting point bias (z)) are presented in the Supplementary Materials along with model comparisons, model identifiability, posterior predictive checks, and parameter recovery exercises.

First, we assessed how relative start time parameters varied across groups as a function of food type in each condition. We found a significant three-way interaction between group, condition, and food type (Table S2; β = −0.32, t = −2.22, p = 0.028), indicating that the two groups differed in how negative affect influenced their attribute bias patterns across food types. Visual inspection of the data suggested this interaction was driven by group differences across food types in the neutral condition that disappeared in the negative affect condition, and by a more pronounced influence of negative affect on attribute biases in BN (Figure 4A). Specifically, in the neutral condition, the HC group showed greater taste bias for low-fat versus high-fat foods (M = −0.430 vs. −0.286), while individuals with BN showed greater taste bias for high-fat versus low-fat foods (M = −0.293 vs. −0.450; Table S3). The negative affect induction eliminated this crossover pattern, increasing taste biases across both food types and groups, but with greater increases observed in the BN group (Tables S2, S3).

Figure 4.

Reduced tastiness weights are abnormally unaffected by negative affect in BN

Our second and third regression models assessed how each attribute weight varied across groups as a function of food type in each condition (Figure 4B-C). For taste weights, a significant two-way interaction between group and condition (Table S4; β = 0.18, t = 2.11, p = 0.036) indicated that although the HC group placed consistently higher weight on taste compared to the BN group, the HC group reduced their weighting of taste information following negative affect induction while the BN group maintained similar taste weights across conditions (Table S5). A main effect of food type (β = −0.23, t = − 3.69,p < 0.001) showed that both groups placed less weight on taste information for high-fat versus low- fat foods. In contrast, healthiness weights were insensitive to negative affect induction across both groups. However, a significant interaction between group and food type (Table S6; β = 0.29, t = 2.94, p = 0.004) indicated that the BN group placed consistently higher weights on health information, particularly for high-fat foods, while the HC group showed no difference in health weighting across food types (Table S7).

More severe symptoms of BN are linked to affect-induced delays in processing healthiness information

To assess the clinical significance of our findings, we sought to connect stDDM parameters to retrospective self-reported symptom severity. For these exploratory analyses, we used median values of the individual-level posterior distributions of the attribute onset parameter τ_s, the parameter which showed an exaggerated response to the affect induction in the BN group. These parameter values were then separately added to mixed-effects negative binomial regression models with Group and Food Type to assess their association with two symptom severity measures in the BN group: the frequency of objectively large binge eating episodes (OBEs) and subjectively large binge eating episodes (SBEs) in the past three months. We examined both types of out-of-control eating episodes because they both contribute to a diagnosis of BN according to ICD-11 (World Health Organization, 2022), and some data suggest that SBEs are more strongly related to negative affect (Brownstone et al., 2013; Brownstone & Bardone-Cone, 2021; Fitzsimmons-Craft et al., 2014). Indeed, we found that that greater negative affect-based decreases in τ_S for high-fat foods relative to low-fat foods were associated with more frequent SBE in the past three months (Table S8; β = −46.39, t = −2.80, p = 0.005). In other words, those who focused for longer on the tastiness of high-fat foods after the negative affect induction reported more SBE episodes in the past three months (Figure 5). Conversely, affect-based changes in τ_S were unrelated to OBE frequency (Table S8).

Figure 5.

Discussion

Negative affect is a known precursor to over-eating and dietary rule-breaking (Frayn & Knäuper, 2018). Healthy adults, and even rodents, can exhibit dysregulated food consumption and altered food choices after periods of stress or other negative emotions (Dionysopoulou et al., 2021; Jacques et al., 2019; Konttinen et al., 2010; Tomiyama, 2019). Negative affect’s influence on eating behavior is particularly pronounced among individuals with clinical diagnoses like BN, who tend to have large and out-of-control eating episodes in states of strong negative emotion (Cardi et al., 2015; Wonderlich et al., 2022). Prior studies have demonstrated that outside of binge-eating episodes, and in states of lower negative affect, individuals with BN typically engage in restrictive eating behaviors, consuming some low-fat foods and largely avoiding high-fat foods (Alpers & Tuschen-Caffier, 2004; Bjorlie et al., 2022; Elran-Barak et al., 2015; Kales, 1990). We previously found that women with BN demonstrated these restrictive food choices on a food-related decision task, regardless of their affective state (Gianini et al., 2019). However, by leveraging a computational model that captures the decision-making process underlying ultimate choices, we found that negative affect more strongly biased individuals with BN than healthy adults, increasing the amount of time that taste information alone influenced their decisions involving food. This affect-induced delay in the consideration of healthiness versus tastiness may help account for the fact that negative affect is a common precursor to taste-driven overeating in healthy adults, and the fact that it is a particularly strong predictor of taste-driven, out-of-control eating in individuals with BN. Consistent with this notion, we found that negative affect-induced delays in accumulating health information relative to taste information for higher fat foods were associated with more frequent out-of-control eating episodes in BN.

Previous studies have shown that in states of neutral affect, individuals with BN are more likely to consume foods that are low-fat and considered to be healthy (e.g., salads, fruits), as well as use healthiness information to guide their food choices (Alpers & Tuschen-Caffier, 2004; Bjorlie et al., 2022; Elran-Barak et al., 2015; Kales, 1990; Neveu et al., 2018; Schnepper et al., 2021). Our results demonstrate that in a neutral mood state, despite an initial tendency to focus on taste information before healthiness information, the BN group had stronger weights on healthiness information and lower, and in fact negatively signed, weights on taste, especially for high-fat foods. This negative weight on taste information (i.e., an aversive assignment to the value of taste) may have facilitated the BN group’s restrictive food choice policy compared to healthy controls. Individuals with BN may compensate for their tendency to focus on the tastiness of high-fat foods by negatively reappraising or deemphasizing the relative importance of taste and highly weighting their food choices based on health.

During binge-eating episodes, which often follow increases in negative affect, individuals with BN tend to consume large amounts of high-fat foods that were previously avoided (Alpers & Tuschen-Caffier, 2001; Berg et al., 2013; Collins et al., 2017; Haedt-Matt & Keel, 2011; Hilbert & Tuschen-Caffier, 2007; Smyth et al., 2007). Although prior studies examined food consumption associated with negative emotions in BN (Cardi et al., 2015), discrete food choice behaviors had not been investigated, and the influence of negative affect on specific aspects of the decision-making process were unclear. We showed that this affect-related change in eating behavior may occur because negative affect disrupts the delicate balance between how much weight tastiness and healthiness carry in influencing food choices and the relative speed with which they begin to exert their influence (Barakchian et al., 2021; HajiHosseini & Hutcherson, 2021; Hutcherson & Tusche, 2022; Lim et al., 2018; Maier et al., 2020; Schubert et al., 2021; Sullivan et al., 2015; Sullivan & Huettel, 2021). When in states of high negative affect, both healthy adults and individuals with BN may consider food’s appealing taste for longer, biasing them towards choosing tastier, high-fat foods. If this temporal advantage for tastiness is long enough, and a positive weight on tastiness is strong enough, individuals with BN could become more vulnerable to out-of-control eating because they could decide to eat tastier or high-fat foods before ever considering their perceived healthiness. This notion is consistent with our finding that greater affect-induced delays in processing healthiness information were associated with more frequent subjective binge episodes— episodes of dysregulated eating that may not include a large amount of food, but like objectively large binge-eating episodes, are dominated by carbohydrates and high-fat foods that are highly palatable (Presseller et al., 2023), and are often spurred by negative affect (Alpers & Tuschen-Caffier, 2001; Berg et al., 2013; Haedt-Matt & Keel, 2011; Hilbert & Tuschen-Caffier, 2007; Smyth et al., 2007).

Our findings reveal new insights into the dynamics of food-related decision-making in both healthy adults and BN and may help to inform new interventions. When attribute weights are inconsistent with a larger goal, decision-makers can use shifts in the timing of attribute consideration to achieve a desired outcome (Amasino et al., 2019; Maier et al., 2020). For example, earlier shifts in the relative onset time of healthiness information could help an individual with stronger weights on tastiness adhere to a diet. Perhaps targeting affect-induced shifts in the timing of attribute consideration could help regulate the over- and under-controlled eating seen in binge-fast cycles. For example, past research in individuals without eating disorders indicates that cueing attention towards the healthiness of foods during a choice task both increased the weight on healthiness information and increased its onset time during decision-making (Barakchian et al., 2021; Maier et al., 2020) but see (Sullivan & Huettel, 2021). Several interventions already focus on modifying attentional biases in individuals with obesity and binge eating (Boutelle et al., 2016; Brockmeyer et al., 2019; Stojek et al., 2018). Our results suggest that such interventions could be targeted even more precisely. One way for individuals to avoid emotion-driven binge eating may be to focus attention away from the taste attributes of tastier, high-fat foods during states of high negative affect. Conversely, during states of low negative affect, increased attention on the positive taste attributes of all foods may be instrumental in motivating individuals with BN to engage in less restrictive eating.

Although we found an aberrant influence of negative affect on the computations that contribute to food choice in BN, in our paradigm, the influence of negative affect was not strong enough to change final choices in either group. Even after the negative affect induction which decreased taste weights for HC and not BN, choice outcomes in BN were ultimately determined by the advantage in weighted evidence for healthiness as opposed to the advantage in relative timing for tastiness. We speculate that this may have been for two reasons. First, the affect induction method may not have been potent enough. In the current study, post-induction increases in negative affect (d_HC = 1.34; d_BN = 1.19) were smaller than those reported after other induction methods, such as viewing affectively salient images with congruent music (d = 3.33; (Zhang et al., 2014)), and they may not have been powerful enough to effect changes in ultimate choices. Although the Food Choice Task was completed immediately after the affect induction, we do not know how long its effect lasted because mood was not repeatedly assessed throughout the task. This and previous studies using task-based designs have only assessed mood ratings immediately following affect inductions (Rouhani et al., 2025; Werthmann et al., 2014), but future studies could include continuous assessments throughout the experiment to better quantify potential group differences in how induced affect changes persist and/or fluctuate. In addition, our negative-affect inducing music and writing exercise may have created an emotional experience that inadequately represents the negative affect that typically precedes emotional eating or binge eating. Qualitative studies of patients with BN indicate that interpersonal stressors and related negative emotions (e.g., resentment) often precipitate binge episodes (Bohon et al., 2021; Wasson, 2003). Perhaps larger increases in negative affect intensity or exposure to the specific types of negative affect which typically precede binge eating could produce a large enough asynchrony in onset timing of attributes and changes in attribute weights to change individuals’ choices (Figure 6). Second, participants knew they would be offered a snack-sized amount of food from a randomly selected trial to consume at the end of the task. This quantity of food may not have provided participants with access to the type of binge-eating experience that negative affect would otherwise precipitate. Further work is needed to test whether the same processes identified in the current study also drive binge eating outside the lab (Wonderlich et al., 2022). Future adjustments to incentive compatibility, with designs that allow for increased consumption beyond snack-sized portions and/or guaranteed access to a wider-array of high-fat foods could clarify whether opportunities to overconsume highly palatable food would generate different patterns of food choice behavior on the task. Additionally, future research should include continuous assessments of other relevant factors such as cravings (Konova et al., 2018; Wonderlich et al., 2017) to enrich the connection between mood, discrete food choice, and binge eating.

Figure 6.

Of note, the modest sample size (n = 46) composed exclusively of female participants may limit the generalizability of our findings. Replication in larger samples that include male participants is needed. Nonetheless, the study’s repeated-measures design benefits from the added statistical power and sensitivity of within-subjects statistical tests, which reduce error variance by controlling for individual differences (Charness et al., 2012; Judd et al., 2017; Westfall et al., 2014). Moreover, we collected a large quantity of data per subject, which enhances the validity of the findings by providing a more precise estimate of the effects of the experimental manipulation, despite the smaller sample size (Baker et al., 2021; Smith & Little, 2018).

Overall, our results advance our understanding of how changes in affective states may shift individuals with BN away from typically restrictive dietary choices and towards binge eating. Specifically, negative affect can bias individuals with BN to focus on taste information for longer, and the strength of this effect for high-fat foods was associated with symptom severity. Our results add to recent data suggesting that computational modeling can detect subtle alterations in decision-making processes in BN that are sensitive to state changes (Berner et al., 2023), and future research should continue examining the potential connections between negative affect and eating-related decision-making using these analytic approaches.

Materials and Methods

Participants

Data were collected from 25 individuals who met DSM-5 diagnostic criteria for BN and 21 healthy controls, group-matched for age and BMI. All participants were female and aged 18-40 years. The sample consisted of 27 participants who identified as Caucasian (58.7%), 7 as Hispanic (15.2%), 6 as Asian (13.0%), 3 as Black (6.5%), and 3 as mixed race (6.5%). For details on recruitment and exclusions, see (Gianini et al., 2019).

Materials and procedures

Participants completed the tasks over two separate study sessions, each lasting approximately four hours. On each study day, participants were instructed to abstain from consuming food or drinks except for a standardized meal consumed two hours preceding the study session.

Participants completed a Food Choice Task (Steinglass et al., 2015) on two separate days, in two counterbalanced states: after a neutral affect induction, and after a negative affect induction. This task is composed of three phases: Healthiness Rating, Tastiness Rating, and Choice. In each phase, participants viewed the same 43 food items, each categorized as either low-fat (< 30% kcals from fat) or high-fat (> 30% kcals from fat). During the Health Rating phase, participants used a 5-point scale to rate the healthiness of each food item from 1 (Unhealthy) to 3 (Neutral) to 5 (Healthy). During the Taste Rating phase, participants used a 5-point scale to rate the tastiness of each food item from 1 (Bad) to 3 (Neutral) to 5 (Good). The order of the Health Rating and Test Rating phases was counterbalanced. After both phases were completed, one food item that was rated Neutral for both health and taste was selected as a Reference Item for use in the Choice phase. If no item was rated Neutral for healthiness and tastiness, then a food item rated 3 on health and 4 or above on taste was selected following previously established procedures (Steinglass et al., 2015). During the Choice phase, participants made choices between the neutral Reference Item and other food items presented on the computer screen. Participants indicated their choice on a 5-point scale from 1 (No – select the Reference Item) to 3 (Indifference) to 5 (Yes – select the shown food item). Participants were informed that one randomly selected trial would be selected for payout at the end of the experiment and would receive the food item they selected on that trial.

The experiment included an affect induction that included a combination of music and autobiographical writing (Werthmann et al., 2014). During the negative affect induction, participants were asked to write about a recent negative experience while listening to “Adagio for Strings” by Samuel Barber for 8 minutes. During the neutral affect induction, participants were asked to write about the route they took to get to the study site while listening to “Dancing with the Sun” by Celia Felix for 8 minutes. The 65-item Profile of Mood States scale was administered before and after the affect inductions to assess changes in affect (McNair et al., 1989).

Participants also completed the following measures: Eating Disorder Examination Questionnaire (EDE-Q; (Fairburn & Beglin, 1994)); Difficulties in Emotion Regulation Scale (DERS; (Gratz & Roemer, 2004)); Urgency, Premeditation, Perseveration, Sensation Seeking, and Positive Urgency Behavior Scale-Negative Urgency subscale (UPPS-P Negative Urgency; (Lynam et al., 2006)); Beck Depression Inventory (BDI; (Beck et al., 1961)); State Anxiety Inventory (STAI; (Spielberger et al., 1983)).

For additional information, please see (Gianini et al., 2019).

Data analysis

Data transformations and exclusions

Following the procedure from (Gianini et al., 2019), participants’ responses from the Choice phase of the FCT were binarized from their 5-point scale. Responses marked 1 or 2 (i.e., “no”) were determined to be choices of the reference item, responses marked 4 or 5 (i.e., “yes”) were determined to be choices of the presented food item, and responses marked 3 (i.e., “indifferent”) were omitted (total n =599; BN n = 292, HC n = 307).

In contrast to the original study, we excluded trials containing outlier responses times (RTs). Outlier response times likely reflect phenomena outside the decision-making processes of interest, including attention lapses or accidental responses (Cousineau & Chartier, 2010; Ratcliff, 1993). We excluded outlier responses using cutoffs of ±3 standard deviations (SD) from the mean (Berger & Kiefer, 2021). For each participant, the mean and standard deviation of their RTs were calculated and trials containing responses larger/smaller than the mean ±3 SDs were excluded. Additionally, we removed trials where the RTs were greater than 10,000 ms or less than 250 ms after the mean ±3 SDs treatment. Using this method, we removed 2.5% of trials (total n =85; BN n = 51, HC n = 34).

Computational modeling

We used a time-varying diffusion decision model (DDM; (Ratcliff & McKoon, 2008) to study the dynamics of the decision-making during the Food Choice Task. This time-varying model introduces a starting time parameter that indicates the delay with which different attributes enter the evidence accumulation process. This model was previously fitted to food choice data (Lombardi & Hare, 2021; Maier et al., 2020) and lottery choice data (Chen et al., 2022) from healthy adults.

The starting time DDM (stDDM) was estimated using a hierarchical, Bayesian framework implemented with the R package Rjags, which uses the JAGS MCMC sampling algorithm (Plummer, 2003). Our fitting scripts were adapted from those developed by Hsiang-Yu Chen, Gaia Lombardo, and Todd Hare (Chen et al., 2022). This fitting method simultaneously estimates both group- and individual-level parameters, which improves the reliability of parameters estimated from data with low trial counts per participant (Ratcliff & Childers, 2015; Wiecki et al., 2013). Parameter estimates were fitted to data that included individual-level response times, choices, z-scored health and taste ratings, and Affect Condition. We specified a model where individual parameters were drawn from four separate group-level distributions that varied both by Group (BN or HC) and by Affect Induction (neutral or negative). Within this model, attribute weight parameters (ω_Taste and ω_Health) and the relative-starting time parameter (τ_S) were estimated separately based on food-types (low-fat or high-fat). Using model comparisons and posterior predictive checks, we confirmed that this model best fit our data compared to simpler models (see Supplementary Materials).

Following previously described protocols (Chen et al., 2022), we used group-level priors for attribute weight parameters (ω_Taste and ω_Health), the relative-starting time parameter τ_s), and the affect induction parameters that were drawn from Gaussian distributions with mean = 0 and standard deviation = 1. These parameter values were then divided by their standard deviations before fitting the model. The priors for boundary separation (α) and non-decision time (τ_ND) were drawn from uniform distributions with ranges 1.0 × 10⁻⁴ to 5 and 0 to 10, respectively. The priors for starting point bias (z) were drawn from a beta distribution where both the shape and scale parameters were set to 2. All individual-level priors were drawn from gamma distributions with shape parameters of 1 and scale parameters of 0.1. Posterior estimates were drawn from three chains, each with 100,000 samples (85,000 discarded as burn-in) and thinning every 10 samples. Convergence among chains was assessed using the Gelman-Rubin statistic (Gelman & Rubin, 1992) using a threshold of 1.1. Model fit was quantified with the Widely Applicable Information Criterion (WAIC) (Watanabe, 2010), which penalizes for model complexity based on the number of parameters to prevent overfitting. For parametric tests involving parameter estimates, we used the medians of the individual-level posterior distributions. See the Supplementary Materials for model information on model comparisons, model identifiability, posterior predictive checks, and parameter recovery.

Multilevel linear regressions were run to evaluate our hypotheses regarding the influence of Food Type, affect induction, BN diagnosis, and their interactions on parameter estimates. These models were implemented using the lme4 (Bates et al., 2015) and lmertest (Kuznetsova et al., 2013) R packages. Each regression model included Food Type (coded −1/1 for low-fat/high-fat foods), Affect Condition (coded −1/1 for Neutral/Negative), and Group (coded −1/1 for HC/BN), and their interactions as independent variables. For these models, the intercepts were treated as random effects and statistical significance was evaluated using α = 0.050.

To unpack observed interaction effects, we ran exploratory simple effects analyses to separately evaluate the influence of independent variables at specific values of other independent variables (Aiken et al., 1991; Brambor et al., 2006; Jaccard & Turrisi, 2003; Winer et al., 1991).

We used the R package emmeans to compute estimated marginal means for specified factor combinations (Lenth, 2025), evaluating statistical significance using α = 0.050.

Additional fixed-effects regressions were run to assess the relationship between parameter estimates and symptom severity. For these exploratory analyses, we used negative binomial models to predict the three- month frequency of self-report Objective Binge Episodes (OBE) and Subjective Binge Episodes (SBE). We implemented these models using the R package glmmTMB (Brooks et al., 2017), with a zero-inflated term included for the model predicting SBE. For each of these two models, individual-level parameter estimates from each condition and food-type were included as independent variables. We applied a Bonferroni correction to control for the FWE of these exploratory analyses, using α = 0.025.

Supplemntary Methods

Behavioral analyses

We used a similar analytic approach to assessing choice and response time data to the methods described in (Gianini et al., 2019). A multilevel logistic regression models was fit to trial-level choice data (coded 1/0 for yes/no) using Food Type (coded -1/1 for low-fat/high-fat foods), Affect Condition (coded -1/1 for Neutral/Negative), Group (coded -1/1 for HC/BN), and their interactions as independent variables. We fit a second logistic regression model to trial-level choice data, also using Affect Condition, Group, and their interaction as independent variables. In this second model, we also included z-scored health and taste ratings, and their independent interactions with group, as independent variables.

We assessed self-control by selecting trials where participants made decisions involving highly appetitive unhealthy foods (i.e., taste rating = 5, health rating = 1) or unappetizing healthy foods (i.e., taste rating = 1, health rating = 5). The use of self-control (coded 1 for rejecting appetitive unhealthy foods or selecting unappetizing healthy foods and 0 otherwise) was assessed using a multilevel logistic regression with Affect Condition, Group, and their interaction entered as independent variables.

A similar multilevel linear regression was fit to trial-level response time data (log-transformed) using the same independent variables. This model tested the main effects and interactions of Food Type, Affect Condition, and Group with a dummy choice variable (coded -1/1 for choice of reference item or food item).

For all models, any within-individual variables (i.e., the intercept, main effects of z-scored health and taste ratings, Affect Condition, and Food Type, and their interactions) were treated as random effects. Regression analyses were conducted using the R packages lme4 (Bates et al., 2015) and nmle (Pinheiro et al., 2022) and marginal means were estimated using emmeans (Lenth et al., 2024) and sjPlot (Lüdecke et al., 2024).

Model comparison

We assessed the model fit of variants of the time-varying DDM to determine which model parameters best explained the potential influence of food type on behavior. The null model (M0) assumed decision parameters did not depend on food type. The attribute weight model (M1) assumed both ω_taste and ω_health varied based on food type. The attribute starting time model (M2) assumed τ_s varied based on food type. Finally, the full model (M3) assumed ω_taste, ω_health, τ_s varied based on food type. All models were estimated using the same procedures described in the Computational model subsection of the Methods section located in the main text.

Model identifiability and posterior predictive checks

We then evaluated the identifiability of each model. For each model, a set of simulated datasets was generated reflecting the structure and parameters of the corresponding model. First, each model was fit to the empirical data using the procedures described in the Materials and Methods section of the main text, under the subsection Computational Modeling. Data were simulated using the empirical trial structure and the medians of the individual-level posterior parameter distributions. Each simulated dataset was fit using all four models (M0, M1, M2, and M3), regardless of the generating model. For model fitting, we employed the same Bayesian hierarchical framework that was used in previous steps. We evaluated model identifiability by determining the frequency with which the model that generated the data was also identified as the best-fitting model, based on WAIC (Watanabe, 2010) implemented in the loo package in R (Vehtari et al., 2017, 2024). This allowed us to assess the extent to which each model was uniquely identifiable and distinguishable from alternative models. A model was considered successfully recovered when it was the best-fitting model for the data it generated.

In addition, we performed posterior predictive checks to ensure that the models could capture key features of the observed data. We assessed the simulated choice and response time patterns of the model variants using nonparametric equivalent aligned rank tests (Wobbrock et al., 2011) implemented in the ARTool package in R (Kay et al., 2021). For this procedure, we calculated both the empirical and the simulated mean choice proportions and median response times for each participant across Affect Conditions and Food Types. Aligned rank tests were run separately on mean choice proportions and median response times evaluating the interaction between Group, Affect Condition, Food Type, and Source (empirical or simulated data) for each model. We subsequently ran the multi-level logistic regression models predicting trial-level choice from the interaction between Group, Affect Condition, and Food Type on the simulated data to confirm we could recover the results from the empirical data.

Parameter recovery

We conducted a parameter recoverability exercise to ensure that the values of the winning model could be accurately recovered from data that had similar qualities to the empirical data included in the current study. For this exercise, we took the mean values of the subject-level parameters (χ_fit) estimated in both the neutral and negative affect conditions and simulated data using the trial-level health and taste ratings from the empirical data used during estimation. This synthetic data was then used to fit the winning model using the procedures described previously. We extracted this new set of parameters, once again calculating the mean values of the subject-level parameters (χ_rec). We ran Pearson-correlations between χ_fit and χ_rec to assess parameter recovery.

Mood analysis

We analyzed the specific emotions that changed following the mood inductions. As described in the main text, mood was assessed using the Profile of Mood States (POMS; (McNair et al., 1989)). The POMS has the following subscales: Anger, Confusion, Depression, Fatigue, Tension, and Vigor.

Alternative symptom severity measure

As recent studies suggest that negative urgency (i.e., acting impulsively when experiencing negative affect) contributes to binge-eating pathology (Wonderlich et al., 2024), we examined whether decision parameters might correlate with symptom severity measures related to eating-disorder-linked behaviors outside the lab. We used linear regressions to evaluate the extent to which subject-level parameters (i.e., τ_s, ω_taste, ω_health) could predict scores on the negative urgency subscale of the UPPS-P Impulsive Behavior Scale (Lynam et al., 2006) and the restraint subscale of the Eating Disorder Examination Questionnaire (EDE-Q; (Fairburn & Beglin, 1994)).

Supplementary Results

Analysis of choice patterns

We replicated the findings from the original analysis, confirming that individuals with BN were more likely than the HC group to choose the reference item (Table S9; β = −0.48, z = −2.63, p = 0.009), with the BN group selecting the reference item in approximately 68.3% of trials (95% CI: 57.0% −77.8%) and the HC group selecting it in 45.3% of trials (95% CI: 33.9% −58.3%). We also replicated the Group-by-Food Type interaction reported in the original analysis (β = −0.27,z = −1.99,p = 0.047). Those in HC group chose approximately 63.1% (95% CI: 52.0% −73.0%) of low-fat foods and 46.0% (95% CI: 27.8% −65.3%) of high-fat foods, while individuals with BN chose 53.0% (95% CI: 42.6% −63.1%) of low-fat foods and 16.1% (95% CI: 8.3% −28.8%) of high-fat foods. While individuals with BN were less likely to choose the presented foods over the neutral reference item in general, both groups chose high-fat foods less often than low-fat foods (β = −0.62, z = −4.58, p =4.75 × 10⁻⁶).

Similar to the original analysis, we did not observe a significant main effect of Affect Condition (β =−0.02, z = −0.26, p = 0.793), an interaction between Affect Condition and Group (β = 0.07, z = 0.82, p = 0.410), or an interaction Affect Condition and Food Type (β = −0.03, z = −0.62, p = .535) on food choice. We also did not observe a significant three-way interaction between Affect Condition, Food Type, and Group on choice (β = −0.10, z = −1.91, p = 0.056).

In a separate multilevel regression (Table S10), we examined the relationships among food choice, taste and health ratings, BN diagnosis, and negative affect. We confirmed the original paper's finding that health ratings more strongly influenced food choice among the BN group compared to the HC group (β = 0.64, z = 2.77, p = 0.006). We additionally observed that taste ratings were less influential among the BN group compared to the HC group (β = −0.65, z = −2.78,p = 0.005).We also reproduced the original paper's finding of a significant, negative interaction between Affect Condition and taste ratings on choice, such that negative affect reduced the influence of taste ratings on choice among both groups (β = −0.45,z = −2.57,p = 0.010).

Furthermore, we examined trials with an opportunity for self-control (Table S11). Reproducing the original analyses, we found that the BN group used self-control more often than the HC group (β =0.38, z = 2.14, p = 0.033). We did not observe a main effect of Affect Condition on self-control (β =0.08,z = 0.70,p = 0.486), nor an interaction between Affect Condition and Group (β = 0.15,z = 1.27, p = 0.204).

We next examined RT patterns using multilevel regression models fit to trial-level RT data (Table S12). We did not observe significant main effects of Affect Condition or its interactions with other variables. An interaction between Food Type and choice on RT indicated that, across groups, rejections were relatively faster for high-fat foods than for low-fat foods, while selections were relatively slower for high-fat foods than for low-fat foods (β = 0.03, t = 2.70, p = 0.007). Additionally, we found a positive interaction between Group and choice, suggesting that the BN group was slower than the HC group to choose presented foods over the neutral reference item (β = 0.03, t = 2.44, p = 0.015).

Modeling results

Model comparison

These different model variants are described in Table S13. The winning model (M3) included terms representing the influence of Food Type on both attribute weights (ω_taste, ω_health) and attribute starting time (τ_s).

Model identifiability

The results of the model recovery exercise are summarized in Table S14. Models M0, M1, and M3 all had perfect recoverability. However, M2 had poor recoverability, as it was always mis-identified as M3.

Posterior predictive check

The data from M3 best replicated the results from the empirical data. While all models had biases in predicting choice data, these were smallest in M3. M0 (Table S15) was significantly biased in predicting choice based on Food Type (F(1,352) = 11.97, p = 0.001) and was biased in predicting response times (F(1,352) = 7.88, p = 0.005). Both M1 (Table S16) and M2 (Table S17) were similarly biased in predicting choice based on Food Type (M1: F(1,352) = 45.07, p = 7.63 × 10⁻¹¹; M2: F(1,352) = 4.90,p = 0.028), but M3 (Table S18) was not (F(1,352) = 1.26,p = 0.262).

In the empirical data, a multilevel logistic regression found significant, negative main effects of Group and Food Type, and a negative Group × Food Type interaction (Table S9). M0 (Table S19) was able to reproduce the main effect of Group (β = −0.34, t = −2.74, p = 0.006) and the Group × Food Type interaction (β = −0.32, t = −4.34, p = 1.41 × 10⁻⁵), but not the main effect of Food Type (β =−0.01, t = −0.19, p = 0.848). M1 (Table S20) produced a marginal negative effect of Group (β =−0.28, t = −1.93, p = 0.053) and the negative Group × Food Type interaction (β = −0.22, t = −3.27, p = 0.001), but the significant effect of Food Type had the wrong sign (β = 0.42, t = 6.35, p = 2.17 × 10⁻¹⁰). M2 (Table S21) produced the negative Group effect (β = −0.42, t = −3.08,p = 0.002) and the negative Group × Food Type interaction (β = −0.36, t = −3.36, p = 7.93 × 10⁻⁴), but not the significant effect of Food Type (β = −0.16, t = −1.52, p = 0.128). Only M3 (Table S22) could reproduce all three patterns: a negative effect of Group (β = −0.36, t = −2.68, p = 0.007), a negative effect of Food Type (β = −0.36, t = −4.16, p = 3.25 × 10⁻⁵), and a negative interaction between the two (β = −0.27, t = −3.11,p = 0.002).

Parameter recovery

As illustrated in Figure S1, recovery was satisfactory for all parameters of M3.

Non-decision time

We did not observe significant effects of either Group or Affect Condition in non-decision time (τ_ND; Table S23). However, we found a significant Group by Affect Condition interaction, suggesting that the BN group, compared to the HC group, had faster non-decision times following the negative affect induction versus the neutral affect induction (β = −0.09, t = −2.29, p = 0.026).

Boundary separation

For the boundary separation parameter (α; Table S24), there was a significant difference between the HC and BN groups (β = −0.39, t = −3.09, p = 0.003). Specifically, the BN group exhibited reduced response caution compared to the HC group. We found a significant, negative effect of Affect Condition (β = −0.36, t = −4.24, p = 1.06 × 10⁻⁴), but a trending, positive interaction between Group and Affect Condition (β = 0.22, t = 1.94, p = 0.059). A simple effects analysis (Table S25) confirmed that the negative affect induction significantly reduced boundary separation among the HC group (M = 0.35, t = 4.15, p = 1.00 × 10⁻⁴), but not the BN group (M = 0.14, t = 1.73, p = 0.091).

Starting point bias

There were no Group or Affect Condition differences in starting point bias (z; Table S26).

Mood analysis

Anger subscale

We found that the BN group had significantly higher scores on the Anger subscale across time (Table S27; β = 7.59, t = 3.16, p = 0.002). We also found a significant, positive interaction between Affection Condition and Timing (β = 5.71, t = 3.04, p = 0.003), indicating that the negative affect induction increased anger ratings across Groups.

Confusion subscale

We found that the BN group had significantly higher scores on the Confusion subscale across time (Table S28; β = 6.07, t = 4.75, p = 1.03 × 10⁻⁵). We also found a significant, positive interaction between Affection Condition and Timing (β = 2.30, t = 2.42, p = 0.017), indicating that the negative affect induction increased confusion ratings across Groups.

Depression subscale

We found that the BN group had significantly higher scores on the Depression subscale across time (Table S29; β = 16.06, t = 4.63, p = 1.84 × 10⁻⁵). We also found a significant, positive interaction between Affection Condition and Timing (β = 5.58, t = 2.44, p = 0.016), indicating that the negative affect induction increased depression ratings across Groups.

Fatigue subscale

We found that the BN group had significantly higher scores on the Fatigue subscale across time (Table S30; β = 6.04, t = 334, p = 0.001). No other significant contrasts emerged.

Tension subscale

We found both a significant, negative interaction between Group and Timing (β =−2.50, t = 1.67, p = 0.021) and a significant, positive interaction between Group, Affect Condition, and Timing (β = 3.82, t = 2.53, p = 0.013). Taken together, these indicate that only the BN group's tension ratings changed after the affect induction, but in a condition-dependent way: decreasing tension following the neutral affect induction and increasing tension following the negative affect induction.

Vigor subscale

We found that the BN group had significantly lower scores on the Vigor subscale across time (Table S32; β = −10.40, t = −5.73,p = 2.33 × 10⁻⁷). We also found a significant, negative interaction between Affection Condition and Timing (β = −4.93, t = −3.58, p = 0.001), indicating that the negative affect induction decreased vigor ratings across Groups.

Alternative symptom severity measures

There were no significant relationships between decision parameters and negative urgency scores (Table S33) or restraint scores (Table 34).

Supplementary Tables

Supplementary Table S1.

Supplementary Table S2.

Supplementary Table S3.

Supplementary Table S4.

Supplementary Table S5.

Supplementary Table S6.

Supplementary Table S7.

Supplementary Table S8.

Supplementary Table S9.

Supplementary Table S10.

Supplementary Table S11.

Supplementary Table S12.

Supplementary Table S13.

Supplementary Table S14.

Supplementary Table S15.

Supplementary Table S16.

Supplementary Table S17.

Supplementary Table S18.

Supplementary Table S19.

Supplementary Table S20.

Supplementary Table S21.

Supplementary Table S22.

Supplementary Table S23.

Supplementary Table S24.

Supplementary Table S25.

Supplementary Table S26.

Supplementary Table S27.

Supplementary Table S28.

Supplementary Table S29.

Supplementary Table S30.

Supplementary Table S31.

Supplementary Table S32.

Supplementary Table S33.

Supplementary Table S34.

Supplementary Figures

Supplementary Figure S1.

Supplementary Figure S2.

Acknowledgements

We thank the women who participated in this study for their time. Collection of these data was supported by the National Institute of Mental Health grant T32-MH096679-01A1 (PI: LG). Data analysis and preparation of this manuscript were supported by career development awards from the NIH (K12AR084233; PI: KH; K23MH118418; PI: LAB) and a NARSAD Young Investigator Grant from the Brain & Behavior Research Foundation (PI: LAB). This work was supported in part through the computational and data resources and staff expertise provided by Scientific Computing and Data at the Icahn School of Medicine at Mount Sinai and supported by the Clinical and Translational Science Awards (CTSA) grant UL1TR004419 from the National Center for Advancing Translational Sciences. Research reported in this publication was also supported by the Office of Research Infrastructure of the National Institutes of Health under award number S10OD026880 and S10OD030463. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. We would like to thank Todd Hare for helpful discussions regarding model development.

Additional information

Code Availability

Custom code in this work is available at https://github.com/blairshevlin/Computations_BN_Food_Choice.

Author Contributions

BRKS: Methodology, Formal Analysis, Validation, Visualization, Writing - Original Draft Preparation, Review, and Editing

LG: Original Study: Conceptualization, Funding Acquisition, Methodology, Investigation, Project Administration, and Data Curation; Current Paper: Writing - Review and Editing

JES: Original Study: Resources and Supervision; Current Paper: Writing - Review and Editing

KF: Original Study: Methodology and Data Curation; Current Paper: Writing - Review and Editing

ECL: Writing - Review and Editing

KH: Data Curation, Writing - Review and Editing

LAB: Conceptualization, Methodology, Formal Analysis, Supervision, Writing - Original Draft Preparation, Review, and Editing