Intertemporal Pavlovian biases and links with mental health in the general population

Floor Burghoorn; Anouk Scheres; Karin Roelofs; Bernd Figner

doi:10.3758/s13415-025-01326-9

. 2025 Aug 11;25(5):1306–1329. doi: 10.3758/s13415-025-01326-9

Intertemporal Pavlovian biases and links with mental health in the general population

Floor Burghoorn ^1,^2,^✉, Anouk Scheres ¹, Karin Roelofs ^1,³, Bernd Figner ^1,³

PMCID: PMC12464115 PMID: 40790000

Abstract

Previous research has found the anticipation of immediate (compared with delayed) rewards to enhance instrumental approach behaviour and interfere with instrumental inhibition. This so-called intertemporal Pavlovian bias has been proposed as a mechanism contributing to impatient actions. In the present preregistered study, we aimed to replicate this intertemporal Pavlovian bias effect, and examined whether individual differences in the strength of this bias were associated with mental health symptoms characterized by increased or decreased intertemporal impatience. A sample of 389 participants completed an intertemporal go/no-go learning task, as well as questionnaires that assessed symptoms of alcohol use disorder, nicotine dependence, depression, anxiety, attention-deficit/hyperactivity disorder (ADHD), disordered eating, impulsivity, and body mass index. We replicated the intertemporal Pavlovian bias effect on the probability of go responding (a response bias), and reinforcement learning models again pointed to a Pavlovian cue-response bias as computational mechanism driving this effect. Contrary to our expectations, the response bias was not associated with mental health symptoms. We also observed an intertemporal Pavlovian bias effect on the vigour of correct go responses, with faster responses in anticipation of immediate rewards (a response time or RT bias). A stronger RT bias was associated with ADHD hyperactivity/impulsivity and nonplanning impulsivity symptoms, while a reversed RT bias (with faster incorrect go responses in anticipation of delayed rewards) was associated with eating disorder symptoms and total and motor impulsivity. Together, our results support the robustness of the intertemporal Pavlovian response bias and point towards the possible relevance of the RT bias for several mental health problems.

Supplementary Information

The online version contains supplementary material available at 10.3758/s13415-025-01326-9.

Keywords: Intertemporal choice, Delay discounting, Reinforcement learning, Pavlovian bias, Mental health, Computational psychiatry

Intertemporal decisions refer to choices between smaller short-term and larger long-term outcomes. For instance, we may choose between watching another episode of our favourite series versus going back to writing a difficult paper to obtain our degree, or we may be tempted to smoke a cigarette instead of abstaining to improve our long-term health. Individuals vary in the degree of intertemporal impatience they exhibit in these situations, i.e., their tendency to choose short-term over long-term outcomes. In research, the most common approach to quantify intertemporal impatience is to ask participants to make a series of choices between sooner-smaller and later-larger monetary rewards (also known as delay discounting tasks, referring to the subjectively discounted value of a reward as its delay until delivery increases; Lempert et al., 2018). Intertemporal impatience has been associated with a large variety of maladaptive behaviours and mental health problems, with meta-analyses showing increased impatience in individuals diagnosed with attention-deficit/hyperactivity disorder (ADHD; Jackson & Mackillop, 2016; Marx et al., 2021; Patros et al., 2016; Pauli-Pott & Becker, 2015), substance and gambling addiction (MacKillop et al., 2011), obesity (Amlung et al., 2016), major depressive disorder, schizophrenia, borderline personality disorder, bipolar disorder, bulimia nervosa, binge-eating disorder (Amlung et al., 2019), and decreased impatience in individuals diagnosed with anorexia nervosa (Amlung et al., 2019), compared with individuals without mental health diagnoses. Following these findings, intertemporal impatience has been proposed as a transdiagnostic process that may form a risk factor contributing to the development and maintenance of these disorders (supported by studies showing an etiological role of impatience), and may explain some of the comorbidity between disorders (Amlung et al., 2017, 2019; Bickel et al., 2012, 2019; Lempert et al., 2018; Scheres et al., 2024). In addition to these differences between patients and controls, studies have observed continuous associations between intertemporal impatience and the severity of mental health problems in general population samples, in line with the proposed dimensionality of intertemporal impatience and mental health constructs (Amlung et al., 2016, 2017; Burghoorn et al., 2025; Keidel et al., 2024; Levin et al., 2018; Levitt et al., 2022).

Intertemporal Pavlovian bias

The relevance of intertemporal impatience across mental health problems raises the question about which psychological mechanisms may contribute to impatience and thus may be implicated in the behaviours and mental health symptoms that have been characterized by intertemporal impatience. In a recent study, we provided evidence for an intertemporal Pavlovian bias as a possible mechanism contributing to impatient intertemporal decisions (Burghoorn et al., 2024). Pavlovian biases refer to the influence of anticipated outcomes (e.g., rewards or punishments), signalled by cues in the environment, on instrumental (often goal-directed) actions. Previous research has demonstrated that the anticipation of a reward elicits a Pavlovian approach response, which is adaptive when approach responses are required to achieve a goal, but maladaptive when inhibitory or withdrawal responses are required to achieve a goal (Guitart-Masip et al., 2014). In contrast, the anticipation of a punishment has been shown to elicit a Pavlovian withdrawal response, enhancing instrumental approach but interfering with instrumental withdrawal. Pavlovian responses may either align or conflict with instrumentally required actions. A key difference between Pavlovian and instrumental action control is that whereas Pavlovian control revolves around relatively simple cue-outcome associations, instrumental control revolves around more complex cue-action-outcome associations. Thus, only instrumental actions (particularly when they are goal-directed) are determined by their effect on the desired outcomes or goals. Although Pavlovian actions are often shaped through their general adaptiveness (e.g., it is generally adaptive to approach rewards or reward-predicting cues; Dayan et al., 2006), they can result in actions that persist even when maladaptive to reaching one’s goal (e.g., when a reward requires inhibition). In other words, while instrumental control is computationally costly and flexible, Pavlovian control is computationally cheap and inflexible (Dorfman & Gershman, 2019; Huys et al., 2012; Millner et al., 2018).

Until recently, research on Pavlovian biases exclusively focused on the valence and magnitude of anticipated outcomes (see Guitart-Masip et al., 2014 for a review), comparing the anticipation of rewards versus punishments (and sometimes also comparing different reward and/or punishment magnitudes). In Burghoorn et al. (2024), however, we demonstrated that the timing of the anticipated reward also influenced instrumental actions in a Pavlovian manner. To test this, we developed an intertemporal variant of a go/no-go learning task that orthogonalizes the required instrumental action (go/no-go) and the timing of the reward that is available upon giving the correct response (immediate/delayed). This resulted in four trial types: go to win immediate reward, go to win delayed reward, no-go to win immediate reward, and no-go to win delayed reward trials. The immediate and delayed rewards were preference-matched per participant by using a choice titration procedure, allowing us to test for the effect of reward timing beyond individual differences in subjectively discounted reward value. We observed that the anticipation of immediate rewards increased go responding compared with the anticipation of delayed rewards, thus facilitating instrumental approach but interfering with instrumental inhibition. This intertemporal Pavlovian bias may contribute to the often-experienced difficulty to inhibit ourselves in the face of immediate gratification, possibly at the cost of long-term goals. For instance, the sight of your favourite unhealthy snack may elicit an approach response (i.e., eating the snack), giving you an immediate reward (i.e., the taste of the snack) but undermining your long-term health goal. Note that irrespective of the reward timing, any reward may elicit a Pavlovian approach response and may therefore contribute to intertemporally impatient actions (as also discussed by Dayan et al., 2006; Huys et al., 2012). Importantly, however, the results of Burghoorn et al. (2024) extend this by demonstrating that immediate rewards exert a stronger approach response compared with delayed rewards, even when the two rewards are matched based on revealed choice preferences.

Intertemporal Pavlovian bias and mental health: The present study

In the current study, we asked how such an intertemporal Pavlovian bias may be implicated in those mental health problems that have been characterized by increased or decreased intertemporal impatience. For instance, a stronger difficulty to inhibit one’s response to cues signalling immediate gratification (e.g., a drug or snack), at the cost of long-term (health) goals, may contribute to a vulnerability to develop substance use addiction, obesity, or binge eating episodes. Similarly, a stronger behavioural activation towards distracting but immediately rewarding cues (e.g., an incoming text), at the cost of long-term occupational or educational achievements, may be associated with ADHD symptoms. In contrast, excessive control over Pavlovian responses towards immediate reinforcers (e.g., food) and/or a preoccupation with future (e.g., weight) goals may be associated with symptoms of anorexia nervosa.1 Providing support for the general relevance of Pavlovian biases in mental health, several studies have found valence-driven Pavlovian biases (i.e., comparing the anticipation of rewards versus punishments) to be associated with individual differences in substance use (Garbusow et al., 2022), mood and anxiety symptoms (Huys et al., 2016; Mkrtchian et al., 2017; Nord et al., 2018; Peterburs et al., 2022; Scholz et al., 2020; but also see Moutoussis et al., 2018), suicidal thoughts and behaviours (Millner et al., 2018), ADHD hyperactivity/impulsivity symptoms (Geurts, Den Ouden, et al., 2022), borderline personality disorder symptoms (but only for neural Pavlovian bias-related signals, not for behaviour; Geurts, Van Den Heuvel, et al., 2022), schizophrenia (Albrecht et al., 2016; Montagnese et al., 2020; Morris et al., 2015; but also see Metts et al., 2022), and psychopathic tendencies (Geurts, Von Borries, et al., 2022), although the direction of the reported statistically significant associations (i.e., positive or negative) has been somewhat inconsistent.

The present study had two main goals. First, using the intertemporal go/no-go task developed previously (Burghoorn et al., 2024), we aimed to replicate the intertemporal Pavlovian bias effect, thereby examining its robustness. Second, we investigated the associations between this intertemporal Pavlovian bias and symptoms of alcohol use disorder, nicotine dependence, ADHD, depression, trait anxiety, disordered eating, body mass index (BMI), and impulsivity, all of which have previously been associated with increased or decreased intertemporal impatience. We investigated these associations in a general population sample, thus treating mental health and intertemporal impatience as dimensional constructs, and tested for their continuous associations. Third and finally, while not a central goal of the present study, we tested whether the associations between the mental health symptoms and the level of intertemporal impatience (as assessed using a choice titration procedure, explained further below) were similar to those previously reported in the literature.

Study hypotheses

We expected to replicate the intertemporal Pavlovian bias effect and thus predicted that go responding in the go/no-go task would increase in anticipation of an immediate reward (i.e., in the immediate reward trials) compared with the anticipation of a preference-matched delayed reward (i.e., in the delayed reward trials).

Next, following the literature on intertemporal impatience and mental health discussed earlier and the proposed role of the intertemporal Pavlovian bias as mechanism of impatience, we expected the strength of the intertemporal Pavlovian bias to be associated with increased symptoms of alcohol use disorder, nicotine dependence, ADHD, depression, and BMI. For several of these mental health variables, we not only assessed total symptom scores, but additionally distinguished between different symptom subscales. Specifically, for alcohol addiction, we distinguished between general alcohol use and alcohol-related problems, as research suggests them to be separate constructs (Horváth et al., 2023), and because we expected to observe more interindividual variation in general alcohol use compared with alcohol-related problems. Similarly, for nicotine addiction, we distinguished between nicotine dependence in the full sample, smoking status (i.e., whether someone smokes or not), and nicotine dependence in a subset of smokers. For ADHD, we distinguished between hyperactivity/impulsivity symptoms and inattention symptoms, as recommended by Jackson and Mackillop (2016). We hereby explored the possibility that different subsets of symptoms may be driven by different psychological mechanisms, thus accounting for within-disorder heterogeneity of symptoms and underlying mechanisms. Scheres et al. (2008) observed hyperactivity-impulsivity but not inattention symptoms to be associated with increased intertemporal impatience, raising the possibility that the former is more strongly associated with the intertemporal Pavlovian bias than the latter.

For disordered eating, we distinguished between symptoms reflecting a lack of control over eating and other eating disorder symptoms (revolving around food restriction and a preoccupation with weight loss). We expected a lack of control over eating to be associated with a stronger intertemporal Pavlovian bias, reflecting a stronger effect of cues signalling immediate gratification (e.g., food rewards) on behaviour. This aligns with research showing increased intertemporal impatience in bulimia nervosa and binge-eating disorders, both of which are characterized by uncontrolled binge-eating episodes (Amlung et al., 2019). We did not have a directional hypothesis for the remaining eating disorder symptoms. Because we expected these symptoms to be positively correlated with the symptoms reflecting a lack of control overeating, one may expect a similarly increased Pavlovian bias. At the same time, the strong focus on food restriction and weight loss that characterizes these symptoms may require strong control over Pavlovian responses towards immediate rewards, as well as a strong focus on delayed rewards, possibly resulting in a reduced or even reversed intertemporal Pavlovian bias. This is consistent with previously reported negative associations between anorexia nervosa—particularly the restricting subtype—and intertemporal impatience (see Amlung et al., 2019 and Lempert et al., 2018 for reviews).

Findings on the association between intertemporal impatience and anxiety have been mixed; studies showed increased impatience (Keidel et al., 2024; Rounds et al., 2007; Xia et al., 2017; Zhang et al., 2020), decreased impatience (Steinglass et al., 2017), or no association with impatience (Jenks & Lawyer, 2015; Takahashi, 2004; Worthy et al., 2014). Even within studies, findings have been mixed, depending on the type of analytic approach used (Levitt et al., 2022) or the type of anxiety assessed (Levin et al., 2018). Similarly, while some studies on valence-driven Pavlovian biases observed an increased Pavlovian bias in association with anxiety—albeit in the punishment domain (Mkrtchian et al., 2017; Peterburs et al., 2022)—others found a decreased bias (Scholz et al., 2020), the latter of which was theorized to result from a more ruminative and careful decision style. Because the present study only included rewards (not punishments), and following the inconsistency in previous research findings on anxiety, intertemporal impatience, and valence-driven Pavlovian biases, we did not have a directional hypothesis regarding the association between the intertemporal Pavlovian bias and anxiety.

Finally, we expected the intertemporal Pavlovian bias to be associated with increased impulsivity, reflecting an impaired ability to override the Pavlovian approach response elicited by immediate rewards. This hypothesis follows the strong conceptual link between impulsivity and intertemporal impatience, as well as several studies showing an empirical association between the two (Baumann & Odum, 2012; Jauregi et al., 2018; Keidel et al., 2024; Malesza & Kalinowski, 2021; Malesza & Ostaszewski, 2016; Zhou et al., 2021). Impulsivity was assessed by using the Barratt Impulsiveness Scale (BIS-11; Patton et al., 1995), which, in addition to a total score, results in separate scores for attentional impulsivity, motor impulsivity, and nonplanning impulsivity. Because the intertemporal Pavlovian bias involves a conditioned approach (go) response, we expected the bias to be most strongly associated with motor impulsivity.

Our hypotheses on the association between mental health and intertemporal impatience (as assessed using the choice titration procedure) followed the hypotheses on mental health and the intertemporal Pavlovian bias described above.

Methods

The study’s research questions, hypotheses, design, sample size, and analyses were preregistered on the Open Science Framework (https://osf.io/4qczy). The OSF page also contains the study materials, data, and analysis code (https://osf.io/p5bz9/).

Participants

Participants were recruited on Prolific (https://www.prolific.com/) if they indicated to be fluent in English, live in a country that uses euros as its currency (as study rewards were presented in euros), have normal or corrected-to-normal vision, have normal colour vision, and have an approval rate on Prolific of at least 95%. Of the 435 participants who took part in the study, 46 were excluded from data analysis, either because they did not meet our preregistered data quality criteria (n = 44, see Supplementary Information 1 [S1] for details), because they did not complete the entire study (n = 1), or did not complete the study in English (n = 1). A simulation-based power analysis showed that the remaining sample of 389 participants (140 females, 243 males, 5 nonbinary, 1 other; M_age = 31.37, SD_age = 9.07) would provide 90–95% power to replicate the intertemporal Pavlovian bias effect observed in Burghoorn et al., (2024; unstandardized effect of reward of b = 0.17 on log-odds scale), and at least 95% power to replicate an interaction between the intertemporal Pavlovian bias and total alcohol use disorder scores observed in a pilot study (N = 51; unstandardized interaction of b = 0.34 on log-odds scale; smaller interaction effect sizes of 0.25 and 0.20 would be detected with 87% and 77% power, respectively). Following our preregistered trial-level exclusion criteria (see S1), 0.32% of the go/no-go trials of the final sample were excluded.

The study fell under a research line that received ethics approval from the local institutional review board prior to data collection (number: ECSW-2024–053), and the study was performed in accordance with the ethical standards of the Declaration of Helsinki. Digital informed consent was obtained from all individual participants. Participation was compensated with £5.25 (Prolific uses Great British Pounds as its currency). In addition, participants took part in a performance-contingent lottery, in which they could win one of the rewards they earned during the go/no-go task (described below).

General procedure

The experimental procedure was programmed in jsPsych (version 7.0.0; de Leeuw & Gilbert, 2023) and could be completed in a Mozilla Firefox, Safari, or Microsoft Edge browser on a desktop or laptop computer. The experimental timeline is displayed in Fig. 1. The complete experiment took approximately 35 min.

Fig. 1 — Experimental timeline. *Note.* Experimental timeline, displaying the order of tasks administered. Whether the reward valuation task was administered before (v1) or after (v2) the go/no-go task, and whether the valuation task was a rating or ranking task, was counterbalanced across participants

Choice titration I

The choice titration procedure was used to derive a participant-specific immediate reward that was preference-matched to a delayed reward of €28 in 120 days. Details are reported in Burghoorn et al. (2024). Summarized, participants first completed the Monetary Choice Questionnaire (Kirby et al., 1999), which consists of 27 choices between varying immediate and delayed monetary rewards. Choices on this task were used to derive a participant-specific discount rate, which was then used to compute the starting amount of the immediate reward in the subsequent adaptive choice titrator. Each trial of the adaptive choice titrator offered a choice between an immediate reward (€X today, with X being an even integer between €0 and €28) versus a fixed delayed reward of €28 in 120 days. If a participant chose the immediate [delayed] reward, the immediate reward amount on the next trial was decreased [increased] with €2. The titrator continued until participant reached a stable window of six trials during which the immediate reward did not deviate by more than €2. Participants who failed to reach stability after 50 trials were excluded from data analyses (n = 5). Participants who indicated to prefer €0 today over €28 in 120 days were asked to confirm their choice in a subsequent, identical trial. If they again preferred €0, they were excluded from the analyses (n = 1), because this indicated that they preferred not to receive any reward, undermining a critical premise of the study. The titration procedure was incentivized through a lottery at the end of the experiment, in which participants had the chance of winning a monetary reward that was determined by their choice preferences (described below).

Intertemporal Go/No-Go task

Next, participants completed the intertemporal go/no-go task as detailed in Burghoorn et al., (2024; originally adapted from Scholz et al., 2022) and as displayed in Fig. 2. The task was framed as a gem game, during which participants had to learn which two gems (i.e., cues) to collect (go response) and which two gems to leave behind (no-go response). For two of the four gems, a correct response resulted in the delayed larger reward of €28 in 120 days. For the other two gems, a correct response resulted in the participant-specific immediate reward of €X today. For all four gems, an incorrect response resulted in no reward. Orthogonalizing the required action (go/no-go) and available reward (immediate/delayed) resulted in four conditions: go to win immediate reward trials; go to win delayed reward trials; no-go to win immediate reward trials; and no-go to win delayed reward trials. For each participant, the four cues were randomly assigned to conditions. The required response (go/no-go) for each of the cues had to be learned by trial and error. Although the available reward (immediate/delayed) could also be learned by trial and error, it was additionally signalled by a coloured edge (orange/blue) around the cue. The purpose of this coloured edge was to increase the salience of the available reward (immediate/delayed), and to ensure that any reward effects would be present early on in the task (hereby following Burghoorn et al., 2024; Scholz et al., 2022; Swart et al., 2017, 2018). Which edge colour signalled which reward was counterbalanced across participants and was instructed to the participants before the start of the task.

Fig. 2 — Go/No-Go task. *Note.* Design of the go/no-go task. A: The go/no-go task consisted of four conditions, each signalled by a unique visual cue: go to win immediate reward, go to win delayed reward, no-go to win immediate reward, no-go to win delayed reward. The cue signalled the required instrumental action (go/no-go) and the reward available upon giving a correct response (immediate/delayed). Participants had to learn the required instrumental action by trial and error; the reward was instructed through the coloured edge around the cue, and could also be learned by trial and error. Cues were randomly assigned to conditions, and which cue edge (blue/orange) indicated which reward (immediate/delayed) was counterbalanced across participants. Each condition was presented 50 times in pseudorandom order (i.e., in random order except that the same condition could not be presented more than twice in a row). B: Example of a go to win immediate reward trial. The trial started with a fixation cross (inter-trial interval), after which the cue was presented. Upon cue presentation, participants had to respond within 600 ms, after which feedback was provided. If participants gave a correct response for an immediate reward cue (such as that presented in B), the immediate reward was presented to go into a chest close to the participants’ agent on a timeline. If participants gave a correct response for a delayed reward cue (not presented in B), the delayed reward was presented to go into a chest far away from the participants on the timeline. If participants gave an incorrect response (regardless of the cue), both chests remained closed. Feedback was probabilistic; in only 80% of the trials, the feedback presented corresponded to the correctness of the participant’s response

Upon cue presentation, participants had 600 ms to make a response by pressing the space bar (go response) or doing nothing (no-go response). After the response window, the outcome was presented for 1,700 ms, by presenting a reward going into a reward chest that was either close to the participant on a timeline (immediate reward) or far away on the timeline (delayed reward), or by presenting only closed chests (no reward). Outcome feedback was probabilistic, such that in 80% the trials, the outcome presented to the participant was consistent with the correctness of their response. Each of the four conditions was presented 50 times in pseudorandom order (i.e., random order with the exception that the same condition could not be presented more than twice in a row). The 200 trials were divided in four blocks of 50 trials, divided by 20-s breaks.

Participants received extensive task instructions and were required to reach 80% performance in a round of five practice trials before proceeding to the full task. Participants who required more than six practice rounds were excluded from the analyses (n = 2). The task included four attention checks (separated by a minimum of 25 trials), during which a crosshair target appeared on the screen for 1,000–2,000 ms (jittered), and participants had to press the space bar within 1,500 ms upon disappearance of the target. To incentivize task performance, participants were informed that they would be entered into a lottery at the end of the experiment, during which they had the chance to win the outcome they received on one randomly selected go/no-go trial (hereby incentivizing response accuracy) and that their response speed increased their chance of winning the lottery (hereby incentivizing response speed). See Burghoorn et al. (2004) for the equation used to determine the per-participant lottery chance; 13 participants won the lottery.

Mental health measures

The instruments used to assess the mental health symptoms, and their descriptive statistics, are outlined in Table 1. A detailed description of all questionnaires can be found in S2. The order in which the questionnaires were administered was randomly determined per participant. The STAI-T-5 and BIS-11 contained one attention check, instructing participants which answer to give. After completing the mental health questionnaires, participants also completed a few questions on their menstrual cycle (if applicable); these data were not of interest (and therefore not analysed) for the present study.

Table 1.

Mental health measures

Variable	Scale	Possible range	Observed range	M (SD)	Mdn	Cronbach’s α
Total alcohol use disorder	AUDIT Saunders et al. (1993)	0–40	0–29	3.61 (4.20)	2	.86
Alcohol use	AUDIT	0–12	0–12	2.43 (2.15)	2	.75
Alcohol problems	AUDIT	0–28	0–19	1.18 (2.49)	0	.81
Smoking status	“Do you smoke (e-)cigarettes?”	Yes / No	16% smokers
Nicotine dependence in full sample	FTND Heatherton et al. (1991)	0–11	0–8	0.57 (1.58)	0	.86^a
in smokers (n = 63)		0–10	0–7	2.54 (2.23)	2
Depression	DASS-21 Lovibond and Lovibond (1995)	0–21	0–21	5.55 (5.23)	4	.93
Trait anxiety	STAI-T-5 Zsido et al. (2020)	0–15	0–15	5.98 (3.35)	6	.85
Total ADHD	ASRS-v1.1 Kessler et al. (2005)	0–72	0–62	27.42 (9.89)	27	.88
ADHD inattention symptoms	ASRS-v1.1	0–36	0–31	14.58 (5.69)	14	.85
ADHD hyperactivity / impulsivity symptoms	ASRS-v1.1	0–36	0–31	12.84 (5.58)	13	.80
Total disordered eating	EDE-QS Gideon et al. (2016)	0–36	0–29	7.26 (6.70)	5	.88
Lack of control over eating	EDE-QS	0–6	0–6	0.86 (1.28)	0	.75
Remaining eating disorder symptoms	EDE-QS	0–30	0–26	6.40 (5.82)	5	.85
Total impulsivity	BIS-11 Patton et al. (1995)	0–90	5–64	30.07 (10.42)	29	.86
Attentional impulsivity	BIS-11	0–24	0–24	8.59 (3.91)	9	.76
Motor impulsivity	BIS-11	0–33	1–21	9.65 (4.07)	9	.68
Non-planning impulsivity	BIS-11	0–33	1–26	11.84 (5.04)	12	.77
Body Mass Index^b			16.07 – 61.86	24.17 (4.78)	23.15

Open in a new tab

Overview of the mental health variables assessed in the study, the instruments used to assess them, and their descriptive statistics. M = mean, SD = standard deviation, Mdn = median. Cronbach’s α was computed for all scales consisting of multiple items

^a Given the limited subsample of smokers, we only computed Cronbach’s α in the total sample

^bOne impossible BMI score was removed from the sample. Therefore, all analyses involving BMI scores were performed on 388 participants

Secondary measures

In addition to the choice titration, go/no-go task, and mental health questionnaires, which formed the three main parts of the experiment, we administered several other short tasks, described in the following sections.

Reward valuations

The choice titration procedure resulted in an immediate and delayed reward that were matched based on revealed choice preferences. To examine whether the two rewards were also valued similarly outside a choice context, we administered a separate reward valuation task. We included two versions of this task, a rating and ranking version, allowing us to compare valuations across different preference elicitation methods. Which of the two versions of the valuation task a participant completed, and whether they completed this before or after the go/no-go task, was counterbalanced across participants.

Rating task

The rating task was identical to that used in Burghoorn et al. (2024). Participants were asked to rate how attractive they found each of the two rewards, one by one, using a slider on a visual analogue scale ranging from very unattractive (0, left endpoint) to very attractive (100, right endpoint). The left endpoint additionally included as anchor a delayed reward that was €1 lower in amount than the immediate reward of the preference-matched reward pair, and 1 day longer in delay than the delayed reward (i.e., 121 days), hereby representing a relatively very unattractive reward. The right endpoint included as anchor an immediate reward that was €1 higher in amount than the delayed reward of the preference-matched reward pair (i.e., €29), representing a relatively very attractive reward. The two reward ratings were presented in random order.

Ranking task

The ranking version of the valuation task was adapted from Luo et al. (2009). Participants were asked to rank the two rewards based on how attractive they found each of them. They were instructed to give the reward that they found most attractive rank 1 and the reward that they found least attractive rank 2, or if they found both rewards equally attractive to give both rewards rank 1. In line with Luo et al. (2009), participants were instructed not to think about which reward they would choose if they compared them but simply to think about how attractive they found each reward. The two rewards were simultaneously presented on the screen; which reward was presented on the top or bottom of the screen was randomly determined.

Cue ratings

To examine pre-existing differences in subjective valuation of the cues (i.e., gems) used in the go/no-go task, we asked participants to rate how attractive they found each of the cues. Ratings were provided using a slider on a visual analogue scale ranging from very unattractive (0, left endpoint) to very attractive (100, right endpoint). The cue rating task was again presented after the go/no-go task, allowing us to examine whether the cue ratings changed over the course of the experiment. In both cue rating tasks, the cues were presented in random order.

Choice titration II

To examine whether the degree of intertemporal impatience, as assessed using the first choice titrator, remained stable across the experiment, we again administered the adaptive part of the choice titration procedure after the go/no-go task. The starting value of the immediate reward was identical to the adaptive titrator administered before the go/no-go task.

Data analyses

Statistical models

The majority of the data were analysed with Bayesian mixed-effects models, using the brms package (version 2.21.0; Bürkner, 2018) in R (version 4.4.1; R Core Team, 2024). The statistical models used for our primary analyses are outlined below; the models for all secondary analyses are specified in the respective results or Supplementary Information sections. In all models, brms’ weakly informative default priors were used, categorical predictors were sum-to-zero coded, and continuous predictors were mean-centered. A Gaussian model family was used unless mentioned otherwise. We accounted for by-participant random variation using a maximal random-effects structure. Effects were denoted as statistically significant if the 95% highest density interval (HDI) did not include 0 (except when the HDI levels were corrected for multiple testing, as detailed below). If we computed Pearson correlations, these were denoted as statistically significant when p < 0.05 (except when correcting for multiple testing). Visualizations of the results were created using the packages emmeans (version 1.10.3; Lenth, 2019) and ggplot2 (version 3.5.1; Wickham, 2016).

Intertemporal Pavlovian bias models

Response bias

Our main statistical model on the intertemporal Pavlovian bias included response (go/no-go) as dependent variable, a fixed intercept, required action (go/no-go), reward (immediate/delayed), task block (1–4, modelled as centered linear predictor), and their interactions as fixed effects. Random effects included a random intercept, a random slope for all fixed effects, and all random correlations varying over participants. Responses were modelled with a Bernoulli model family to account for their binary nature.

RT bias

As preregistered, we also explored whether any intertemporal Pavlovian bias effect was observed on the response times (RTs), reflecting an effect of the available reward on response invigoration. To this end, we reran our main intertemporal Pavlovian bias model, but with response times (from both correct and incorrect go responses) as dependent variable. A shifted lognormal family was used to account for the non-normal RT distribution.

Mental health moderation

To examine whether the intertemporal Pavlovian bias was associated with the mental health symptoms, we tested whether the mental health variables moderated the Pavlovian bias. To this end, we reran our Pavlovian bias models (both for responses and RTs), but this time included a mental health score as additional fixed effect, allowing it to interact with all other fixed effects. To prevent convergence issues resulting from overly large models, we ran a separate model for each mental health variable, and corrected for multiple testing by applying a Benjamini–Hochberg correction to the HDIs (with a false discovery rate of 5%; Benjamini & Hochberg, 1995). Our main test involved the two-way interaction between the reward effect and the mental health score. We additionally examined whether this two-way interaction was further moderated by the required action, task block, or both.

Intertemporal impatience and mental health

We also tested whether the level of intertemporal impatience, as assessed using the first choice titrator, was associated with any of the mental health variables, in line with the literature on intertemporal impatience and mental health (Amlung et al., 2019; Lempert et al., 2018). Intertemporal impatience was quantified as the reward amount of the immediate reward preference-matched to the delayed reward of €28 in 120 days. Because a higher immediate reward amount indicates more patience, we reverse-coded these patience scores to create an impatience score (1–28, with higher scores indicating more impatience). We subsequently computed the Pearson correlations between the immediate reward amount derived from the choice titrator, and the mental health variables. A Benjamini–Hochberg correction was applied to correct for multiple testing. Finally, we computed the bivariate correlations between all mental health variables to gain insight into possible comorbidity between mental health symptoms.

Reinforcement learning models

To study the computational mechanisms that may underlie the intertemporal Pavlovian response bias, and its possible association with mental health symptoms, we fitted a series of nested reinforcement learning (RL) models to the data of the go/no-go task (following Burghoorn et al., 2024, inspired by Guitart-Masip et al., 2012b). The first and most simple model was a Rescorla-Wagner model (M0), in which action weights (w_t) are fully determined by action values (Q_t).:

Action values are updated on a trial-by-trial basis, based on prediction errors, i.e., the difference between the expected reward (Q_t-1) and the obtained reward (r_t−1), scaled by the learning rate (α). A softmax function was used to transform action weights into the probability of making a go response (p). The inverse temperature parameter τ reflects the response stochasticity, i.e., the degree to which responses are determined by action weights:

The next model, M1, includes an extended softmax function with ξ that captures irreducible noise in action selection, due to, e.g., motor or attentional lapses:

M2 adds a go bias parameter b to the computation of the action weight, capturing a general tendency to give go responses:

M3 captures the hypothesized intertemporal Pavlovian bias through a cue-response bias π, which reflects the elicitation of a Pavlovian go response by cues that signal immediate (versus delayed) rewards, implemented through an increased action weight of the go response:

Here, V( Inline graphic ) represents the reward signalled by the cue. In line with previous research, the values of V() were fixed at 1 for immediate rewards and − 1 for delayed rewards, capturing the instruction of the reward identity by the coloured edge around the cue (Burghoorn et al., 2024; Scholz et al., 2022; Swart et al., 2017, 2018; van Nuland et al., 2020).

In Burghoorn et al. (2024), we fitted two additional models, M4 and M5, which reflect alternative and/or complementary computational mechanisms possibly driving the intertemporal Pavlovian bias. We observed that whereas these models provided a slight advantage compared to M3 in terms of model fit, M3 was the only model that showed successful parameter recovery, model recovery, and posterior predictive performance. In the present study, we again fitted M4 and M5 to examine whether the results were consistent with our previous study or whether these models could be successfully validated in the larger sample of the current study. M4 revolves around a learning bias: It reflects enhanced learning of (1) go responses that are followed by immediate (versus delayed) rewards and (2) no-go responses that are followed by delayed (versus immediate) rewards. The learning advantage is reflected by an increased learning rate:

M5 forms a combination of M3 and M4 by including both the cue-response and the learning bias.

Model fitting, comparison, and validation

The models were fitted using maximum a posteriori (MAP) estimation, which aims to find the participant-specific posterior mode. As in Burghoorn et al. (2024), the learning rate and irreducible noise parameters were constrained between 0 and 1, the inverse temperature was constrained between 0 and 50, and the go bias and Pavlovian bias parameters were constrained between − 3 and 3. A Gamma(3,0.3) prior was used for the inverse temperature parameter, and a Gaussian(0,1) prior was used for the go bias and the cue-response bias parameters. Parameters were optimized with a differential evolution algorithm implemented in the DEoptim package (version 2.2.8; Mullen et al., 2011). Models were compared using Akaike’s Information Criterion (AIC), with smaller values indicating a better fit, and model frequency, the proportion of participants for which each model had the lowest AIC. We validated the best-fitting models using parameter recovery, model recovery, and posterior predictive checks.

Associations with mental health

We computed Pearson correlations between the parameters of the winning RL model and the mental health variables assessed in the study, with a specific focus on the Pavlovian bias parameter(s). A Benjamini–Hochberg correction was applied to the threshold for statistical significance (i.e., the α-level).