Enhanced decision-making in nicotine dependent individuals who abstain: A computational analysis using Hierarchical Drift Diffusion Modeling

Abstract

Background:

Variability in decision-making capacity and reward responsiveness may underlie differences in the ability to abstain from smoking. Computational modeling of choice behavior, as with the Hierarchical Drift Diffusion Model (HDDM), can help dissociate reward responsiveness from underlying components of decision-making. Here we used the HDDM to identify which decision-making or reward-related parameters, extracted from data acquired in a reward processing task, contributed to the ability of non-treatment-seeking smokers to abstain from cigarettes during a laboratory task.

Methods:

80 adults who smoke cigarettes completed the Probabilistic Reward Task (PRT) - a signal detection task with a differential reinforcement schedule - following smoking as usual, and the Relapse Analogue Task (RAT) - a task in which participants could earn money for delaying smoking up to 50 min - after a period of overnight abstinence. Two cohorts were defined by the RAT; those who waited either 0-minutes (n=36) or the full 50-minutes (n=44) before smoking.

Results:

PRT signal detection metrics indicated all subjects learned the task contingencies, with no differences in response bias or discriminability between the two groups. However, HDDM analyses indicated faster drift rates in 50-minute vs. 0-minute waiters.

Conclusions:

Relative to those who did not abstain, computational modeling indicated that people who abstained from smoking for 50 minutes showed faster evidence accumulation during reward-based decision-making. These results highlight the importance of decision-making mechanisms to smoking abstinence, and suggest that focusing on the evidence accumulation process may yield new targets for treatment.

Introduction

It is critical to define factors associated with continued substance use as abstinence rates remain very low. For instance, tobacco smoking abstinence rates have been estimated to be as low as 10% for those who attempt to quit (U.S. Department of Health and Human Services, 2020). Disruptions in reward function have long been identified as one factor associated with relapse, where one’s return to using nicotine would reverse a withdrawal-induced decrement in reward processing for non-drug rewards (Pergadia et al., 2014). However, the primary function of rewards is to affect downstream decisions, suggesting that a careful inquiry into decision-making mechanisms might provide insight into relapse vs. abstinence. Decision-making is a complex process that can be parsed using computational models. To identify decision-making mechanisms linked with continued nicotine use, we modeled decision-making during a reward task in people who smoke cigarettes, who either did or did not abstain from smoking in a laboratory-based abstinence task.

Reward-based decision-making can be measured using the Probabilistic Reward Task (PRT; Pizzagalli, Jahn, & O’Shea, 2005). The PRT, which is sensitive to dopaminergic modulation (Pizzagalli et al., 2008), assesses reward responsiveness by measuring the ability to modulate behavior based reward delivery (Pizzagalli et al., 2005). Using the PRT, Pergadia and colleagues (2014) found that nicotine withdrawal reduced reward responsiveness in both clinical and preclinical samples. The authors suggested that blunted reward responsiveness during nicotine withdrawal may contribute to relapse, as acute nicotine re-exposure mitigated the withdrawal-induced reward deficit. Furthermore, nicotine is known to enhance reward responsiveness, as measured by the PRT, in non-smoking individuals (Barr, Pizzagalli, Culhane, Goff, & Evins, 2008), and nicotine is thought to normalize deficits in reward responsivity in those with a history of major depressive disorder (Janes et al., 2015). While performance on the PRT provides a quantifiable way to measure reward responsiveness following acute and chronic nicotine administration, the primary dependent variable produced by the PRT—namely, response bias—tells us little about the latent processes that contribute to reward-related decision-making.

The drift diffusion model (DDM; Ratcliff, 1978; Ratcliff & McKoon, 2008) can address this problem. When fit to trial-level choice and response time (RT) data, the DDM provides quantitative estimates of key decision-making mechanisms. In this way, the DDM can identify underlying cognitive processes that give rise to reward-related behaviors (Eikemo, Biele, Willoch, Thomsen, & Leknes, 2017; Johnson, Hopwood, Cesario, & Pleskac, 2017; Lawlor et al., 2020), which may in turn help uncover deficits in clinical samples that cannot be detected with traditional analyses of behavior (White, Ratcliff, Vasey, & McKoon, 2010). Briefly, the DDM conceptualizes behavior in two-alternative forced choice tasks, such as the PRT, as reflecting a process of evidence accumulation to decision boundaries reflecting two separate responses. On every trial, the model proposes that once a stimulus requiring one of two responses is presented, participants begin extracting evidence from perceptual systems that should indicate which response to make. Each individual sample of evidence is noisy, thus this process is performed repeatedly until enough evidence has accrued such that either response boundary is reached. At that point, the corresponding response is executed. Importantly, the DDM is applied to both choice and response time (RT) data, thus it offers a well-constrained model of behavior (i.e., relative to models that explain choices but not RTs).

The DDM has four main parameters: decision threshold, which indicates the distance between the two decision boundaries, such that wider thresholds would support slower but more accurate choices than narrow ones; drift rate, indicating average speed of the evidence accumulation process; starting point bias, which indicates whether the evidence accumulation process begins midway between the two boundaries or is, instead, shifted towards one of them so as to increase the likelihood of hitting that boundary; and non-decision time, which captures the time needed to perceive the stimuli and to respond once a decision has been made. Recent work has demonstrated that DDM has been successfully applied in clinical contexts, including in people who smoke (Copeland, Stafford, & Field, 2023) and may relate to (or even underlie) executive function deficits in several forms of psychopathology (Lawlor et al., 2020; Sripada & Weigard, 2021). This highlights the clinical utility of understanding the underlying mechanisms of choice in the context of tobacco use. In the current study, we estimated these four parameters by analyzing PRT data with the Hierarchical Drift Diffusion Model (HDDM; Wiecki, Sofer, & Frank, 2013), a Bayesian version of the DDM that uses group-level priors to constrain parameter fits at the level of individuals, which has been shown to support more precise estimates in smaller samples.

To determine whether decision-making processes or reward sensitivity itself modulates the choice to smoke, we assessed behavior on the PRT in non-treatment-seeking individuals who smoke and modeled their behavior using the HDDM. Individuals were divided into two groups based on their performance on the Relapse Analogue Task (RAT; McKee, 2009). This behavioral task directly pits smoking rewards against non-smoking rewards by creating a condition where abstinent individuals are given the option to engage in smoking or delay smoking in exchange for small amounts of money (McKee, 2009). In our study, we contrasted individuals who decided to smoke immediately (0-minute waiters), vs. those who delayed smoking for the entire test period of the RAT (50-min waiters); thus providing a test of the hypothesis that decision-making mechanisms, captured by HDDM, are relevant to smoking-related decisions. Identifying what decision-making or reward-related parameters contribute to the ability to abstain (or not) could ultimately inform which cessation strategy could be most effective and facilitate individualized treatment plans (Caponnetto & Polosa, 2008).

Methods

Study sample

Analyses were conducted on a total of 80 non-treatment seeking adults who smoked cigarettes daily. Individuals were recruited from the community as part of a larger parent study on smoking and personality at the University of Southern California (USC) (Leventhal et al., 2014). Participants were included if they were 18 years old or older, smoked regularly for 2+ years, and currently smoked more than 9 cigarettes/day. Exclusion criteria were a current substance dependence other than nicotine (based on the 4^th Diagnostic and Statistical Manual of Mental Disorders), mood disorder, psychotic symptoms or use of psychiatric medications, breath carbon monoxide (CO) levels of <10 ppm at baseline, or breath alcohol concentration (BAC) > 0. All procedures were approved by the USC Institutional Review Board. Informed consent was obtained at the onset of the study, and participants were debriefed and paid $300 total for their participation.

Procedures

The current work is based on two visits of the larger parent study: a baseline visit (no restriction on smoking), and an abstinent visit. For the abstinent visit, individuals were instructed not to smoke after 8pm the night before the visit (minimum 16 hours abstinence), and abstinence was verified by assessing breath CO levels; levels of < 10 ppm were considered abstinent. The baseline and abstinence visits were separated by 2 – 14 days. The PRT was conducted at the baseline visit following smoking as usual, while the RAT was administered at the abstinence visit.

Probabilistic Reward Task (PRT)

Participants completed the PRT during their baseline visit to assess their reward responsivity and reward-based decision-making. Following procedures outlined in Pizzagalli et al. (2005), participants completed three blocks of 100 trials with a 30-second break between each block; the task lasted approximately 30 minutes total. Each trial began with the presentation of a cross for 500ms (Figure 1), followed by presentation of a mouthless cartoon face. After a 500ms delay, either a short (11.5 mm) or long (13.0 mm) mouth was presented on the face for 100ms, and participants were asked to identify which type of mouth was presented by pressing either the “z” or the “/” key for short vs. long (assignment of key lengths was counterbalanced across subjects). For each block, 40 correct trials were followed by reward feedback (“Correct!! You won 5 cents”). Asymmetrical reinforcement was applied such that correct identification of one type of mouth (the ‘rich’ stimulus) was rewarded three times more frequently (30 of 40 trials) than correct identification of the other type of mouth (the ‘lean’ stimulus; 10 of 40 trials). For half the participants short and long mouths served as the rich and lean stimuli, respectively, whereas for the other half of participants, these contingencies were reversed. No feedback was provided on non-reward trials. All participants “earned” $6.00 at the end of the PRT session.

Figure 1. Schematic diagram of the Probabilistic Reward Task.

In each trial of the task, participants decided whether a short or long mouth stimulus was presented by pressing the “z” or “/” key.

Relapse Analogue Task (RAT)

Participants completed the RAT to assess the relative reward value of smoking during their abstinent visit. Participants were given a tray containing 8 cigarettes, a lighter, and an ashtray. They were instructed that they could begin smoking at any point over the next 50 minutes, but for every 5 minutes they remained abstinent they would receive $0.20. Thus, participants were eligible to receive a maximum of $2 for remaining abstinent throughout the entire 50-min period (maximum reward was set to $2 based on pilot work indicating that this value produced maximum inter-individual variation). The smoking delay period was operationalized as the number of minutes (0–50 min) the individual remained abstinent. The RAT ended once participants chose to smoke or 50-minutes passed. Participants were informed prior to the RAT that the study visit would end for everyone at 4:00 pm, which helped prevent the possibility that individuals might smoke during the RAT to end the visit sooner.

Analyses

Group segmentation

Upon inspection of the RAT data’s distribution, it was evident that most participants (from the larger parent study) fell into two groups: subjects who smoked immediately (0-min waiters) or did not smoke for the entire 50 min period (50-min waiters). Therefore, all subsequent analyses focused on determining differences in behavioral and computational parameters between these two groups.

PRT Calculations and Quality Control

Outliers were defined by extreme response times (RTs) (< 150ms, > 2500ms), or if the log transformed RT exceeded the participant’s mean ±3 S.D. Participant datasets were excluded if any of the 3 blocks contained: > 20 RT outliers, < 24 rich or 7 lean rewards, a rich-to-lean reward ratio lower than 2.5, or lower than 40% correct accuracy.

Behavioral metrics of performance on the PRT included response bias, which captures the tendency to report having seen the rich stimulus, discriminability (the ability to accurately differentiate between stimuli), and response time (Pizzagalli et al., 2005). Cumulative reward totals (i.e., number of rewards received over all blocks) were also recorded. Previously published signal detection analyses (Pizzagalli et al., 2005) were used to calculate response bias and discriminability:

$$\text{Response Bias: log }b=\frac{1}{2}\text{log}\left(\frac{\text{RICHcorrect}*\text{LEANincorrect}}{\text{RICHincorrect}*\text{LEANcorrect}}\right)$$

$$\text{Discriminability: log }d=\frac{1}{2}\text{log}\left(\frac{\text{RICHcorrect}*\text{LEANcorrect}}{\text{RICHincorrect}*\text{LEANincorrect}}\right)$$

Analyses of behavioral metrics were conducted in SPSS Version 29. Separate 2-way ANOVAs evaluated main effects of Group (0-min waiters vs. 50-min waiters) and Block (1, 2, 3), as well as their interaction, for response bias and discriminability. A 3-way ANOVA evaluated the main effects of Group, Block, and Stimulus-type (rich vs. lean), as well as their interactions, for RT. A between-groups t-test evaluated group differences in cumulative reward totals. Pearson’s correlations evaluated relationships between cumulative reward and response bias versus discriminability, averaged over all trials, and the strength of these relationships was compared using a test for dependent correlations (Meng, Rosenthal, & Rubin, 1992). Multiple linear regressions measured the relationship between behavioral metrics (response bias, discriminability, cumulative reward) and HDDM parameters (see below).

Hierarchical Drift Diffusion Modeling (HDDM)

The HDDM was fit to trial-level RT and choice data following published recommendations (Wiecki et al., 2013), and as done in our prior publications (Dillon et al., 2022; Lawlor et al., 2020). Briefly, the HDDM is initialized with priors that reflect established findings in the DDM literature, and the Markov Chain Monte Carlo (MCMC) method is used to fit the model to the data by estimating the joint posterior distribution for all parameters. The HDDM was “stimulus-coded” rather than “accuracy coded”, meaning that the response boundaries corresponded to “short” vs. “long” responses rather than accurate vs. inaccurate responses; this is the recommended approach when a response bias is expected, as is the case with the PRT. All four HDDM parameters were allowed to vary by group during model-fitting. We drew 10,000 samples from the posterior distribution, discarding the first 1,000 “burn-in” samples (Kruschke, 2014). The model was constructed such that individual parameter estimates were constrained by group-level prior distributions. Trace and autocorrelation plots were inspected to assess convergence. To evaluate model quality, the estimated parameters were used to generate simulated data (posterior predictive checks). Summary statistics from the actual data fell well within 95% intervals of the simulated data, indicating good fit.

Analyses were computed in Jupyter Notebooks (Kluyver et al., 2016) and focused on drift rate (v) and starting point (z). These parameters were of greatest interest because they capture the efficiency of decision-making vs. the impact of the asymmetric reinforcement rates on behavior, respectively. Additional HDDM parameters incorporated into the model included non-decision time (t; time needed for stimulus perception and response execution) and threshold (a; decision boundary between two choices). Consistent with prior studies (e.g., Lawlor et al., 2020), differences between groups were determined by inspecting each parameter’s posterior distribution and identifying areas with < 5% overlap between groups. Because these compare Bayesian posterior distributions, we refer to the HDDM outcomes as q-values rather than p-values.

Results

Group Characteristics

The distribution of time delay data resulted in a bimodal distribution, whereby subjects either smoked immediately at the start of the RAT (0-minute waiters, n = 36), and subjects that abstained from smoking for the entirety of the RAT (50-minute waiters, n = 44). Demographic and smoking-related characteristics revealed no statistically significant difference between groups (see Table 1). Withdrawal symptoms were assessed using the Minnesota Nicotine Withdrawal Scale (Hughes & Hatsukami, 1986) at both the abstinent and non-abstinent sessions. At the abstinent session, 0-minute waiters (M = 2.144, SD = 1.227) experienced more withdrawal symptoms relative to 50-minute waiters (M = 1.498, SD = 1.023); t₇₈ = 2.567, p = 0.012. There was no difference in withdrawal symptoms at the non-abstinent session (p = 0.234).

Table 1.

Demographic and smoking characteristics

Variable	M (SD), [n]		Statistics
	0-minute	50-minute
Age	43.69 (10.59)	40.64 (10.88)	t = 1.27, p = 0.209
Gender
Female	[14]	[11]	X2 = 1.778, p = 0.228
Male	[22]	[33]
Race/ethnicity			X2 = 4.979, p = 0.547
Black	10	17
White	19	23
Multiracial	2	1
Other	2	1
Hispanic	3	2
Educationa
Less than high school	[5]	[4]	X2 = 4.11, p = 0.250
High school or GED	[13]	[9]
Some college	[10]	[19]
College or higher	[6]	[11]
Incomeb			X2 = 6.439, p = 0.376
Less than $15,000	20	17
At least $15,000 but less than $30,000	7	8
At least $30,000 but less than $45,000	4	3
At least $45,000 but less than $60,000	0	2
At least $60,000 but less than $75,000	0	2
At least $90,000 but less than $105,000	0	1
Greater than $120,000	1	0
FTNDc	5.83 (1.94)	4.93 (2.33)	t = 1.86, p = 0.067
Cigarettes per dayd	17.79 (5.44)	16.50 (7.52)	t = 0.85, p = 0.398
Years of smoking	24.03 (11.52)	20.91 (11.09)	t = 1.23, p = 0.222
Number of cigarettes before PRT visit	5.28 (4.12)	4.32 (3.08)	t = 1.19, p = 0.237

^aN = 3 missing data

^bN = 15 missing data; only income categories for which there are data for at least one group are reported

^cFTND: Fagerstrom Test for Nicotine Dependence

^dN = 1 missing data

Group Differences in PRT Measures

A Group x Block ANOVA on response bias revealed a main effect of Block (F_2,156 = 7.802, p < 0.001, ɳ² = 0.091), whereby Block 3 elicited significantly higher response biases relative to Block 1 (p < 0.001) and Block 2 (p = 0.011; see Figure 2a). There was no main effect of Group (p = 0.550) nor a Block x Group interaction (p = 0.574). When controlling for overall discriminability, the Block x Group interaction remained non-significant (F_2,154 = 0.368, p = 0.686).

Figure 2. Response bias, discriminability, and RT data.

(A) Response bias was larger in block 3 vs. blocks 1 and 2 in both groups; (B) Discriminability was lower in block 1 vs. blocks 2 and 3, and was higher in the 50-minute group; (C) Across all blocks, RTs were faster during the presentation of the rich vs. lean stimulus; * p = 0.11, ** p = 0.002, *** p < 0.001

A 2-way ANOVA on discriminability revealed main effects of Block (F_2,156 = 8.440, p < 0.001, ɳ² = 0.098) and Group (F_2,78 = 10.511, p = 0.002, ɳ² = 0.119), but no interaction (p = 0.859). Post-hoc tests indicated that discriminability was lower in Block 1 vs. Block 2 (p = 0.002) and Block 3 (p < 0.001; see Figure 2b). The Group effect was driven by higher discriminability in 50-minute waiters (M = 0.662, SD = 0.252) vs. 0-minute waiters (M = 0.478, SD = 0.252; p = 0.002).

A 3-way ANOVA on RT revealed main effects of Block (F_2,156 =18.675, p < 0.001, ɳ² = 0.193), due to faster responses in Blocks 2 and 3 vs. Block 1 (p’s < 0.001). There was also a mean effect of Stimulus (F_2,156 = 58.694, p < 0.001, ɳ² = 0.431), as all subjects responded faster to rich compared to lean stimuli (p < 0.001), but there was no main effect of Group (p = 0.475). Moreover, there was a Block x Stimulus interaction (F_2,156 =6.783, p = 0.001, ɳ² = 0.80). For rich and lean stimuli, subjects responded faster in Block 3 vs. Block 1 (rich: p < 0.001; lean p = 0.001) and Block 2 vs. Block 1 (rich: p < 0.001; lean p < 0.001), but there was no difference between Block 2 and 3 (rich: p = 0.421; lean: p = 0.553; see Figure 2c). The between-stimulus difference within each block was also significant across all comparisons (all p’s <0.001) with responses to rich vs lean stimuli relatively small in block 1 (mean difference = 38.123), and similar in blocks 2 (mean difference = 69.403) and 3 (mean difference = 71.673). All other interactions were non-significant (all p’s > 0.05).

50-minute waiters (M = 117.522, SD = 2.663) accumulated more rewards across all blocks of the PRT relative to 0-minute waiters (M = 116.972, SD = 2.602), but this difference was not statistically significant (t₇₈ = 0.929, p = 0.356). Correlations between cumulative reward totals and signal detection metrics indicated that higher discriminability was associated with more rewards (r = 0.512, p <0.001, see Supplement Fig 1b), while response bias and cumulative reward totals were unrelated (r = −0.037, p = 0.746, see Supplement Fig 1a). Moreover, there was a significant difference between these two correlations (Z = 3.571, p < 0.001).

HDDM Parameters

The HDDM analyses revealed faster drift rates (v) for 50-minute waiters (Figure 2a) relative to 0-minute waiters, with little overlap between the posterior distributions (q < 0.001). By contrast, the posterior distributions for starting point bias (z), non-decision time (t), and threshold (a) showed considerable overlap between groups (q = 0.481, q= 0.227, q = 0.951 respectively, see Figure 3b–d). This result was replicated when drift rate was estimated separately for the rich and lean stimuli, such that 50-minute waiters demonstrated faster drift rates than 0-minute waiters for both the rich (q = 0.002) and lean (q = 0.002) conditions (see Supplement Fig. 2).

Figure 3. Posterior distributions of for HDDM parameters extracted from the PRT.

Posterior distributions for (A) drift rate, (B) starting point bias, (C) non-decision time, and (D) threshold. *** < 0.001 percent overlap between the two groups.

Relationship between HDDM parameters and PRT measures

Multiple linear regression of HDDM parameters on response bias indicated that the overall regression was statistically significant (R² = 0.415, F_4,75, = 13.289, p < 0.001), whereby starting point bias predicted response bias (β = 0.639, p < 0.001, see Supplement Fig 1c), while other HDDM parameters did not (all p’s > 0.05). The overall regression for HDDM parameters on discriminability was also significant (R² = 0.950, F_4,75, = 356.221, p < 0.001), whereby drift rate (β = 1.092, p < 0.001) and, to a lesser extent, threshold (β=0.339, p < 0.001) and starting bias (β = −0.071, p = 0.015) predicted discriminability, while non-decision time did not (β = −0.006, p = 0.817).

Linear regression of cumulative reward total on group was not significant (R² = 0.011, F_1,78, = 0.864, p =0.356); however, including the individual HDDM parameters into the regression (R² = 0.330, F_5,74, = 7.306, p < 0.001) indicated that drift rate significantly predicted cumulative reward (β = 0.568, p < 0.001, see Supplement Fig 1d), while no other HDDM parameters were predictive of cumulative reward (all p’s > 0.05).

Discussion

Computational metrics of decision-making can disentangle underlying processes that are not directly observable through basic behavioral measures alone (Copeland et al., 2023; Mandali, Sethi, Cercignani, Harrison, & Voon, 2021; Mandali, Weidacker, Kim, & Voon, 2019; Ratcliff, 1978; Ratcliff & McKoon, 2008). In this study, we used the HDDM to tease apart reward sensitivity from decision processes in order to determine how these relate to the choice to smoke during a period of abstinence. People who abstained from smoking for the duration of the RAT demonstrated faster evidence accumulation (i.e., higher drift rates) for choices in the PRT relative to those who smoked immediately. There were, however, no group differences in the other HDDM parameters. Moreover, the groups differed in discriminability but not response bias or RT (although the presence of a strong response bias in both groups indicated that the task elicited the expected learning of reward contingencies), while greater discriminability was related to higher drift rates clarifying the relationships between signal detection metrics and underlying HDDM parameters. Together these data indicated that decision-making accuracy, as assessed by discriminability and supported by drift rate, differed reliably between people who chose (or not) to smoke, whereas there were no group differences in reward responsivity (as assessed by response bias and starting point bias).

As expected, signal-detection metrics of behavior indicated that all subjects learnt the contingencies of the task, confirming that the PRT effectively elicited preferential responding towards the more rewarded stimuli. This is evident in the fact that there was a significant response bias towards the rich stimulus in both groups, and in the fact that RTs were faster in response to rich vs. lean stimuli. Given the reinforcement schedule of rich compared to lean stimuli, it follows that subjects would develop a preference for the more highly rewarded stimulus (Macmillan & Creelman, 2004; McCarthy, 1991; Pizzagalli et al., 2005). It was thus plausible that people who chose to smoke more quickly on the RAT might show weaker reward responsiveness as assessed by these variables (response bias, RT), but this was not the case. There were no group differences in response bias or RT, and no group difference emerged for the starting bias parameter from the HDDM, which captures the preferential tendency to respond “rich” in this task. Together these results indicate that the PRT drove the expected response bias in all participants—this effect simply did not differ between the two groups.

Instead, group differences emerged in discriminability and drift rate, suggesting that computational modeling can tease apart underlying variability in decision-making that maps on to smoking choice behavior during abstinence. Drift rate indicates the speed with which evidence about the stimuli is accumulated, in order to cross one of the decision boundaries (Wiecki et al., 2013). Here we found that people who were able to abstain from smoking for a short period of time more efficiently accumulated evidence in the PRT, relative to people who were unable to abstain. Moreover, drift rate was strongly related to discriminability, which also differed between the two groups, and these two variables—discriminability and drift rate—were highly (positively) predictive of cumulative reward totals. Discriminability was also predicted by threshold (albeit to a lesser extent), together indicating that HDDM captured discriminability by combining both fast evidence accumulation and wider separation between threshold to discern stimuli. Overall, this set of findings indicates that, relative to individuals who chose to smoke immediately, individuals who were able to abstain from smoking also made more accurate (and more frequently rewarded) decisions in the PRT, with this behavioral effect driven by underlying differences in the efficiency of decision-making (as indexed by drift rate).

These findings have clinical relevance given that group differences not only affected choice behavior during abstinence (during RAT performance), but were apparent even after subjects had smoked (because the PRT data were collected when all participants had smoked). Thus, these results suggest that there may be individual differences in cognitive function that are relevant to abstinence and that can be defined before someone attempts to quit; it may be necessary to use computational models, such as the HDDM, to identify these differences as they may not be captured by standard metrics of behavior (like response bias). The findings also speak to the ecological validity and clinical utility of this design, whereby differences between groups on the PRT were evident in advance of the RAT – suggesting that abstinence sessions and reward sessions need not be conducted concurrently for meaningful group and individual differences to emerge.

The neurobiological underpinnings for faster evidence accumulation in one group relative to another, while not directly measured in this study, may rely on individual variability in neuronal signaling. It is well known that dopamine relates to reward (Schultz, 2010) and its function is implicated in substance use disorders (Volkow & Morales, 2015). A recent meta-analysis reported that in people with various addictions, lower dopamine function (as measured by PET) was related to a higher propensity to devalue delayed rewards (Castrellon et al., 2019). Variability in dopaminergic function may therefore explain the natural separation into 0- and 50-minute waiters on the RAT – where lower dopaminergic function may contribute to an inability to delay cigarette reward in the former. Pre-clinical (Picciotto, 2003; Wittenberg, Wolfman, De Biasi, & Dani, 2020) and clinical work (Brody et al., 2004; Takahashi et al., 2008) also demonstrates that nicotine use stimulates dopamine release, and thus the choice to smoke immediately for the 0-minute waiters during the RAT may reflect a drive to “normalize” the dopamine deficit that may be observed among this group. Indeed, 0-minute waiters demonstrated higher withdrawal symptoms at the abstinence session (i.e., when the RAT was performed), which may have contributed to their choice to smoke immediately in this task.

Variance in dopaminergic function may also have contributed to the observed group difference in the speed of evidence accumulation. The dopamine system has been implicated in the evidence accumulation process in prior work (Beste et al., 2018), and dopamine broadly facilitates behaviors aimed at attaining rewards (Goto & Grace, 2008; Pessiglione, Seymour, Flandin, Dolan, & Frith, 2006). Studies that have examined evidence accumulation in groups with known dopamine dysfunction (e.g., schizophrenia, obsessive-compulsive disorder, ADHD, depression; Banca et al., 2015; Fosco, White, & Hawk, 2017; Lawlor et al., 2020; Moustafa et al., 2015) have generally found slowed drift rates, suggesting that slow evidence accumulation may be a transdiagnostic marker of—or risk factor for—psychopathology (Sripada & Weigard, 2021). Given that chronic nicotine use can decrease baseline dopaminergic function (Perez, Ly, McIntosh, & Quik, 2012), and that all subjects in the current study were chronic smokers (i.e., regularly smoked for an average for 22 years), it is plausible that group differences in evidence accumulation may be related to variability in baseline dopaminergic function.

While the current study provides compelling evidence for differences in reward-based decision-making between two groups of smokers, it is not without limitations. First, individuals included in our sample were not required to be treatment seeking, and, as a result, likely differed in their baseline levels of motivation to quit smoking. However, the small monetary reward offered in the RAT clearly separated the behavior of groups who were and were not willing to abstain, which mapped onto group differences in computational measures previously measured during satiety. Second, given that we did we did not measure neurobiological correlates of PRT behavior, links to dopaminergic variability are speculative. To link these behavioral findings with their neurobiological underpinnings, future studies using neuroimaging techniques are necessary. Third, higher IQ has also been tied to higher drift rate (Ratcliff, Thapar, & McKoon, 2010), which was not measured in this sample. However, education level - which is highly correlated with IQ (Ritchie & Tucker-Drob, 2018) - did not differ between groups, suggesting there would likely be no relationship between IQ and drift rate in our sample.

To note, some research has shown that increasing the value of an alternative (i.e., monetary) reinforcer may reduce the likelihood of choosing to smoke (Bisaga, Padilla, Garawi, Sullivan, & Haney, 2007; Cassidy, Tidey, Kahler, Wray, & Colby, 2015; Stoops, Poole, Vansickel, & Rush, 2011). However, even if the value of the monetary reinforcement on the RAT was increased (or even decreased), leading to reduced variability in groupings, computational parameters would likely remain unchanged as the PRT was completed when all participants were satiated. Instead, the relationship between abstinence and drift rate would be obscured, as there would be no way to differentiate PRT results if groupings disappeared. Thus, computational parameters can provide a clearer understanding of variability in reward processing even during satiety in smokers. Indeed, DDM has been shown to discriminate tobacco-related choice in current and ex-smokers (Copeland et al., 2023). In a clinical setting, this suggests we can use DDM parameters to track how mechanisms underlying tobacco-related value and choice change from treatment initiation, through relapse, and to successful abstinence. This may lead to more targeted treatment based modulating specific parameters that contribute to successful cessation.

Here we demonstrate that computational modeling can provide insight into choice behavior in people who smoke cigarettes. Specifically, our results indicate that individuals who were willing to abstain from smoking in exchange for financial compensation more efficiently accumulated evidence (as measured by drift rate) for reward-based decisions, during satiety, relative to those who did not abstain. Slow evidence accumulation has been shown to cut across psychiatric illnesses, suggesting it can be used as a transdiagnostic marker for vulnerability to psychopathology (Sripada & Weigard, 2021). Our results thus suggest cognitive factors tied to decision-making may relate to an individual’s ability to remain abstinent from smoking in exchange for non-drug rewards, and highlight the utility of using computational modeling to understand the mechanisms that underlie decision-making in people who smoke.

Highlights.

• Computational models clarify reward-based decisions

• Evidence accumulation differentiates those who can and cannot forgo nicotine use

• Variance in decision making during satiety is related with the choice to smoke