Substance abusers often are impaired on laboratory measures of decision-making (Bechara et al., 2001
; Petry, 2003
; Petry, Bickel, & Arnett, 1998
; Rogers et al., 1999
). For example, in a laboratory decision-making task known as the Iowa Gambling Task (IGT; Bechara, Damasio, Damasio, & Anderson, 1994
), substance abusers often make choices that lead to small, immediate gains at the cost of larger losses over time (S. Grant, Contoreggi, & London, 2000
). Cannabis (marijuana) users, like other substance-using populations, perform more poorly than non-using controls on the IGT (Lamers, Bechara, Rizzo, & Ramaekers, 2006
; Whitlow et al., 2004
), even after prolonged abstinence from the drug (Bolla, Eldreth, Matochik, & Cadet, 2005
). This impairment may be due to underlying deficits or differences in psychological processes (e.g., memory impairments, loss insensitivity, etc.), but pinpointing such processes can be difficult with traditional behavioral measures from the IGT. Recent work has attempted to disentangle component processes of the IGT by means of computational cognitive models (Busemeyer & Stout, 2002
; Garavan & Stout, 2005
; Yechiam, Busemeyer, Stout, & Bechara, 2005
). In this report, we use mathematical models of choice behavior on the IGT to better understand the risk taking behavior of cannabis users. We present a comparison of two such models, and then compare estimated model parameters of chronic cannabis users and controls to identify the particular psychological processes which may be impaired in cannabis users.
For the IGT, the participant must make a series of choices from four decks of cards with the goal of maximizing his or her net payoff across trials. On each trial, the participant selects a card from any of the four decks and is informed how much (s)he won or lost by choosing that card. Every choice leads to a gain that sometimes is coupled with a simultaneous loss (see ). Selecting from the two “disadvantageous” decks will result in a larger per-selection gain, but on average leads to a net loss over ten selections, whereas selecting from the two “advantageous” decks results in a smaller per-selection gain but an overall net gain over ten selections. To perform well on the IGT the participant learns to select primarily from advantageous decks on the basis of the net gains and losses they experience across the task. Thus, the IGT incorporates cognitive (i.e., learning and memory) and motivational processes (i.e., responsivity to gains and losses) associated with the anticipation of outcomes following choices over time. The decision to use or abstain from also drugs invokes processes related to learning from previous experiences with the drug, and the perceived rewards (i.e., pleasure) and punishments (i.e., financial, interpersonal, legal trouble) associated with drug use.
Payoffs used in the Iowa Gambling Task
Computational cognitive models allow us to disentangle the processes contributing to IGT performance and to identify specifically those processes which may account for the poorer overall performance of an individual or group on the task (Busemeyer & Stout, 2002
). Our research group has developed a mathematical model called the Expectancy Valence Learning (EVL) model (Busemeyer & Stout, 2002
) to investigate the psychological processes underlying individuals’ decisions on the IGT. The model has three assumptions. First, a utility function represents an individual’s subjective evaluation of gains and losses. Second, a learning rule allows the development of expectancies for each deck that are updated on the basis of experienced utilities. Third, these expectancies determine the probabilities that the participant will choose a given deck on each trial via a choice rule. The EVL model is based on principles derived from the judgment and decision-making literature and yields theoretically-derived dependent measures (model parameters) that describe psychological processes underlying IGT performance. These parameters reflect the degree to which the decision maker attends to gains versus losses (Attention to Gains parameter), his or her learning rate (Recency parameter), and the degree of consistency between deck selections and the expected outcomes associated with each deck (Consistency parameter). By applying the model to several datasets from clinical populations who demonstrate impaired IGT performance, we have identified distinctive patterns within the empirical data which differentiate various groups of drug abusers, subjects with Huntington’s disease, and subjects with orbitofrontal brain lesions from their respective control groups (for a review, see Yechiam et al., 2005
Using the EVL model of IGT performance, our group has shown previously that disruptions in psychological processes may underlie the poorer performance of cocaine abusers (Stout, Busemeyer, Lin, Grant, & Bonson, 2004
) and polysubstance abusers (Stout, Rock, Campbell, Busemeyer, & Finn, 2005
) on the IGT. With regard to cannabis users specifically, a recent analysis of decision processes in a sample of 21 young (mean age = 24 years) cannabis-using college students found no significant differences between that group and non-using controls on any EVL model parameters (Bishara et al., in press). In addition, a review of the EVL modeling of the IGT performance in various clinical samples included a brief summary of 25 chronic cannabis abusers who differed from controls on the Recency and Attention to Gains parameters (Yechiam et al., 2005
). This report includes the 17 chronic cannabis abusers from that report who had been abstinent only long enough for the acute effects of the drug to have worn off (i.e., they were no longer intoxicated, or high) but before they would have started having withdrawal symptoms, and extends this previous work principally by allowing an evaluation of a new model that may have better explanatory ability for IGT behavior.
Investigations of cognition among chronic cannabis abusers have identified disruptions in psychological processes which could contribute to their poorer IGT performance. For instance, chronic users are impaired relative to non-users on neuropsychological measures of memory and learning (for reviews, see I. Grant, Gonzalez, Carey, Natarajan, & Wolfson, 2003
; Solowij & Battisti, 2008
). This impairment could compromise their ability to maintain and update representations of the expectancy for each deck across IGT trials. The effects of chronic cannabis use on sensitivity to reward and punishment are less clear, although acute administration studies have shown that cannabis exposure is associated with increased risk-taking and decreased sensitivity to choice outcomes (Lane & Cherek, 2002
; Lane, Cherek, Tcheremissine, Lieving, & Pietras, 2005
). These results are supported by a recent functional magnetic resonance imaging (fMRI) study which showed that chronic cannabis users exhibit patterns of neural activity consistent with hypersensitivity during reward anticipation and hyposensitivity to loss outcomes (Nestor, Hester, & Garavan, in press
). Chronic users may show a similar pattern of behavior on the IGT, manifested as a bias toward the disadvantageous decks. Lastly, chronic cannabis users score highly on personality measures related to risk-seeking, which may lead them to make impulsive selections that are inconsistent with their expectancies regarding deck outcomes (Satinder & Black, 1984
We recently developed the Prospect Valence Learning (PVL) model, which is a modification of the EVL model (Ahn, Busemeyer, Wagenmakers, & Stout, 2008
. The PVL model employed in this report uses the same learning rule as the EVL model, but uses a different utility function and a different choice rule. Ahn et al. (2008)
showed that the PVL model resulted in better post-hoc fits, simulation performance, and generalizability than comparison models when applied to IGT data from healthy normal subjects. There are two main purposes of this report. The primary purpose is to compare the new PVL model to the EVL model using both a clinical population and a control population for the first time. This is an important step for two reasons: first, we need to determine whether the superior predictive power of the PVL model over the EVL models continues to hold for clinical populations; second, we need to examine if the parameters estimated from the PVL model are more or less informative than the parameters estimated from the EVL model with respect to revealing important differences between clinical and control populations.
The equations for the EVL and PVL models are shown in . The models explain choices in the IGT in slightly different ways. First consider the concept ‘valuing a card’ shown in . As each card is selected in the IGT, the decision maker assesses the value of that card. The decision maker’s valuation of a card will vary depending on the relative amount of attention the (s)he pays to gains versus losses. Some individuals will only register gains, others will only register losses, and still others will attend to both wins and losses, with the weighting of attention varying across decision-makers. The Attention to Gains parameter (w;
see ) in the EVL model captures the relative amount of attention a decision maker pays to gains compared to losses on a given trial. If w
= 0 all attention is paid to losses, whereas if w
= 1 all attention is paid to gains. Based on the level of attention to gains, the decision maker generates a value for that card. In the PVL model, the subjective utilities are represented by a non-linear prospect utility function. The shape parameter (α) governs the curvature of the utility function (0 < α < 1: as α approaches 1, subjective utility increases in direct proportion to the outcome value; as α approaches 0, subjective utility increases in a stepwise fashion so all gains are subjectively equal and all losses are subjectively equal). The utility function of the PVL model also contains a loss-aversion parameter (λ) which indicates the subject’s sensitivity to losses compared to gains (0 < λ < 10: as λ approaches 0 losses are experienced as neutral events with utility = 0; for λ = 1 losses and gains have the same impact; for λ > 1 losses have greater impact than gains on the subjective utility of an outcome, leading to loss aversion). The advantage of using the PVL’s non-linear utility function is that it accounts for the gain/loss frequency effect. That is, winning $100 five times is often perceived as better than winning $500 once, even though the net gain is equivalent (Erev & Barron, 2005
). The EVL’s linear utility function assumes that both of these events have the same overall utility. Therefore, the PVL model explains participants’ preferences for decks with low net-loss frequency (e.g., Deck B) over decks with high net-loss frequency (e.g., Deck A) even if their expected values are the same (Ahn et al., 2008
EVL and PVL model equations for estimating parameters. Model-fitting selects parameter values that maximize the likelihood of the decision maker’s responses, given the model.
Next consider the concept ‘creating an expectancy’ shown in . With the experience of each card’s payoff, the decision maker can then revise the expectancy about the deck from which the card was chosen. Each time a new card is drawn, the old deck expectancy is updated based on the value of the new card. The Recency parameter (A
) is a parameter of the delta learning rule (Rescorla & Wagner, 1972
) for both the EVL and PVL models. The Recency parameter (0 < A
< 1) is an index of learning rate, indicating how much weight is given to past experiences with a given deck versus how much weight is placed on the value of the most recent selection from that deck. A high Recency parameter indicates that the value of the most recent card selection has a large influence on the expectancy for that deck, and forgetting of previous card selections is rapid. In contrast, a low Recency parameter indicates that the value of the most recent card selection has a small influence on the expectancy for that deck, and forgetting is more gradual. In this way, expectancies about each deck develop as each new card is selected.
The third concept in is the ‘probability of choosing a deck.’ In order to select a deck on each trial, the decision maker compares the current expectancies for each deck. A good decision maker makes choices consistent with his or her deck expectancies as the trials progress and as confidence in the expectancies increases with experience. The Consistency parameter (c
) is an indicator of the fidelity between the decision maker’s selections and expectancies as the task progresses. A high value indicates that the decision maker’s choices are deterministic, resulting in maximal choices from the deck with the highest expectancy. A low Consistency value indicates that the decision maker chooses more randomly, possibly reflecting impulsivity or boredom with the task. The EVL model uses a trial-dependent choice rule in which the consistency increases or decreases over trials (−5 < c
< 5; Busemeyer & Stout, 2002
). In contrast, the PVL model uses a trial-independent choice rule in which consistency remains constant over trials (0 < c
In summary, for the current study, we applied the EVL and PVL models to the IGT performance data obtained from 17 chronic, heavy cannabis users and 15 control subjects who had only minimal lifetime exposure to cannabis. Empirical results from ten of the subjects in this sample were included in a previous report (Whitlow et al, 2004
), which revealed poor gambling task performance in the cannabis group as compared to the control group. We replicate this finding in an enlarged sample of chronic cannabis users. We then report a comparison of the ability of the EVL and PVL models to account for each group’s performance on the IGT. Finally, we present an analysis of the individual differences in psychological processes underlying cannabis users’ poor performance on the IGT.