Examining the Neurochemical Underpinnings of Animal Models of Risky Choice: Methodological and Analytic Considerations

Abstract

Because risky choice is associated with several psychiatric conditions, recent research has focused on examining the underlying neurochemical processes that control risk-based decision making. Not surprisingly, several tasks have been developed to study the neural mechanisms involved in risky choice. The current review will briefly discuss the major tasks used to measure risky choice and will summarize the contribution of several major neurotransmitter systems to this behavior. To date, the most common measures of risky choice are the probability discounting task (PDT), the risky decision task (RDT), and the rat gambling task (rGT). Across these three tasks, the contribution of the dopaminergic system has been most studied, although the effects of serotonergic, adrenergic, cholinergic, and glutamatergic ligands will be discussed. Drug effects across these tasks have been inconsistent, which makes determining the precise role of neurotransmitter systems in risky choice somewhat difficult. Furthermore, procedural differences can modulate drug effects in these procedures, and the way data are analyzed can alter the interpretations one makes concerning pharmacological manipulations. By taking these methodological/analytic considerations into account, we may better elucidate the neurochemistry of risky decision making.

Risky choice primarily refers to choosing between certain and uncertain outcomes. For example, subjects can choose between a small, certain reward or a large, probabilistic reward. Alternatively, subjects can choose between a large reward that is sometimes associated with punishment and a small reward not associated with punishment. In the first case, consistently choosing the large reward when the probability of its delivery is low is considered to reflect risky choice. In the second case, choosing the large reward when the probability of receiving punishment is high also reflects risky decision making. Because risky choice is associated with several psychiatric conditions, including attention-deficit/hyperactivity disorder (see Dekkers, Pompa, Agelink van Rentergem, Bexkens, & Huizenga, 2016 for a recent meta-analysis), borderline personality disorder (Schuermann, Kathmann, Stiglmayr, Renneberg, & Endrass, 2011; Svaldi, Phillpsen, & Matthies, 2012), obsessive-compulsive disorder (Grassi et al., 2015), pathological gambling (Brand et al., 2005; Madden, Petry, & Johnson, 2009), and substance use disorders (Brevers et al., 2014; Schutter, van Bokhoven, Vanderschuren, Lochman, & Matthys, 2011), understanding the underlying neurobiological processes of risky decision making has received considerable attention in recent years. Because Winstanley and Floresco (2016) already describe the neuroanatomic structures involved in such decision making, the goals of the current review are as follows: 1) provide a brief overview of the most commonly used behavioral tasks to measure risky choice in rodents; 2) describe the neurotransmitter systems that have been implicated in risky choice; and 3) discuss some of the methodological and analytic considerations that need to be taken into account when studying risk-based decision making.

I previously published a literature review (Yates, 2018) discussing how using different measures of impulsive choice and different statistical analyses can alter interpretations of drug effects in delay discounting procedures. Because some measures of impulsive choice (e.g., delay discounting) share some similarities to certain measures of risky choice (e.g., probability discounting), some of the topics raised in this review will be similar to those discussed in Yates (2018). However, this review is important for addressing some of the limitations that have been observed in behavioral pharmacology experiments assessing the neurochemical basis of risky choice. Before reviewing the literature, I want to note that the Institutional Animal Care and Use Committee (IACUC) at NKU (protocol number 2017–04; “Contribution of NMDA NR2B Subunit to Risky Choice and Amphetamine Reward”) has approved my research.

Preclinical Models of Risky Choice

The three most common preclinical models of risky choice will be discussed in this section: the probability discounting task (PDT), the risky decision task (RDT), and the rat gambling task (rGT). Please see Winstanley and Floresco (2016) for schematics illustrating how each of these procedures is designed.

Probability Discounting Task (PDT)

In the PDT, animals choose between a small, guaranteed reinforcer and a large, uncertain reinforcer. Although the probability of obtaining the large reinforcer can be manipulated across sessions (e.g., Mobini, Chiang, Ho, Bradshaw, & Szabadi, 2000), most studies decrease the probability within a session (e.g., Cardinal & Howes, 2005; St Onge & Floresco, 2009). Typically, sessions are composed of blocks of trials, in which subjects are exposed to forced-choice trials and free-choice trials. During forced-choice trials, only one manipulandum is presented to each subject. During free-choice trials, both manipulanda are available. In the PDT, subjects can maximize food intake by selecting the alternative with the higher expected utility. For example, the expected utility of one food pellet delivered with a probability of 1 is 1. The expected utility of four pellets delivered with a probability of 0.25 is 1 as well. If a 1:4 magnitude ratio is used, subjects should respond for the large magnitude reinforcer when the probability of its delivery is greater than 0.25. At a probability of 0.25, subjects should show indifference. At a probability less than 0.25, subjects should respond for the small magnitude reinforcer. The dependent variable is the percentage of trials in which the subject selects the large magnitude reinforcer during each block of trials. As the probability of obtaining the large magnitude reinforcer decreases across the session, choice for this alternative normally decreases. Subjects that consistently choose the large, probabilistic reinforcer are considered to exhibit greater risky choice relative to subjects that respond more for the small, certain alternative.

Risky Decision Task (RDT)

Instead of manipulating the probability of receiving the large magnitude reinforcer, the RDT alters the probability subjects receive shock (positive punishment) for choosing this alternative (e.g., Simon, Gilbert, Mayse, Bizon, & Setlow, 2009). Similar to the PDT, RDT sessions are composed of multiple blocks of trials, which consist of forced-choice and free-choice trials. Determining which alternative is optimal is somewhat difficult with the RDT. If the optimal alternative is the one that maximizes reinforcement, subjects should respond for the large magnitude reinforcer. Indeed, there is some evidence that subjects are not always sensitive to shock (see the Advantages and Disadvantages of Each Task section below). However, if the goal is to minimize exposure to foot shock, subjects should respond for the alternative associated with the small magnitude reinforcer. The dependent variable is identical to that observed in the PDT: the percentage of trials in which the large magnitude reinforcer is chosen during each block of trials. As the probability of receiving shock increases across the session, subjects are less willing to select this alternative. Here, subjects that consistently choose the alternative paired with positive punishment are labeled as high risk taking.

Rat Gambling Task (rGT)

The rGT is an animal analog of the Iowa Gambling Task (IGT; Bechara, Damasio, Damasio, & Anderson, 1994). In contrast to the PDT and the RDT, in which subjects choose between two alternatives, the rGT incorporates four alternatives that differ in reinforcer magnitude, probability of earning each reinforcer, and time-out duration (Zeeb, Robbins, & Winstanley, 2009). The four alternatives are labeled as P1, P2, P3, and P4. If a subject chooses the P1 alternative, it can receive either one food pellet (with a probability of 0.9) or a time-out period of 5 s (with a probability of 0.1). For the P2 option, there is a probability of 0.8 that the subject receives two pellets and a probability of 0.2 of receiving a 10-s time-out. The P3 option is associated with either delivery of three pellets (with a probability of 0.5) or a 30-s time-out (with a probability of 0.5). Finally, selecting the P4 option can lead to delivery of four pellets (with a probability of 0.4) or a 40-s time-out (with a probability of 0.6). In the rGT, selecting the P2 option is considered to be optimal, as subjects can earn the most food by consistently selecting this alternative (due to the high probability of receiving two pellets and the relatively short time-out periods). The dependent variable is typically the percentage choice for each alternative.

Advantages and Disadvantages of Each Task

One advantage of each task is that individual sessions can be completed within a relatively short amount of time (45 min for the PDT, 60 min for the RDT, 30 min for the rGT). Also, behavioral training can be completed within 30–40 sessions for each task. The PDT is perhaps the simplest task to use as it requires no special equipment (other than an operant conditioning chamber). Despite its simplicity, Yang, Cheng, and Liao (2018) argue that this task does not provide a true assessment of risk taking. Instead, this task merely provides a measure of decision making based on reward probability. According to Yang et al. (2018), tasks that measure risk taking should be designed such that the expected utility between the large magnitude and small magnitude reinforcers is equivalent. For example, instead of delivering one pellet guaranteed (expected utility of 1) or 4 pellets with a probability between 12.5–100% (expected utility of 0.5–1), subjects can have a choice between a one-pellet reinforcer delivered with a probability of 1 and a 4-pellet reinforcer delivered with a probability of 0.25. This would result in a situation in which the expected utility is equal across both conditions.

In contrast to the PDT and rGT, the RDT measures the contribution of positive punishment to risky choice as opposed to negative punishment. Therefore, this task may be beneficial for modeling behaviors such as substance abuse (e.g., individuals that share needles risk contracting a sexually transmitted disease). Although RDT sessions can be completed within a traditional operant chamber like the PDT, special equipment is needed (shock generators). These generators can be fairly expensive (e.g., the shock generators I just purchased for my laboratory are approximately $1,000 each). Another potential limitation is that subjects may not always be sensitive to shock. For example, Orsini, Willis, Gilbert, Bizon, and Setlow (2016) compared male and female rats in the RDT and found that male rats showed near exclusive preference for the large magnitude reinforcer even when shocks were delivered after each food delivery (greater than 80% responding for this alternative). It is important to note that the insensitivity observed in males is most likely due to the fact that Orsini et al. (2016) had to lower the shock intensity to prevent a complete suppression in behavior for female subjects.

Because the rGT is an animal analog of the IGT, some have argued that this task is beneficial due to the fact that it provides a more realistic account of risk-based decision making (Zeeb et al, 2009). Another advantage of this procedure is that it can simultaneous measure motor impulsivity in the form of premature responses and preservative responses. One drawback to this task is that a special apparatus is typically used in the rGT; specifically, the rGT is measured in a chamber with four nosepoke apertures (although, this task could be potentially conducted in a standard operant conditioning chamber containing four manipulanda). Another drawback to this task is that interpreting drug effects can be difficult. For example, Zeeb, Wong, and Winstanley (2013) observed an increase in the P1 option following administration of the dopamine D₂ receptor antagonist eticlopride in rats that live in an isolated condition, but this drug failed to alter choice for the other alternatives. Because the P2 option is considered to be optimal, does an increase in P1 responding without a change in P2 responding measure a change in risky choice?

Neurochemical Underpinnings of Risky Choice

Because studies assessing the neurochemical basis of risky choice have focused primarily on the dopaminergic system, this section will discuss the results of dopaminergic ligands in detail. Table 1 summarizes the results of studies assessing the contribution of the dopaminergic system to risky choice. Table 1 also lists the characteristics of the task, and the results of each pharmacological manipulation. Although the dopaminergic system has received considerable attention, the contribution of other neurotransmitter systems will be described as well.

Table 1.

Effects of Amphetamine (AMPH) and Dopamine-Selective Ligands on Performance in Each Major Risky Choice Task

Task	Probability presentation order	Reinforcement signaled?	Drug	Mechanism of action	Dose	Outcome	Citation
PDT	Probabilities decrease	No	AMPH	DAT/NET reverser	0.125–1.0 mg/kg	↑risky choice (averaged across dose)	St Onge & Floresco (2009)
PDT	Probabilities decrease	No	AMPH	DAT/NET reverser	0.125–0.5 mg/kg	↑risky choice (each dose)a	Floresco & Whelan (2009)
PDT	Probabilities decrease	No	AMPH	DAT/NET reverser	0.5 mg/kg	↑risky choice	St Onge et al. (2010)
PDT	Probabilities increase	No	AMPH	DAT/NET reverser	0.5 mg/kg	↑risky choice	St Onge et al. (2010)
PDT	Probabilities pseudorandomized	No	AMPH	DAT/NET reverser	0.5 mg/kg	No effect	St Onge et al. (2010)
PDT	Probabilities decreaseb	Yes	AMPH	DAT/NET reverser	0.1–1.0 mg/kg	↑risky choice (0.3 & 1.0)	Oinio et al. (2017)
RDT	Probabilities increase	No	AMPH	DAT/NET reverser	0.33–1.5 mg/kg	↓risky choice (1.5)	Simon et al. (2009)
RDT	Probabilities increase	No	AMPH	DAT/NET reverser	1.5 mg/kg	↓risky choice	Simon et al. (2011)
RDT	Probabilities increase	No	AMPH	DAT/NET reverser	0.3–1.5 mg/kg	↓risky choice (1.0c & 1.5d)	Orsini et al. (2016)
RDT	Probabilities increase	No	AMPH	DAT/NET reverser	2.0–20.0 μg in mPFC	No effect	Orsini et al. (2018)
RDT	Probabilities decrease	No	AMPH	DAT/NET reverser	2.0–20.0 μg in mPFC	Impairs discriminability of reinforcer magnitudes (10.0)	Orsini et al. (2018)
rGT	N/A	No	AMPH	DAT/NET reverser	0.3 mg/kg	No effect	Zeeb et al. (2009)
rGT	N/A	No	AMPH	DAT/NET reverser	1.0 mg/kg	↑Pl; ↓P2	Zeeb et al. (2009)
rGT	N/A	No	AMPH	DAT/NET reverser	1.5 mg/kg	↑Pl/P4; ↓P2	Zeeb et al. (2009)
rGT	N/A	No	AMPH	DAT/NET reverser	0.3 mg/kg	↓P2e; ↑P3f	Zeeb et al. (2013)
rGT	N/A	No	AMPH	DAT/NET reverser	1.0 mg/kg	↑Ple; ↓P2g	Zeeb et al. (2013)
rGT	N/A	No	AMPH	DAT/NET reverser	1.5 mg/kg	↑Pl/P4e; ↓P2e; ↑P3f	Zeeb et al. (2013)
rGT	N/A	No	AMPH	DAT/NET reverser	1.0 mg/kg	↑Pl; ↓P2	Silveira et al. (2015)
rGT	N/A	No	AMPH	DAT/NET reverser	0.5–5 mg/kg	↑Pl; ↓P2	Barrus & Winstanley (2016)
rGT	N/A	Yes	AMPH	DAT/NET reverser	0.5–5 mg/kg	No effect	Barrus & Winstanley (2016)
PDT	Probabilities decrease	No	SKF 81297	D1-like agonist	0.1 & 0.3 mg/kg	↑risky choice (0.3)	St Onge & Floresco (2009)
PDT	Probabilities decrease	No	SKF 81297	D1-like agonist	1.0 mg/kg	↓risky choice @ 50% but ↑risky choice (0.01) @ 25%	St Onge & Floresco (2009)
PDT	Probabilities decrease	No	SKF 81297	D1-like agonist	0.1 & 0.4 μg in mPFC	No effect	St Onge et al. (2011)
PDT	Probabilities decreasel	No	SKF 81297	D1-like agonist	0.1–0.5 mg/kg	No effect	Oinio et al. (2017)
PDT	Probabilities decrease	Yes	SKF 81297	D1-like agonist	0.1 & 0.5 mg/kg	No effect	Wallin-Miller et al. (2018)
PDT	Probabilities decrease	No	SKF 38393	D1-like agonist	1.0–10.0 mg/kg	No effect	Smith et al. (2018)
PDT	Probabilities decrease	Yes	SKF 38393	D1-like agonist	1.0–10.0 mg/kg	No effect	Smith et al. (2018)
RDT	Probabilities increase	No	SKF 81297	D1-like agonist	0.1–1.0 mg/kg	No effect	Simon et al. (2011)
rGT	N/A	No	SKF 81297	D1-like agonist	0.03–0.3 mg/kg	No effect	Zeeb et al. (2009)
PDT	Probabilities decrease	No	SCH 23390	D1-like antagonist	0.005 mg/kg	↓risky choice	St Onge & Floresco (2009)
PDT	Probabilities decrease	No	SCH 23390	D1-like antagonist	0.01 mg/kg	Impairs discriminability of reinforcer magnitudes	St Onge & Floresco (2009)
PDT	Probabilities decrease	No	SCH 23390	D1-like antagonist	0.1 & 1.0 μg in mPFC	↓risky choice (1.0)	St Onge et al. (2011)
PDT	Probabilities decrease	No	SCH 23390	D1-like antagonist	0.003 & 0.01 mg/kg	↓risky choice	Smith et al. (2018)
PDT	Probabilities decrease	Yes	SCH 23390	D1-like antagonist	0.003 & 0.01 mg/kg	No effect	Smith et al. (2018)
RDT	Probabilities increase	No	SCH 23390	D1-like antagonist	0.005–0.03 mg/kg	No effect	Simon et al. (2011)
rGT	N/A	No	SCH 23390	D1-like antagonist	0.001–0.01 mg/kg	No effect	Zeeb et al. (2009)
rGT	N/A	No	SCH 23390	D1-like antagonist	0.001–0.01 mg/kg	No effect	Zeeb et al. (2013)
PDT	Probabilities decrease	No	Bromocriptine	D2-like agonist	1.0 & 5.0 mg/kg	↑risky choice (5.0)	St Onge & Floresco (2009)
PDT	Probabilities decrease	No	Quinpirole	D2-like agonist	1.0 & 10.0 mg/kg	Impairs discriminability of reinforcer magnitudes (10.0)	St Onge et al. (2011)
PDT	Probabilities decrease	No	Quinpirole	D2-like agonist	0.003–0.03 mg/kg	↑risky choice (0.01)	Oinio et al. (2017)
PDT	Probabi litics decrease	No	Quinpirole	D2-like agonist	0.1 & 0.5 mg/kg	↑risky choiceh	Wallin-Miller et al. (2018)
PDT	Probabilities decrease	Yes	Quinpirole	D2-like agonist	0.1 & 0.3 mg/kg	↓ sensitiivty to reinforcer magnitude	Smith et al. (2018)
PDT	Probabilities decrease	No	Quinpirole	D2-like agonist	0.1 & 0.3 mg/kg	↓ sensitiivty to reinforcer magnitude	Smith et al. (2018)
RDT	Probabilities increase	No	Bromocriptine	D2-like agonist	1.0–5.0 mg/kg	↓risky choice (3.0 & 5.0)	Simon et al. (2011)
rGT	N.A	No	Bromocriptine	D2-like agonist	1.0–5.0 mg/kg	No effect	Zeeb et al. (2009)
rGT	N/A	No	Quinpirole	D2-like agonist	0.0125–0.125 mg/kg	No effect	Zeeb et al. (2009)
PDT	Probabilities decrease	No	Eticlopride	D2-like antagonist	0.01 & 0.03 mg/kg	↓risky choicei	St Onge & Floresco (2009)
PDT	Probabilities decrease	No	Eticlopride	D2-like antagonist	0.1 & 1.0 μg in mPFC	↓risky choice (1.0)j	St Onge et al. (2011)
PDT	Probabilities decrease	Yes	Eticlopride	D2-like antagonist	0.003–0.017 mg/kg	No effect	Smith et al (2018)
PDT	Probabilities decrease	No	Eticlopride	D2-like antagonist	0.003–0.017 mg/kg	No effect	Smith et al (2018)
RDT	Probabilities increase	No	Eticlopride	D2-like antagonist	0.01 −0.05 mg/kg	No effect	Simon et al. (2011)
rGT	N/A	No	Eticlopride	D2-like antagonist	0.01 mg/kg	↑P2; ↓P3/P4	Zeeb et al. (2009)
rGT	N/A	No	Eticlopride	D2-like antagonist	0.03 mg/kg	↓P4	Zeeb et al. (2009)
rGT	N/A	No	Eticlopride	D2-like antagonist	0.06 mg/kg	No effect	Zeeb et al. (2009)
rGT	N/A	No	Eticlopride	D2-like antagonist	0.001 mg/kg	↑P2e	Zeeb et al. (2013)
rGT	N/A	No	Eticlopride	D2-like antagonist	0.003 mg/kg	↑P1k	Zeeb et al. (2013)
rGT	N/A	No	Eticlopride	D2-like antagonist	0.01 mg/kg	↓ P1/P4f	Zeeb et al. (2013)
rGT	N/A	No	L-74l,626	D2-like antagonist	0.1–1.0 mg/kg	No effect	Di Ciano et al (2015)
PDT	Probabilities decrease	No	PD128,907	D3-like agonist	0.1 –0.5 mg/kg	↓risky choice (0.25)l	St Onge & Floresco (2009)
PDT	Probabilities decrease	No	Pramipexole	D3-like agonist	2.0 mg/kgm	↑risky choice	Rokosik & Napier (2012)
PDT	Probabilities decrease	No	Pramipexole	D3-like agonist	0.3 & 1.2 mg/kgn	↑risky choice	Holtz et al. (2016)
PDT	Probabilities decrease	No	Pramipexole	D3-like agonist	0.1 & 0.3 mg/kg	Impairs discriminability of reinforcer magnitudes (0.3)	Pes et al. (2017)
rGT	N/A	No	PD128,907	D3-like agonist	0.03–1.0 mg/kg	No effect	Di Ciano et al. (2015)
rGT	N/A	No	PD128,907	D3-like agonist	0.01–0.1 mg/kg	No effect	Barrus & Winstanley (2016)
rGT	N/A	Yes	PD128,907	D3-like agonist	0.01–0.1 mg/kg	↑P3(0.1)	Barrus & Winstanley (2016)
PDT	Probabilities decrease	No	Nafadotride	D3-like antagonist	0.5–2.0 mg/kg	No effect	St Onge & Floresco (2009)
rGT	N/A	No	SB-277011-A	D3-like antagonist	0.3–3.0 mg/kg	No effect	Di Ciano et al. (2015)
rGT	N/A	No	SB-277011-A	D3-like antagonist	0.5–5.0 mg/kg	No effect	Barrus & Winstanley (2016)
rGT	N/A	Yes	SB-277011-A	D3-like antagonist	0.5–5.0 mg/kg	↓P3 (0.5 & 1.5)	Barrus & Winstanley (2016)
PDT	Probabilities decrease	No	PD168,077	D3-like agonist	0.5–5.0 mg/kg	No effect	St Onge & Floresco (2009)
rGT	N/A	No	PD168,077	D3-like agonist	0.5–10.0 mg/kg	No effect	Di Ciano et al. (2015)
rGT	N/A	No	PD168,077	D3-like agonist	0.5–5.0 mg/kg	No effect	Barrus & Winstanley (2016)
rGT	N/A	Yes	PD168,077	D3-like agonist	0.5–5.0 mg/kg	No effect	Barrus & Winstanley (2016)
PDT	Probabilities decrease	No	L745,870	D3-like antagonist	0.5–5.0 mg/kg	No effect	St Onge & Floresco (2009)
rGT	N/A	No	L745,870	D3-like antagonist	0.5–5.0 mg/kg	No effect	Di Ciano et al. (2015)
rGT	N/A	No	A-381393	D3-like antagonist	0.5–5.0 mg/kg	No effect	Barrus & Winstanley (2016)
rGT	N/A	Yes	A-381393	D3-like antagonist	0.5–5.0 mg/kg	No effect	Barrus & Winstanley (2016)

Note. PDT = probability discounting task; RDT = risky decision task; rGT = rat gambling task; mPFC = medial prefrontal cortex; DAT = dopamine transporter; NET = norepinephrine transporter; N/A = not applicable.

^aEffect at 0.125 mg/kg dose observed in rats chronically treated with amphetamine only

^bOnly two probabilities were tested: 100% and 25%

^cAffected female rats only

^dAffected male rats only, as females showed a complete suppression in behavior

^eAffected pair-housed rats only

^fAffected environmentally enriched rats only

^gAffected pair-housed and environmentally enriched rats, but not in isolated rats

^hThe 0.1 mg/kg dose increased risky choice in rats previously treated with testosterone, whereas the 0.5 mg/kg dose increased risky choice in rats previously treated with testosterone and vehicle

ⁱThe 0.03 mg/kg dose caused a decrease in responding when the probability was set to 100%

^jThis dose appears to have decreased choice at the 100% probability

^kAffected isolated rats only

^lThe highest dose impaired discriminability of the large and small magnitude reinforcers

^mPramipexole was administered twice daily for 13 days

ⁿPramipexole was delivered via osmotic pump.

Dopamine

Amphetamine, which is known to reverse dopamine, as well as norepinephrine, transporters, differentially alters risky choice across the various tasks described above. Specifically, amphetamine increases risky choice in the PDT (Floresco & Whelan, 2009; Oinio et al., 2017; St Onge, Chiu, & Floresco, 2010; St Onge & Floresco, 2009) and decreases optimal choice in the rGT, as rats are less willing to choose the P2 option (Barrus & Winstanley, 2016; Silveira, Malcom, Shoaib, & Winstanley, 2015; Zeeb et al., 2009; Zeeb et al., 2013). However, amphetamine typically decreases risky choice in the RDT (Mitchell, Vokes, Blakenship, Simon, & Setlow, 2011; Orsini et al., 2016; Simon et al., 2009; Simon et al., 2011; but see Orsini et al., 2018). Inconsistencies are also observed across tasks following administration of dopamine selective agonists/antagonists. The dopamine D₁ receptor antagonist SCH 23390 decreases risky choice in the PDT (Smith, Hofford, Zentall, & Beckmann, 2018; St Onge, Abhari, & Floresco, 2011; St Onge & Floresco, 2009), but has no effect on choice in the RDT (Simon et al., 2011) or rGT (Zeeb et al., 2009; Zeeb et al., 2013). Overall, stimulating D₁ receptors does not alter performance in the PDT (Smith et al., 2018; St Onge et al., 2011; Wallin-Miller, Kreutz, Li, & Wood, 2018), in the RDT (Simon et al., 2011), or in the rGT (Zeeb et al., 2009), but one study reported an increase in risky choice following administration of SKF 81297 in the PDT (St Onge & Floresco, 2009).

Similar to the inconsistencies described above, discrepancies have been observed in risky choice following administration of dopamine D₂ ligands. Specifically, stimulating D₂ receptors with either quinpirole or bromocriptine increases risky choice in the PDT (Oinio et al., 2017; St Onge & Floresco, 2009; Wallin-Miller et al., 2018), decreases risky choice in the RDT (Simon et al., 2011), but has no effect in the rGT (Zeeb et al., 2009). Concerning the effects of D₂ antagonists, eticlopride increases risk aversion in the PDT (St Onge et al., 2011; St Onge & Floresco, 2009; but see Smith et al., 2018 for null effects) and increases optimal choice in the rGT, as evidenced by an increase in responses for the P2 option and a decrease in responses for the P3/P4 options (Zeeb et al., 2009; but see Di Ciano et al., 2015 for null effects with the D₂ antagonist L-741,626 in the rGT). Although blocking D₂ receptors alters risky choice in the PDT and the rGT, there is no evidence that antagonizing this receptor alters RDT performance (Simon et al., 2011).

Most of the studies assessing the contribution of the dopaminergic system to risky choice have primarily examined D₁ and D₂ receptors although some have tested the effects of D₃/D₄ ligands. Stimulating D₃ receptors with PD128,907 selectively increases risk aversion in the PDT at an intermediate dose (0.25 mg/kg) but impairs discriminability of the large and small magnitude reinforcers at a higher dose (0.5 mg/kg; St Onge & Floresco, 2009). However, stimulating these receptors with pramipexole results in increased risky choice (Holtz, Tedford, Persons, Grasso, & Napier, 2016; Rokosik & Napier, 2012). Finally, Pes et al. (2017) observed that pramipexole decreases responding for the large, probabilistic reinforcer when its delivery is set to 50% (note: Pes et al., 2017 did not test subjects at the 100% probability) but increases responding for this alternative when its delivery is more probabilistic (6.25%). These results suggest that pramipexole may cause a general impairment in PDT performance. In the rGT, stimulating D₃ receptors increases choice for the P3 option only (Barrus & Winstanley, 2016). D₃ receptor antagonists do not affect risky choice in the PDT (St Onge & Floresco, 2009) but decrease choice of the P3 option in the rGT (Barrus & Winstanley, 2016). Finally, D₄ receptors do not appear to mediate risky choice (Barrus & Winstanley, 2016; Di Ciano et al., 2015; St Onge & Floresco, 2009).

Serotonin

The contribution of the serotonergic system to risky choice has been studied exclusively in the rGT. Stimulating 5-HT_1A receptors with 8-OH-DPAT increases choice for the P1 and P3 options at the intermediate dose (0.3 mg/kg), but increases responses for the P3 option only at the highest dose (0.6 mg/kg; Zeeb et al., 2009). Although blocking 5-HT_1A receptors does not affect risky choice, it does attenuate the effects of 8-OH-DPAT on rGT performance (Zeeb et al., 2009). When examining the contribution of 5-HT₂ receptors, Adams, Barkus, Ferland, Sharp, and Winstanley (2017) found that blocking the 5-HT_2A receptor or stimulating the 5-HT_2C does not alter risky choice, whereas blocking 5-HT_2C receptors increases responses for the P2 option, thus decreasing risky choice. However, one important consideration is that blocking 5-HT_2C receptors increases premature responses, a measure of motor impulsivity. Thus, the 5-HT_2C receptor may not be a viable target for treating disorders characterized by excessive risk.

Norepinephrine

Thus far, the contribution of adrenergic receptors to risky choice has been explored in the PDT only. Atomoxetine, a selective norepinephrine transporter inhibitor, increases risky choice (Yang, Pan, & Li, 2016). Although Montes, Stopper, and Floresco (2015) did not find a significant effect of atomoxetine on choice, they observed an increase in risky decision making in rats with a low baseline level of risky choice. Atomoxetine’s ability to increase risky choice may be mediated by beta receptors, as propranolol (beta receptor antagonist), but not prazosin (alpha 1 antagonist), blocks the atomoxetine-induced risky choice (Yang et al., 2016). Administration of an alpha 2 agonist increases risk aversion (Montes et al., 2015). Interestingly, the effects of an alpha 2 antagonist are dependent on the order in which probabilities are presented; yohimbine increases risky choice when the probabilities decrease across the session but decreases risky choice when the probabilities increase across the session (Montes et al., 2015; note: the effects of probability presentation order on drug effects will be discussed in more detail in the Methodological Considerations in Risky Choice Tasks section).

Acetylcholine

The effects of nicotine (nicotinic receptor agonist) on risky choice are inconsistent across procedures. Whereas nicotine increases risky choice in the PDT (Mendez, Gilbert, Bizon, & Setlow, 2012), it increases risk aversion in the RDT (Mitchell et al., 2011), although early-life administration of nicotine does not affect behavior in this task (Mitchell et al., 2012). Nicotine does not affect performance in the rGT (Silveira et al., 2015). Although stimulating nicotinic receptors alters risky choice, albeit in different ways across the PDT and the RDT, blocking nicotinic receptors does not alter risky choice in either the PDT or in the rGT (Mendez et al., 2012; Silveira et al., 2015).

Stimulating muscarinic receptors does not alter risky choice (Mendez et al., 2012; Silveira et al., 2015), but it does increase premature responses in the rGT (Silveira et al., 2015). Conversely, blocking muscarinic receptors tends to impair decision making in these tasks. In the PDT, scopolamine flattens the discounting curve, suggesting impairment in discriminability of each reward alternative (Mendez et al., 2012). Unlike scopolamine, atropine does not flatten the discounting curve; however, atropine increases suboptimal choice by increasing preference for the large magnitude reinforcer when the probability of its delivery is low (Mendez et al., 2012). In the rGT, scopolamine increases choice of the P1 option (Silveira et al., 2015). In this case, rats are not maximizing the number of pellets they can earn following scopolamine administration.

Glutamate

The contribution of the glutamatergic system to risky choice has been primarily examined using the PDT in my laboratory. Most of our work has focused on the N-methyl-D-aspartate (NMDA) receptor, as our earliest work found that blocking AMPA receptors did not affect PDT performance (Yates, Batten, Bardo, & Beckmann, 2015). So far, results with uncompetitive receptor antagonists (e.g., MK-801, ketamine) have been somewhat mixed. MK-801 (i.e., dizocilpine) increases risky choice when the probability of obtaining the large magnitude reinforcer decreases across the session (Yates et al., 2015; Yates et al., 2016), but decreases risky choice when the probabilities increase across the session (Yates et al., 2016). When the probabilities decrease across the session, ketamine decreases preference for the large magnitude reinforcer, even when its delivery is guaranteed; however, this effect is only observed when a fixed ratio (FR) 1 schedule of reinforcement is used (Yates et al., 2015). When an FR 10 schedule is used, this effect is not observed (Yates et al., 2016). Although ketamine does not alter PDT performance when an FR 10 schedule of reinforcement is used, Yates et al. (2016) did find that ketamine decreases risky choice under this schedule, as long as the probability of receiving the large magnitude reinforcer increases across the session.

Recent research has focused on the contribution of the NR2B subunit of the NMDA receptor to risky choice. Similar to the effects of ketamine described above, the NR2B-selective antagonist ifenprodil does not alter PDT performance when the probabilities decrease across the session, but decreases risky choice when the probabilities increase (Yates et al., 2016). Probability presentation-dependent effects are also observed following administration of Ro 63–1908 and CP-101,606, two highly selective NR2B subunit antagonists. When the probabilities decrease across the session, an intermediate dose of Ro 63–1908 (0.3 mg/kg)/CP-101,606 (1.0 mg/kg) increases risky choice; however, the highest dose of each drug (1.0 and 3.0 mg/kg, respectively) decreases risky choice (Yates et al., in press). Finally, in the rGT, blocking the NR2B subunit does not alter risky choice, but it does increase motor impulsivity in this task (Higgins, Silenieks, MacMillan, Zeeb, & Thevarkunnel, 2018).

Methodological Considerations in Risky Choice Tasks

Given the methodological differences across risky choice tasks, it is not quite surprising that discrepant findings have been reported in the literature. Even when focusing on a single task, one important consideration is that modifications can be made to a procedure, which can alter how a drug mediates risky choice. Some of these methodological considerations will be discussed in detail below.

Using Ascending/Descending Schedules

Using ascending and descending schedules is applicable to the PDT and to the RDT. In the PDT, the probability of obtaining the large magnitude reinforcer is generally decreased across the session only (e.g., Mendez et al., 2012; Smith et al., 2018; St Onge & Floresco, 2009; Wallin-Miller et al., 2018; Yang et al., 2016; Yates et al., 2015). This is somewhat problematic because studies have shown that the order in which probabilities are presented can alter how pharmacological manipulations affect risky choice in the PDT. For example, I have already discussed the opposite effects drugs such as yohimbine (Montes et al., 2015), MK-801 (Yates et al., 2016), and Ro 63–1908 (Yates et al., in press) have on performance in the PDT using ascending/descending schedules. Similarly, St Onge et al. (2010) found that amphetamine increases risky choice when the probability of earning the large reinforcer decreases across the session but observed decreased risky choice when the probabilities increase across the session. In the RDT, infusions of amphetamine into medial prefrontal cortex (mPFC) do not alter risky choice when the probability of receiving shock increases the session, but intra-mPFC infusions decrease preference for the large magnitude reinforcer when the probability of receiving shock decreases across the session (Orsini et al., 2018). Interestingly, although Yates et al. (2016) and Yates et al., (in press) reported differential baseline levels of risky choice in rats trained on ascending/descending schedules, St Onge et al. (2010) found no such differences between these two groups (note: St Onge et al., 2010 reported increased risky choice in rats trained on a pseudo-randomized schedule). Furthermore, baseline differences are not apparent between rats trained on the ascending and descending schedules in the RDT (Orsini et al., 2018; Simon et al., 2009). Thus, the differential drug effects observed across ascending/descending schedules are not necessarily due to differences in baseline levels of risky choice.

At first glance, the opposite drug effects observed for drugs such as amphetamine may reflect increased perseverative responding. For rodents that start each session with the highest probability are more likely to respond on the alternative with the large, probabilistic reinforcer relative to the small, guaranteed reinforcer. As such, when the probability of obtaining the large magnitude reinforcer decreases across the session, rodents may perseverate on the lever associated with the large magnitude reinforcer. Conversely, when rodents begin a session in which the probability of receiving the large magnitude reinforcer is low, they are more likely to respond for the small, guaranteed alternative. When the probability of receiving the large magnitude reinforcer increases across the session, subjects may perseverate on the lever associated with the small magnitude reinforcer. St Onge et al. (2010) offer an alternative explanation for the discrepant effects observed with amphetamine. They argue that the discrepant effects of amphetamine reflect increased perseverative responding on the perceived relative value of the probabilistic reinforcer as the probability changes. For example, when the probability of receiving the large magnitude reinforcer is initially set at 1, rats treat this option as more advantageous even as the probability of earning that reinforcer decreases throughout the session. Conversely, when the probability of earning the large magnitude reinforcer is initially set to a value of 0.0625, rats perceive this option as less advantageous, even as the probability of receiving this alternative increases across the session.

Signaling Wins and Losses

In the PDT and in the rGT, wins and losses can be signaled with stimuli, which, in turn, can modulate drug effects in these procedures (note: signaling the delivery of shock in the RDT could be done, but this has not been tested, yet). For example, Smith et al. (2018) show greater probability discounting in rats that do not receive a signal indicating the outcome of each trial; additionally, administration of the dopamine D₁ receptor antagonist SCH 23390 increases risk aversion in the unsignaled group only, although SCH 23390 tended to increase risky choice in the signaled group. The lack of effect observed for the signaled group could be due to the fact that rats trained in this condition exhibited high basal levels of risky choice, this reflecting a potential ceiling effect. In the rGT, amphetamine increases choice of the P1 option and decreases choice of the P2 option, but only when rats are in the unsignaled group (Barrus & Winstanley, 2016). Barrus and Winstanley (2016) also show that PD128907, a dopamine D₃ agonist, increases choice of the P3 option, but only when rats receive an audio-visual cue when they receive reinforcement.

One potential explanation for the discrepant drug effects discussed above is that the dopaminergic system is involved in guiding responses for conditioned (secondary) reinforcers. Cues that signal reinforcement can become reinforcers in themselves, and drugs that target the dopaminergic system have been shown to increase responding for these secondary reinforcers (Robbins, Watson, Gaskin, & Ennis, 1983). The results described above are not necessarily consistent with the idea that the discrepant results are due to the use of conditioned reinforcers. For most of the studies described above, dopaminergic ligands altered choice in the unsignaled group only. Even though PD128907 increased choice for the P3 option in the cued version of the rGT, interpreting these results is somewhat difficult because the change in P3 responding was not accompanied by significant decreases in responding for the other alternatives. Thus, what does a change in responding for the P3 option reflect if there is no change in responding for the optimal P2 option? Overall, more research is needed to determine why signaling wins seems to negate the effects of dopamine receptor ligands on choice.

Schedule of Reinforcement

Across all three tasks, an FR 1 schedule of reinforcement is typically used. There is evidence that changing the FR schedule can alter risky choice in animals. Kaminski and Ator (2001) show that risky choice is inversely related to the FR schedule. As the FR requirement increases from an FR 1 to an FR 16, rats are less likely to choose the risky option. Additionally, there is some indirect evidence that changing the schedule of reinforcement can modulate how drugs affect PDT performance. When an FR 1 schedule is used in the PDT, MK-801 increases risky choice (Yates et al., 2015); yet, this effect is not replicated when an FR 10 schedule is used (Yates et al., 2016). As already discussed above, when the probability of obtaining reinforcement decreases across the session, ketamine decreases responding for the large magnitude reinforcer when an FR 1 schedule is used (Yates et al., 2015), but has no effect when an FR 10 is used (Yates et al., 2016). Because these results come from different studies, caution is needed when making interpretations about the role of response requirements on drug effects in these procedures. Future studies that directly compare drug effects across different response requirements are required to provide further support for the claim that response requirement can modulate drug effects.

Analytic Considerations in Risky Choice Tasks

Due to the multiple ways in which PDT/RDT data can be analyzed, this section will focus primarily on these tasks.

Analyzing PDT/RDT Data

Raw proportion of responses for the large magnitude reinforcer.

The raw proportion of responses for the large probabilistic (PDT)/large unsafe (RDT) reinforcer is often analyzed with two-way/three-way ANOVAs. In the PDT/RDT, the raw proportion of choices for the large magnitude reinforcer are plotted as a function of probability (either probability of receiving reinforcement or receiving shock) and are analyzed with probability and drug dose as within-subjects factors. Significant interactions can be probed with follow-up ANOVAs and/or paired-samples t-tests. Although ANOVAs are commonly used, this analysis has multiple limitations that have been described in detail elsewhere (Yates, 2018; Young, Clark, Goffus, & Hoane, 2009). This review will briefly cover a couple of the most relevant issues of using ANOVAs to determine drug effects in risky choice tasks. First, because significant interactions have to be probed with additional ANOVAs and/or paired-samples t-tests, the alpha level needs to be adjusted to control for Type I error. For example, in most PDT/RDT experiments, four or five probabilities are used; therefore, alpha needs to be divided by 4 or 5, resulting in an adjusted alpha level of .0125 or .01. Due to the decreased power that comes from adjusting the alpha level, some studies fail to include Bonferroni corrections (St Onge et al., 2010; St Onge & Floresco, 2009) or simply do not report the results of post hoc tests (e.g., Mitchell et al., 2011; Simon et al., 2011). Although other corrections can be used to control for Type I error that are less conservative than Bonferroni (e.g., Sidak, Holm-Bonferroni), the argument made above is still applicable.

Second, there are situations in which a subject does not respond during one block of trials following administration of one drug dose. In ANOVA, a subject’s entire dataset is excluded if there are any missing data. This is problematic because real, informative data are being discarded. For example, in one of our experiments, one rat treated with the intermediate dose of ketamine (5.0 mg/kg) did not respond during two blocks of trials (Yates et al., 2016). As such, if we had used ANOVA analyses, this subject’s entire dataset would have been excluded due to listwise deletion, despite the fact this subject completed each trial following administration of every other dose. One way to handle missing data is to use some form of imputation, but there are drawbacks to using imputation (see Young, 2017 for a discussion). An alternative to handling missing data will be discussed in more detail below.

Using hyperbolic/exponential functions.

Because the proportion of responses for the large magnitude reinforcer decreases as a function of probability (of either food delivery or shock) in a hyperbolic/exponential fashion, the equations V = A/(1 + hθ) or V = Ae^-hθ can be used, respectively. Here, V represents the subjective value of the large magnitude reinforcer, A is the intercept of the function and indicates how much a subject responds for the large reinforcer relative to the small reinforcer when both reinforcers are made available (note: in the RDT, the intercept denotes how much the subject chooses the large reinforcer relative to the small reinforcer when the probability of receiving shock is 0), h is the slope of the function and denotes how quickly animals stop responding for the large magnitude reinforcer as a function of probability/shock, and θ is odds against. In the PDT, odds against are calculated as 1/probability of receiving reinforcement-1 (Rachlin, Raineri, & Cross, 1991). In the RDT, odds against would be calculated as 1/probability of receiving safe reinforcement-1.

Using nonlinear regression to analyze PDT data has been used twice (Yates et al., 2015; Yates et al., 2016). Yates et al. (2015) fit a hyperbolic discounting function to determine how the NMDA receptor uncompetitive antagonists MK-801 and ketamine altered risky choice. MK-801 selectively decreased h parameter estimates without altering A parameter estimates, suggesting that MK-801 selectively alters risky choice. Conversely, ketamine did not alter the h parameter but decreased the A parameter; thus, ketamine appears to have impaired the discriminability of the reinforcer magnitudes. Yates et al. (2016) used an exponential discounting function as opposed to the hyperbolic function to determine how probability presentation order modulates the ability of NMDA receptor antagonists to alter A/h parameter estimates. In contrast to the findings reported by Yates et al. (2015), MK-801 did not affect A or h parameter estimates when the odds against obtaining reinforcement increased across the session, and ketamine failed to decrease A parameter estimates (Yates et al., 2016). However, MK-801 and ketamine increased h parameter estimates when the odds against obtaining reinforcement decreased across the session (Yates et al., 2016).

In both the Yates et al. (2015) and Yates et al. (2016) studies, the A and h parameter estimates were derived via nonlinear regression and then subjected to a secondary analysis (e.g., ANOVA, Friedman test). Using two-stage approaches to analyzing data can be problematic (see Jonsson, Wade, & Karlsson, 2000; Yates, 2018; Young, 2017 for discussions) as they have the same limitations as discussed above (e.g., listwise deletion of an entire subject’s dataset when partially missing data are included, increased Type I error). Also, in cases in which an animal shows extreme sensitivity to probabilistic reinforcement (e.g., stops responding for the large magnitude reinforcer as soon as the probability of earning that alternative decreases), the h parameter obtained for that subject can significantly increase within-groups variability. Table 2 illustrates this phenomenon. Using data collected from my laboratory (Yates et al., in press), the hyperbolic discounting function was fit to individual subject data following administration of Ro 63–1908. The h parameter estimates shown on the left side of Table 2 (note: I will discuss the right side of Table 2 in a subsequent paragraph) were derived via GraphPad Prism, a program that allows one to create figures and conduct some statistical analyses/curve fitting. When using this approach, the mean h parameter (averaged across each dose) is 2.444, and the standard error of the mean (averaged across each dose) is 0.894. Normally, h parameter estimates are log-transformed before being subjected to a secondary analysis (e.g., Yates et al., 2015; Yates et al., 2016). Although log transformations can reduce within-groups variability, there is a better approach to generating h parameter estimates without having to rely on data transformations.

Table 2.

Individual subject h parameter estimates derivedfrom the hyperbolic discounting function using either nonlinear regression via GraphPad or nonlinear mixed effects modeling (NLME) via R

GraphPad					NLME
Subject	Vehicle	0.1 mg/kg	0.3 mg/kg	1.0 mg/kg	Subject	Vehicle	0.1 mg/kg	0.3 mg/kg	1.0 mg/kg
401	0.657	0.099	0.061	0.164	401	0.381	0.138	0.059	0.105
402	0.371	0.121	0.035	0.227	402	0.492	0.172	0.035	0.153
403	5.545	3.904	3.904	5.545	403	1.893	1.634	1.531	1.730
404	8.009	3.904	0.166	0.283	404	1.131	0.506	0.135	0.558
409	0.198	0.170	0.046	0.072	409	0.354	0.123	0.054	0.087
410	1.183	7.188	0.228	0.286	410	0.938	0.489	0.252	0.515
411	0.252	0.371	0.178	0.074	411	0.243	0.144	0.173	0.091
412	0.349	0.283	0.148	3.085	412	0.919	0.420	0.145	0.447
417	8.009	8.009	8.009	8.009	417	1.992	1.786	1.723	1.888
418	0.106	0.526	8.009	0.506	418	0.115	0.551	0.983	0.457
419	8.009	8.009	2.571	1.059	419	1.673	1.429	1.339	1.506
420	1.858	0.526	3.494	3.494	420	1.325	1.163	1.137	1.205
Average	2.879	2.759	2.237	1.900	Average	0.955	0.713	0.631	0.728
SEM	0.990	0.953	0.881	0.753	SEM	0.190	0.179	0.189	0.194

Instead of using two-staged approaches to analyze parameter estimates derived from hyperbolic/exponential functions, nonlinear mixed effects modeling (NLME) can be used to analyze PDT/RDT data. Briefly, mixed models include fixed effects and random effects. A fixed effect is a variable associated with the entire group of subjects (e.g., dose), and a random effect allows within-subjects variables to have different effects for each individual subject. For example, including a random effect of dose allows each subject to have a different dose effect, whereas the fixed effect of dose captures the group effect of dosage that arises from those subject estimates. Instead of using error minimization, NLME uses maximum likelihood estimation, in which the goal is to select the most likely value of a parameter as opposed to the value that minimizes error. As such, NLME is advantageous because it better controls for Type I error, has increased statistical power, and accounts for partially missing data (i.e., they do not delete an entire subject’s dataset if some data are missing; see Young et al., 2009 for a discussion of mixed models). Because NLME simultaneously calculates the parameter estimates for each individual subject and for the entire group, this approach is advantageous because it prevents outliers from having a large effect on the results of the analysis (see Young, 2017 for a discussion). Going back to the previous example, using the NLME package in R allows one to minimize the standard error of the mean from 0.894 to 0.188. Going back to Table 2, notice that some subjects (e.g., 404, 417, and 419) had much larger h parameter values (8.009) following some of the treatments compared to most of the other subjects when parameter estimates are generated with GraphPad. When NLME is used, the parameter estimates generated for these rats are much lower and are closer to the average h parameter estimate (see the right side of Table 2; note: this phenomenon is known as the shrinkage effect).

To demonstrate how NLME analyses can lead to different conclusions relative to analyzing the raw proportion of responses for the large magnitude reinforcer via ANOVA, I have applied this analysis to two previously published datasets using either the PDT (Yates et al., in press) or the RDT (Orsini et al., 2016). The Yates et al., (in press) study compared the effects of Ro 63–1908 in rats trained on an ascending schedule or on a descending schedule. When a three-way ANOVA was applied, there was a significant schedule × dose interaction. To probe this interaction, separate two-way ANOVAs were conducted, one for each schedule. However, these follow-up analyses revealed main effects of probability only (Fig. 1a and 1b). When a hyperbolic discounting function is fit to the data via NLME, there is a significant schedule × dose interaction (p = .048). Planned contrasts show that the intermediate dose of Ro 63–1908 (0.3 mg/kg) significantly increases risky choice when the odds against obtaining the large reinforcer increases across the session, whereas Ro 63–1908 does not have a significant effect on choice when the odds against decrease (Fig. 1c and 1d).

Figure 1.

Mean (± SEM) proportion of responses following Ro 63–1908 administration for rats trained on an ascending schedule (a) or on a descending schedule (b). Mean (± SEM) A parameter estimates (left y-axis) and h parameter estimates derived using the hyperbolic function (panels c and d) and the exponential function (panels e and f) for rats trained on an ascending (panels c and e) or a descending (panels d and f) schedule. *p < .05, relative to vehicle. Results of the ANOVA analyses for panels a and b: significant dose × schedule interaction, F(2.273, 49.999) = 4.234, p = .016. The follow-up analyses revealed trends of a dose effect only, F’s of 2.706 and 2.720; p’s of .060 and .061, respectively. Results of the NLME analyses for the hyperbolic function (panels c and d): significant dose × schedule interaction, F(3, 441) = 2.658, p = .048. Results of the NLME analyses for the exponential function (panels e and f): significant dose × schedule interaction, F(3, 441) = 5.137, p = .002. Note: data are derived from Yates et al., (in press).

Although nonlinear regression has not been applied to RDT data, this analysis can be incorporated to determine if drugs alter sensitivity to probabilistic punishment or discriminability of each reinforcer alternative. Orsini et al. (2016) compared the effects of amphetamine between male and female rats. By using ANOVA to analyze the raw proportion or responses, Orsini et al. (2016) argue that amphetamine decreases risky choice in male and female rats (see Figs. 2a and 2b). However, when fitting a hyperbolic discounting function to the data, amphetamine significantly decreases A parameter estimates (i.e., discriminability of reinforcer magnitudes; Fig. 2c) without altering h parameter estimates (i.e., sensitivity to probabilistic shock; Fig. 2d). By using NLME, interpreting the effects of amphetamine changes: instead of altering risky choice, amphetamine appears to impair rats’ ability to differentiate the small magnitude and large magnitude reinforcers. Using NLME in this context is advantageous because there were multiple female rats that did not respond during a block of trials (1 following vehicle injection; 1 following the 0.3 mg/kg dose; 4 following the 1.5 mg/kg dose). With ANOVA analyses, 6 of the 8 female rats would have been excluded from the analysis due to missing data. To compensate for the missing data, Orsini et al. (2016) entered zeroes for subjects that did not respond during a block of trials. At face value, this makes sense because the rats did not respond for the large magnitude reinforcer during these trials. The problem with taking this approach is that entering a zero implies that the subject exclusively preferred the small magnitude reinforcer during a block of trials, which is not the case here. By accounting for missing data, NLME can give a more accurate depiction of how a drug alters performance in the PDT or in the RDT.

Figure 2.

Mean (± SEM) proportion or responses for the large magnitude reinforcer in male rats (a) and female rats (b) performing in the RDT. Mean (± SEM) A parameter (c) and h parameter (d) estimates derived from the hyperbolic discounting function via NLME. *p < .05, relative to vehicle. #p < .05, relative to males. Results of the analyses for panels a and b: significant dose × sex × probability interaction, F(12, 168) = 3.11, p = .001. Results of the analyses for panels c and d: A parameter estimate, main effects of dose, F(3, 276) = 7.912, p < .001, and sex, F(1, 276) = 11.786, p < .001; h parameter estimate, no effects, all F’s ≤ 1.603, all p’s ≥ .207. Note: data were derived from Orsini et al. (2016).

One limitation of using NLME analyses is that one needs to specify which type of function best fits the data. Using one model over the other may lead to situations where data interpretation can change. For example, earlier I described the research my laboratory has done examining the contribution of the NR2B subunit to risky choice. Figure 1c/1d shows that Ro 63–1908 only affects choice in rats trained on the ascending schedule when the hyperbolic discounting function is used. However, if the exponential function is used, the increased risky choice observed following Ro 63–1908 (0.3 mg/kg) remains unchanged, but Ro 63–1908 (1.0 mg/kg) now significantly increases risk aversion in rats trained on the descending schedule (Figs. 1e and 1f). For the PDT, there is a general consensus that hyperbolic functions best describe discounting of probabilistic reinforcers (Rachlin et al., 1991; see Green & Myerson, 2004 for a review). However, research has not examined which model best accounts for performance in the RDT, although it is important to note that applying a hyperbolic function or exponential function does not alter the conclusions drawn from the Orsini et al. (2016) dataset discussed above. Still, future research examining hyperbolic/exponential functions in this task may be useful. Because of the theoretical nature of discounting functions, others have argued for the use of atheoretical models to describe discounting (see below).

Area under the curve (AUC).

AUC (Myerson, Green, & Warusawitharana, 2001) has been used to analyze PDT data in experiments outside of psychopharmacology research (Abidi et al., 2018; Lawyer, Williams, Prihodova, Rollins, & Lester, 2010; Lindbergh, Puente, Gray, MacKillop, & Miller, 2014; Mahoney & Lawyer, 2016; Matusiewicz, Carter, Landes, & Yi, 2013; Ohmura, Takahashi, Kitamura, & Wehr, 2006; Ostaszewski & Bialaszek, 2010; Vanderveldt, Green, & Myerson, 2015; Yi, Johnson, & Bickel, 2005), but some studies have used AUC to determine how pharmacological manipulations alter risky choice (Boutros, Semenova, Liu, Crews, & Markou, 2014; Ohmura, Takahasi, & Kitamura, 2005; Yates et al., in press). Because AUC provides a “theoretically neutral” measure of discounting, one does not need to specify if a hyperbolic function or an exponential function best describe the data. To calculate AUC, the probability and subjective value for each data point need to be normalized. The odds against are expressed as a proportion of the maximum odds against, and the subjective value is expressed as a proportion of the nominal amount. The figure is then subdivided into a series of trapezoids (see Fig. 3 of Myerson et al., 2001), and the area of each trapezoid is equal to (x₂-x₁)[(y₁+y₂)/2], where x₁ and x₂ are successive odds against, and y₁ and y₂ are the subjective values associated with these odds against. AUC is equal to the sum of each trapezoid. AUC values range from 0–1, with 0 indicating an exclusive preference for the small immediate/safe option and 1 indicating an exclusive preference for the large probabilistic/unsafe option.

Figure 3.

Mean (± SEM) AUC values following administration of CP-101,606 in the PDT for rats trained on an ascending schedule (a and c) and on a descending schedule (b and d). For panels a and b, AUCs were calculated as described in Myerson et al. (2001). For panels c and d, AUCs were calculated as described in Borges et al. (2016) using an ordinal scaling transformation of probability. *p < .05, relative to vehicle. Results of LME analyses for panels a and b: no main effect of dose, F(3, 66) = 1.473, p = .230; no significant schedule × dose interaction, F(3, 66) = 2.134, p = .104. Results of LME analyses for panels c and d: no main effect of dose, F(3, 66) = 0.093, p = .964; significant schedule × dose interaction, F(3, 66) = 3.156, p = .031. Note: data are derived from Yates et al., (in press).

One advantage of using AUC relative to the raw proportion of responses is that this simplifies the analysis. Instead of having to conduct a two-way ANOVA on the raw proportion of responses, a one-way ANOVA can be used. However, listwise deletion will still occur for subjects that have partially missing data. One way to circumvent this issue is to use linear mixed effects modeling (LME) instead of ANOVA on AUC data. LME has the same advantages as NLME that were discussed above (see Young et al., 2009). Thus far, LME has only been used once to assess drug effects on AUC values (Yates et al., in press).

Although AUCs are often calculated as described by Myerson et al. (2001), Borges, Kuang, Milhorn, and Yi (2016) suggest using a base-10 logarithmic or ordinal scaling transformation of the x-axis. Although this recommendation was used in the context of delay discounting, using these transformations can be applied to AUCs calculated for the PDT/RDT as well. Transforming the x-axis is important because larger odds against values can disproportionately affect AUCs. To illustrate this point, in a recent study conducted in my laboratory (Yates et al., in press), we used an ordinal scaling transformation as suggested by Borges et al. (2016) to determine the effects of NR2B-selective antagonists on discounting. We found that Ro 63–1908 and CP-101,606 increased AUCs for rats trained on an ascending schedule, whereas these drugs decreased AUCs for rats trained on a descending schedule. However, if we had used the Myerson et al. (2001) method, the effects of CP-101,606 would have disappeared (see Fig. 3). This is most likely due to the fact that the major differences in responding for the large magnitude reinforcer occurred at smaller odds against, which are given a lower weight when the Myerson et al. (2001) method is used. By using an ordinal scaling transformation, each odds against value is weighted more equally. This demonstrates that the way in which AUCs are calculated can affect how drug effects are interpreted.

One major weakness of using AUCs is that generating a single data point to measure risky choice does not allow one to determine which behavioral mechanism is controlling behavior in the PDT/RDT. Consider the following example: Yates et al. (2015) show that ketamine (10.0 mg/kg) decreases choice for the large magnitude reinforcer when the odds against receiving reinforcement are 0 and 1 (100% and 50% probability; Fig. 4a). There is a corresponding decrease in AUCs (Fig. 4b), suggesting that ketamine decreases risky choice. Here, stating that ketamine decreases risky choice does not seem to be the best explanation; instead, this drug seems to impair the discriminability of the large and small magnitude reinforcers. When NLME analyses are applied to the data, ketamine produces a change in the A parameter estimate only (Fig. 4c), which provides support for the claim that ketamine does not alter risky choice per se. Overall, some caution needs to be taken when using AUC analyses. If a drug appears to affect choice for the large magnitude reinforcer when the probability of earning that alternative is 100%, relying exclusively on AUC values can lead to erroneous conclusions.

Figure 4.

(a) Mean (± SEM) proportion of responses for the large, probabilistic reinforcer following ketamine administration. (b) Mean (± SEM) AUC values calculated using the Borges et al. (2016) method. (c) Mean (± SEM) A parameter estimates (left y-axis) and h parameter estimates (right y-axis) derived from the hyperbolic discounting function via NLME analyses. *p < .05, relative to vehicle (c). Results of the analyses for panel a: main effect of dose, F(3, 33) = 13,749, p < .001, and significant dose × probability interaction, F(12, 132) = 4.197, p < .001. Results of the analyses for panel b: main effect of dose, F(3, 33) = 13.043, p < .001. Results of the analyses for panel c: significant effect on A parameter, F(3, 221) = 16.719, p < .001; no effect on h parameter, F(3, 221) = 0.813, p = .488. Note: data are derived from Yates et al. (2015).

Multilevel logistic regression.

One final approach to analyzing PDT/RDT data deserves mention. Instead of averaging choices across a block of trials to generate a proportion of responses for the large magnitude reinforcer, individual choice data can be analyzed with multilevel logistic regression. In the PDT/RDT, subjects make a binary choice; they either select the alternative associated with the small magnitude reinforcer, or they select the alternative associated with the large magnitude reinforcer (note: this is not always the case as subjects can fail to respond entirely during a trial). Therefore, choice during each trial can be coded as either 0 (small magnitude reinforcer) or 1 (large magnitude reinforcer). Logistic regression has been previously applied to delay discounting (Young, 2018) but has not been applied to PDT or RDT data. One important consideration needs to be discussed here. Young (2018) analyzed data that manipulated the delay ratio (i.e., the delay to delivery of the small magnitude reinforcer was always set to 0 s, but the delay to the large magnitude reinforcer changed throughout the session) and magnitude ratio (i.e., the magnitude of each alternative varied in the experiment). In the PDT/RDT, the magnitude ratio is held constant throughout the experiment. Therefore, the R syntax Young (2018) uses to apply the multilevel logistic regression analyses needs to be modified for the PDT/RDT (see Appendix).

Analyzing rGT Data

In the rGT, data are analyzed with repeated measures ANOVAs, and significant interactions are probed with paired-samples t-tests. As discussed above, the major issue with using ANOVA to analyze rGT data is controlling for Type I error, which is rarely done in experiments using this procedure. For example, Zeeb et al. (2009) report that eticlopride (0.01 and 0.03 mg/kg) decreases choice for the P4 option (p < .01 and .03, respectively). Considering there are four comparisons being made (vehicle vs. drug at each option [e.g., P1, P2, etc.]), the adjusted alpha should be .0125. If Bonferroni corrections are used, eticlopride (0.03 mg/kg) does not significantly alter responses for the P4 option. This is not surprising as the change in choice between vehicle and eticlopride (0.03 mg/kg) appears to be minimal (~25% following vehicle vs. ~20% following eticlopride). Similar results have been observed following administration of amphetamine (1.5 mg/kg) at the P4 option (Zeeb et al., 2009), 8-OH-DPAT (0.3 mg/kg) at the P1 and P3 options (Zeeb et al., 2009; note: Zeeb et al., 2009 reported a significant effect of the highest dose of 8-OH-DPAT on the P3 option but did not report inferential statistics supporting this claim), PD128907 (0.1 mg/kg) on the P3 option (Barrus & Winstanley, 2016), and SB277011-A (0.5 mg/kg) on the P3 option (Barrus & Winstanley. 2016). The major disadvantage to using Bonferroni corrections is that power is significantly lowered. This makes detecting real effects difficult due to the lower alpha level being used. Although some of the effects reported in the Barrus & Winstanley (2016) and Zeeb et al. (2009) studies did not control for Type I error, this is not to say that their conclusions are incorrect. The lack of statistical significance following Bonferroni corrections may be the direct result of decreased power.

As discussed above for PDT/RDT analyses, ANOVA is not necessarily the appropriate analysis for rGT data. Data from the rGT are often presented as a proportion (i.e., proportion of responses for each option); however, individual choice for each trial could be recorded and subjected to a multinomial logistic regression analysis. This analysis is similar to multilevel logistic regression, which was described in more detail above. Here, multinomial logistic regression should be used because there are four possible outcomes on each trial (P1 vs. P2 vs. P3 vs. P4) as opposed to two possible outcomes (large magnitude reinforcer vs. small magnitude reinforcer). Future studies utilizing the rGT should incorporate multinomial logistic regression.

Discussion

The goals of the current review were to (1) highlight some of the advantages and disadvantages of each of the major tasks used to measure risky choice, (2) provide a general overview of drug effects across each task, (3) discuss how procedural differences (e.g., order of probability presentation, use of stimuli to signal outcomes of each trial, etc.) can modulate drug effects, and (4) describe how data generated from these tasks are normally analyzed, as well as to provide alternative analyses that can be used. It is important to note that I am not advocating for one task over another. As discussed in this paper, each task has its advantages and disadvantages. Below, I will discuss a few potential explanations that may account for the discrepant drug effects across each task. I will also raise the possibility that one or more of these tasks measure a different construct other than risky choice. Finally, I will offer a few recommendations for assessing the neurochemical basis of risky choice.

Explaining Differential Drug Effects Observed Across Risky Choice Tasks

One interesting observation across the various risky choice tasks is that pharmacological manipulations often produce opposite effects in one task relative to another, particularly when comparing results from the PDT and the RDT. These discrepancies may be due to the fact that these tasks measure dissociable aspects of risky decision making. In the PDT and in the rGT, negative punishment is used (e.g., on some trials, subjects fail to receive the large magnitude reinforcer), whereas in the RDT positive punishment is used (e.g., foot shock). Assuming each task measures a dissociable aspect of risky choice, an important question to consider is what behavioral mechanisms are being altered by pharmacological manipulations in these tasks. In the PDT/RDT, a drug can alter (a) sensitivity to probabilistic/risky reinforcement (e.g., what is normally considered the measure of risky choice), (b) sensitivity to the large magnitude reinforcer (or, the discriminability of the large and small magnitude reinforcers), or (c) both of these parameters. In the rGT, these two parameters, in addition to sensitivity to time-out periods, influence behavior. In the PDT/RDT, determining if a drug alters the discriminability of the large/small magnitude reinforcers and/or sensitivity to probabilistic delivery of reinforcement/shock can be accomplished with nonlinear regression (as discussed above in the Analytic Considerations in Risky Choice Tasks section), although if one wants to better determine how drugs affect sensitivity to reinforcer magnitude, they should test different reinforcer magnitude ratios (e.g., Orduña, 2015).

To determine how drug treatments alter performance in the rGT, Zeeb et al. (2009) conducted control experiments in which they (a) kept the time-out duration constant (10 s) and (b) kept the probability of negative punishment constant (0.2). Even with these control conditions, Zeeb et al. (2009) failed to keep reinforcer magnitude constant. Also, they did not include a control condition in which only magnitude is manipulated (i.e., the probability of receiving punishment and the duration of punishment are held constant). Therefore, across the two control conditions, there are still two parameters that vary, which makes pinpointing the behavioral mechanisms controlling drug effects in this task difficult. For example, amphetamine (1.0 and 1.5 mg/kg) increases P1 choices but decreases P2 choices in the rGT. However, when probability and punishment are held constant, amphetamine does not significantly alter performance in these control groups. Because Zeeb et al. (2009) never tested a group that only manipulated magnitude, they cannot rule out the possibility that amphetamine is impairing rats’ ability to discriminate the 1-pellet and 2-pellet reinforcer alternatives. My own research has shown that amphetamine (1.0 mg/kg) simultaneously impairs the discriminability of a 4-pellet reinforcer and 1-pellet reinforcer and increases risky choice (Yates et al., unpublished results).

Although Zeeb et al. (2009) included some controls for the rGT, it is important to note that many studies using this procedure do not include these controls (e.g., Adams et al., 2017; Higgins et al., 2018; Silveira et al., 2015; Zeeb et al., 2013). One can understand why these controls are not included because testing three control conditions in conjunction with the rGT condition would significantly increase the sample size of each study. However, determining the precise behavioral mechanisms controlling behavior in the rGT is important. As such, future studies using the rGT should try to include the following groups to better determine how pharmacological manipulations alter choice in this procedure: (a) a group where reinforcer magnitude is manipulated only; (b) a group where the probability of receiving reinforcement is manipulated only; and (c) a group where punishment time-out is manipulated only. Alternatively, each control could be embedded in a single session. For example, during a block of trials (note: each block can be set to whatever length the experimenter wants in order to accommodate for the half-life of the drug tested), magnitude and time-out duration are held constant. During another block of trials, magnitude and probability of reinforcement are held constant. Another block of trials can hold time-out duration and probability of reinforcement constant. Finally, the rGT condition (i.e., none of the variables are held constant) can be incorporated. To control for order effects, these blocks of trials can be randomized across subjects. Because this latter suggestion has not been utilized, I cannot say for certain if it would work, but attempting this design would be of interest as it can minimize the number of subjects needed and could provide a better account of how pharmacological manipulations influence decision making in the rGT.

Given the discrepant results observed following dopaminergic ligands, another consideration that may account for these discrepancies relates to error prediction (Hollerman & Schultz, 1998). In their seminal study, Hollerman and Schultz (1998) found that dopaminergic neurons respond more early in training when subjects make more errors and reinforcement is unpredictable. As performance improves (i.e., reinforcement becomes more predictable), dopaminergic activity diminishes. Following training, when unexpected reinforcement is delivered, there is an increase in dopaminergic activity; conversely, when expected reinforcement is omitted, there is a decrease in dopaminergic activity. One major difference between the RTD and the PDT/rGT is that reinforcement is always delivered in the RDT, whereas food delivery is probabilistic in the PDT and in the rGT. These tasks may cause differential dopaminergic activity, which may, in turn, modulate drug effects in these tasks. For example, in the RDT food delivery is expected on each trial. Therefore, dopaminergic activity should remain relatively unchanged across each trial. Additionally, aversive stimuli (e.g., shock) do not appear to dramatically increase dopaminergic activity (Schultz, 1998). In the PDT/rGT, the expected/unexpected delivery of reinforcement may cause increases/decreases in dopaminergic activity, depending on the outcome of each trial. For example, if a rat earns a reinforcer on a trial in which the probability of earning it is low, this unexpected food delivery will result in an increase in dopaminergic activity. Conversely, on trials in which food is expected but an omission occurs, dopaminergic activity will decrease. The differential neuronal activity observed across the RDT and PDT/rGT may partially explain the discrepant results across studies. It is important to note that this potential explanation is speculative, and research is needed to determine if differences between the PDT/rGT and the RDT are due, at least in part, by differential dopaminergic activity following expected and unexpected reinforcement. One limitation to this explanation is that one would expect to see similar drug effects between the PDT and rGT, which is not always the case. Additionally, this account can only explain differential results with dopaminergic ligands. More research is needed to determine why opposite results are sometimes seen following pharmacological manipulations across these tasks.

Although the tasks discussed in this review are purported measures of risky choice, this may not be entirely the case. As discussed earlier in this review, Yang et al. (2018) state that the PDT does not adequately measure risky choice because the expected utility of each reinforcer differs from one another. Any changes in the discounting function following pharmacological manipulations do not necessarily reflect alterations in risky choice; instead, they may merely reflect changes in sensitivity to probabilistic reinforcement. The RDT and rGT have a similar limitation in that these tasks do not equate expected utility across each reinforcer alternative. Unlike the PDT, in which the expected utility of the large magnitude reinforcer decreases across the experimental session, expected utility of each reinforcer remains constant throughout the session.

The finding that drugs can produce opposite effects in PDT experiments using ascending/descending schedules suggest that this task measures behavioral flexibility as opposed to risky choice per se (i.e., drugs such as amphetamine, MK-801, and Ro 63–1908 may just increase perseverative responding). As discussed above, St Onge et al. (2010) argue that drugs such as amphetamine increase perseveration for the perceived relative value of the large, probabilistic reinforcer as opposed to perseveration for a particular lever. Even this explanation does not rule out the possibility that the PDT is measuring a construct other than risky choice. Instead, this task may measure how sensitive animals are to positive/negative contrasts. For example, when the probability of obtaining the large magnitude reinforcer decreases across the session, this is a negative contrast, whereas animals experience a positive contrast when the probabilities of obtaining reinforcement increase across the session. Typically, when opposite drug effects are observed in ascending/descending schedules, preference for the large, risky option increases when the probabilities decrease, whereas preference decreases when the probabilities decrease (St Onge et al., 2010; Yates et al., 2016; Yates et al., 2018). One potential explanation for the opposite effects observed across schedules is that drugs such as amphetamine/MK-801 attenuate contrast effects in animals. One interesting finding I have observed in my research is that the glutamatergic drugs MK-801 and Ro 63–1908 produce the same shifts in probability discounting as they do in delay discounting (a purported measure of impulsive choice). This finding is interesting because upward shifts in probability discounting are interpreted as reflecting increased risky choice, whereas upward shifts in delay discounting are interpreted as decreased impulsivity. Like probability discounting, delay discounting could simply measure positive/negative contrast (e.g., when the delay to the large magnitude reinforcer increases across the session, this can be considered to be an example of a negative contrast). This phenomenon has been observed outside of my laboratory. Similar to the effects observed in the PDT, amphetamine increases preference for a large, delayed reinforcer when the delay to reinforcement increases across the session, but decreases preference when the delays decrease (Tanno, Maguire, Henson, & France, 2014).

The assumption that probability and delay discounting reflect the same underlying process is not entirely plausible. As explained by Green and Myerson (2004), reinforcer amount has opposite effects on the discounting rate of delayed and probabilistic reinforcers (Christensen, Parker, Silberberg, & Hursh, 1998; Green, Myerson, & Ostaszewski, 1999), and other variables such as inflation affect delay discounting without affecting probability discounting (Ostaszewski, Green, & Myerson, 1998). Within behavioral pharmacology experiments, my laboratory has not always shown consistent drug effects across the PDT and delay discounting. In particular, the drug ifenprodil impairs discriminability of the small and large magnitude reinforcers in a delay discounting procedure in which the delay to reinforcement increases across the session (Yates, Gunkel, Rogers, Hughes, & Prior, 2017). However, when the probability of earning the large magnitude reinforcer decreases in the PDT, ifenprodil does not alter choice (Yates et al., 2016).

Like the PDT, the RDT may measure behavioral flexibility. Although baseline performance is similar between subjects trained on ascending and descending schedules (Simon et al., 2009), there is some evidence that probability presentation order can modulate drug effects in this task. When the probability of receiving shock decreases across the session, intra-mPFC amphetamine infusions decrease preference for the large magnitude reinforcer, even when the probability of receiving shock is 0. However, amphetamine infusions do not alter choice when the probability of receiving shock increases across the session (Orsini et al., 2018). Importantly, the lack of effect observed in the ascending schedule is not likely due to floor or ceiling effects. When the mPFC is inactivated via muscimol/baclofen infusions, opposite effects on choice are observed across ascending/descending schedules. As such, Orsini et al. (2018) argue that the mPFC may not necessarily be involved in risky choice, but instead, may be more important for modifying choice in response to shifting contingencies. More work is needed to determine if probability presentation order modulates drug effects in the RDT.

Final Recommendations

Considering drug effects are not always consistent across the major tasks of risky choice, one recommendation is to try to compare how pharmacological manipulations alter performance in more than one task (e.g., compare PDT to RDT). As already discussed in Yates (2018), studies commonly compare two different behavioral measures of impulsive behavior (e.g., impulsive choice tasks [such as delay discounting] vs. motor impulsivity tasks [such as the five-choice serial reaction time task]; Baarendse & Vanderschuren 2012; Broos et al., 2012; Hellemans, Nobrega, & Olmstead, 2005; Higgins et al., 2016; Higgins et al., 2018; Isherwood, Pekcec, Nicholson, Robbins, & Dalley, 2015; Isherwood, Robbins, Nicholson, Dalley, & Pekcec, 2017; Korte et al., 2017; Liu, Wilkinson, & Robbins, 2017; Mendez et al., 2012; Paine, Drigenberg, & Olmstead, 2003; Paterson, Wetzler, Hackett, & Hanania, 2012; Robinson et al., 2008; Schippers, Schetters, De Vries, & Pattij, 2016; Schneider et al., 2011; Sukhotina et al., 2008; Sun, Cocker, Zeeb, & Winstanley, 2012; Talpos, Wilkinson, & Robbins, 2006; Winstanley, Dalley, Theobald, & Robbins, 2004; Winstanley et al., 2007; Wiskerke, Stoop, Schetters, Schoffelmeer, & Pattij, 2011); however, comparisons between two risky choice tasks has not been done to my knowledge. This is important because comparing drug effects across tasks is difficult when these experiments are conducted at different time points in different laboratories using different strains of rodents. By conducting an experiment in which two risky choice tasks are tested in the same laboratory at the same time using the same experimenter(s) can better control for these idiosyncrasies.

Although comparing the effects of a drug in two (or more) risky choice tasks can help us further our understanding of why discrepant drug effects occur across studies (e.g., amphetamine produces opposite effects in the PDT and in the RDT), this is not always feasible due to the large number of subjects required. If using a single task to measure risky choice, investigators should consider some of the issues raised in this review. If using the PDT or the RDT, try to incorporate a supplemental analysis (AUC; parameter estimates derived via NLME; multilevel logistic regression) in addition to analyzing the raw proportion of responses for the large magnitude reinforcer. If the drug does not appear to alter the discriminability of the large and small magnitude reinforcers, using AUC should be sufficient (which is what our laboratory did in the Yates et al., in press study). However, if the proportion of responses for the large magnitude reinforcer appears to change when its delivery is guaranteed, generating parameter estimates via NLME or applying multilevel logistic regression will better describe how drugs alter choice. When using the rGT, an additional control group that controls for reinforcer magnitude needs to be incorporated, and multinomial logistic regression should be used to analyze responses for each alternative. Future studies should also further consider how modifying these procedures (e.g., using ascending/descending schedules, altering the response requirement) can modulate drug effects. Alternatively, randomizing the order in which each probability is presented can control for the probability presentation order effects that have been observed in the literature. By taking these suggestions into account, we may enhance our understanding of the neurochemical basis of risky decision making.

Public Significance Statement:

Risky choice is associated with several psychiatric disorders. As such, elucidating the neurochemical systems involved in risky choice can help lead to better treatment options for those with disorders characterized by excessive risk. Several tasks have been used to measure risky choice in animals, but drug effects across each task have been somewhat inconsistent. Understanding the limitations of each task is important for designing studies that better enhance our knowledge of the underlying neurobiology of risky choice.