The effect of conditioned inhibitors and preexposed cues on the outcome-specific Pavlovian-to-instrumental transfer effect in humans

Method

Results

The Pavlovian and instrumental phases were completed uneventfully. The mean number of R₁ responses and O₁ presentations were 269.9 (SEM = 14.4) and 51.8 (SEM = 2.9), respectively, and for R₂ and O₂ 251.2 (SEM = 13.4) and 54.4 (SEM = 2.9), respectively. The mean proportion of correct responses in the inhibitory training (0 = no correct answers, 1 = all correct answers), for blocks 1 and 2, respectively, were 0.58–0.58 (SEM = 0.05–0.06) for S₁A/S₂B; 0.55–0.76 (SEM = 0.04–0.05) for S₁I₁/S₂I₂; 0.82–0.92 (SEM = 0.05–0.04) for C₁C₂; 0.63–0.87 (SEM = 0.03–0.03) for T₁/T₂; 0.69–0.87 (SEM = 0.05–0.05) for I₁/I₂; and 0.65–0.9 (SEM = 0.05–0.05) for C₁/C₂. An ANOVA with Block (1, 2) and Cue (S₁A/S₂B, S₁I₁/S₂I₂, C₁C₂, T₁/T₂, I₁/I₂, C₁/C₂) showed a significant main effect of Block, F(1, 25) = 61.18, MSe = .03, p < .001, significant main effect of Cue, F(5, 125) = 10.06, MSe = .05, p < .001, and a significant Block × Cue interaction, F(5, 125) = 4.73, MSe = .024, p = .001. The analysis of the interaction confirmed more correct responses in Block 2 relative to Block 1 for compounds S₁I₁/S₂I₂ (p < .001), C₁C₂ (p = .001), T₁/T₂ (p < .001), I₁/I₂ (p < .001), C₁/C₂ (p < .001), but not for S₁A/S₂B (p = .86). In view of the fact that performance on S₁A/S₂B trials did not increase over blocks, we further explored the incorrect responses to these compounds and found that in 83% of the cases participants chose the option “nothing” (rather than the alternative outcome). This was likely caused by S₁ and S₂ also being paired with the absence of the outcomes when presented with the inhibitors, rendering them potentially ambiguous cues.

The rating scores in the summation test to T₁ and T₂, according to whether they were paired with the corresponding CI (T₁l₁ and T₂l₂), the preexposed control cue (T₁C₁ and T₂C₂), the novel control cue (T₁N₁ and T₂N₂), or the alternative inhibitor (T₁l₂ and T₂l₁), are plotted in Figure 1. An ANOVA with Outcome (O₁, O₂) and Cue (Inhibitor, Preexposed, Novel, Alternative) showed a significant main effect of Cue, F(3, 75) = 10.9, MSe = 607.6, p < .001, η p 2 = . 30 . Nothing else was significant, largest F(3, 75) = 1.3, MSe = 184.1, p = .28 for the Outcome × Cue interaction. Post hoc comparisons confirmed higher rating scores to the novel compounds than to the inhibitory, p < .01, preexposed, p < .01, and alternative compounds, p < .001. No other difference was statistically significant. The analysis confirmed that participants rated the likelihood of outcome presentation to be as low during the preexposed cues as during the Pavlovian inhibitors, and lower in the presence of all these compounds than the novel stimuli (Inhibitor = Alternative = Preexposed < Novel); in addition, ratings were unaffected by whether the CI was trained with the same outcome as the transfer cue or the alternative outcome (Inhibitor = Alternative), showing that, just as in Alarcón and Bonardi (2016), the inhibitors were not outcome-specific by this measure.

OPEN IN VIEWER

Figure 1.

Mean rating scores (participants’ expectancies) ± SEM to the questions about O₁ and O₂ for the transfer cues in compound with the inhibitors (T₁l₁, T₂l₂), the preexposed control cues (T₁C₁, T₂C₂), the novel control cues (T₁N₁, T₂N₂), and the alternative inhibitors (T₁l₂, T₂l₁) in the summation test.

The results of the PIT test are presented in Figure 2, which shows that, as expected, the transfer cues (T₁ and T₂) elicited more Same than Different responses when presented with the novel control stimuli (T₁N₁ and T₂N₂). This outcome-specific PIT effect appeared equally substantial in the preexposed stimulus compounds T₁C₁ and T₂C₂, but was reduced with the inhibitory compounds (T₁I₁ and T₂I₂). ANOVA with Outcome (Same, Different) and Compound (Preexposed, Inhibitory, Novel) as factors showed a significant main effect of Outcome, F(1, 25) = 9.7, MSe = 15,389.1, p < .01, η p 2 = . 28 and a significant Outcome × Compound interaction, F(2, 50) = 3.27, MSe = 4,936.2, p = .046, η p 2 = . 12 . The effect of Compound was not significant, F < 1. Exploration of the interaction confirmed higher Same than Different responses for the Preexposed, F(2, 50) = 13.92, MSe = 4,936.2, p < .001, and Novel compounds, F(2, 50) = 21.52, MSe = 4,936.2, p < .001, but not for the Inhibitory compounds, F(1, 50) = 1.33, MSe = 4,936.2, p = .27. Equivalent analysis of preCS responding revealed nothing significant, largest F(2, 50) = 1.13, MSe = 5,327.89, p = .33 for the Outcome × Compound interaction. For the preexposed (T₁C₁, T₂C₂), inhibitory (T₁I₁, T₂I₂), and novel (T₁N₁, T₂N₂) compounds, respectively, the mean preCS Same responses were 185.2 (SEM = 31.5), 192.9 (SEM = 33.1), and 173.8 (SEM = 30.0), and the mean Different responses 176.5 (SEM = 28.6), 167.9 (SEM = 30.3), and 191.5 (SEM = 33.9).

OPEN IN VIEWER

Figure 2.

Same and different responding to the transfer cues in compound with the preexposed control stimuli (T₁C₁/T₂C₂), the inhibitors (T₁I₁/T₂I₂), and novel control cues (T₁N₁/T₂N₂) in the PIT test.

Discussion

This experiment aimed to establish whether, in our procedure, preexposed stimuli would behave like CIs in summation and PIT tests. When compounded with the novel control stimuli N₁ and N₂, the transfer cues T₁ and T₂ elicited expectation of their respective outcomes; moreover, this effect was significantly reduced when they were compounded with the inhibitors l₁ and l₂, and also the preexposed control stimuli C₁ and C₂. This indicates that preexposure of C₁ and C₂ in training had made them as inhibitory as “true” Pavlovian inhibitors. Karazinov and Boakes (2004) also found that preexposed cues were more inhibitory than novel cues in a summation test, but—in contrast to our findings—less so than Pavlovian inhibitors, a difference we also observed in previous work (Alarcón & Bonardi, 2016). Karazinov and Boakes speculated that the difference between the Pavlovian and differential inhibitors stemmed from differences in how they were trained. While the Pavlovian CI was trained both as a single cue and in a compound, the preexposed differential CI was only trained alone. Seeing the preexposed cue in a compound for the first time at test might have produced uncertainty in the participants, making their ratings to this compound more conservative than to the other cues (the ratings to this compound were closer to the centre of a rating scale marked as “Don’t know”). However, in both Alarcón and Bonardi (2016) and the experiment reported here the Pavlovian CIs and the preexposed stimuli were presented the same number of times as single cues and compounds across the task, making it unlikely this factor was responsible. One potentially critical difference between our two studies is that Alarcón and Bonardi (2016)—which was primarily aimed to explore the effect of Pavlovian inhibition on PIT—excluded participants who had not learned the Pavlovian inhibitory discrimination in training. It is possible that considering only the participants that showed inhibition in the summation test might have increased the differences between the preexposed control and the inhibitor in the PIT test; however, we reanalysed the data of the present experiment using the same response criterion as Alarcón and Bonardi (2016), but still found no significant differences between these cues.

In summary, although the summation technique indicates that preexposure produces conditioned inhibition, it is unclear whether or not this is as strong as that produced by Pavlovian training. Moreover—in view of the failure of Karazinov and Boakes to manipulate the degree of differential inhibition as theory would predict—it is also worth considering alternative explanations. For example, perhaps the preexposed cue is becoming associated with the idea that “nothing” is going to happen, and expectation of this neutral event interferes with expectation of the outcome, reducing response ratings. This could have a similar effect in the summation test as the active expectation of outcome omission produced by the Pavlovian inhibitor (cf. Alarcón & Bonardi, 2016).

Consistent with this suggestion, although the results of the summation test suggested the preexposed cues acquired inhibitory properties, the PIT test did not. Here, the selective elevation of instrumental responding produced by T₁ and T₂—outcome-specific PIT—was eliminated only by the CIs, and by neither the preexposed or novel cues, which did not differ (cf. Alarcón & Bonardi, 2016). These results suggest that—in contrast to the results of the summation test—the preexposed cues were not inhibitory. This discrepancy between summation and PIT tests may be due to their differential sensitivity in detecting, perhaps subtle, differences in inhibition. Moreover, our sample size may not have been large enough to detect small differences between the inhibitors and the preexposed controls in the summation test, and between the preexposed controls and the novel cues in the PIT test. For instance, a larger sample size might have revealed that the preexposed controls were inhibitory relative to the novel cues, but less so than the Pavlovian inhibitors, that is, inhibitors < preexposed < novel. Alternatively, it may be that the nature of the inhibitory properties endowed by Pavlovian inhibition training and nonreinforced preexposure is qualitatively different and thus manifests differently depending on how it is tested. This would suggest that simple preexposure does not result in true inhibition, summation tests are sensitive to factors other than pure Pavlovian inhibition, and the PIT test is a purer measure of this property. These alternative explanations make it necessary to continue studying the effect of conditioned inhibition on PIT.

Another feature of Alarcón and Bonardi’s (2016) results was that the inhibitor for O₁ reduced summation and PIT effects just as much for T₂ as for T₁, suggesting that its inhibition was not outcome-specific. We suggested that this might be because only one CI-no US relationship was trained, and so in this study we trained two inhibitors, one for each of the outcomes. However, this had no effect on the result: in the summation test, there was no difference in expectation of the outcomes during the inhibitory compounds, regardless of whether the inhibitor predicted omission of the outcome predicted by the excitor with which it was paired, or the alternative outcome. In fact, there is already some controversy over whether CIs are specific to the sensory properties of the outcomes whose omission they signal, or whether they act via a solely motivational mechanism. Initial studies suggested that CIs trained with a specific US, such as a shock to the left eye, would inhibit responding equally to any CS signalling shock, regardless of the eye to which it was delivered (Nieto, 1984; Pearce et al., 1981; see LoLordo & Fairless, 1985, for a review). In contrast, more recent research, using PIT measures of inhibition, suggests a different story. Quail, Laurent, and Balleine (2017; see also Laurent, Wong, & Balleine, 2015) used a human task to train two inhibitors, one for each of two outcomes (S₁→O₁, S₂→O₂, S₁ l₁→nothing, S₂ l₂→nothing). The effects of l₁ and l₂ were evaluated in a PIT test, both when presented alone, when paired with the CSs with which they were trained (S₁l₁, S₂l₂)—and with the alternative CSs (S₁l₂, S₂l₁). If l₁, for example, were a specific inhibitor for O₁, it should reduce the PIT effect produced by S₁—which predicts the precise outcome whose omission l₁ signalled—but not that produced by S₂, which predicted the alternative outcome. This is what the authors observed.

Unfortunately, there were many differences between this study that demonstrated outcome-specific inhibition and our experiments which did not. One is that we evaluated learning in each trial of the Pavlovian conditioning phase by asking participants to press a key to predict the outcome to be delivered. Because of this, participants had to choose the option “Nothing” for the CIs and the control cues. It is possible that this consequence, being common to both inhibitors, hindered the development of outcome-specific inhibition. Quail et al. (2017), in contrast, evaluated learning at the end of the Pavlovian conditioning phase, and participants could answer if the inhibitor was associated with the absence of a particular outcome, for example, I₁ followed by no food, or with the general absence of the outcomes, for example, I₁ followed by no foods or drinks, which might have encouraged a discrimination between the CIs. In addition, Quail et al. separated their participants based on these answers (Experiment 2), finding that the “specific learners,” who identified specific inhibitor→no outcome relationships, showed high levels of outcome-specificity in the PIT test, whereas the “general learners,” those who identified the inhibitors as signalling the absence of any outcome, did not show any signs of such specificity. It is possible that we might show a similar effect by separating the participants similarly; however, our measures of Pavlovian learning do not allow us to divide our participants into specific or general learners.

Another difference is that Quail et al. did not evaluate their inhibitors’ properties against separately trained test excitors, and only used a PIT test (Rescorla, 1969). Conversely, our first experiment that demonstrated non-outcome-specific inhibition in the PIT test only trained one inhibitor, while in the present experiment we did not evaluate the alternative compounds in the PIT test. Our evidence for non-outcome-specific inhibition came only from a summation test, which Quail et al. did not include. Thus, there is no real conflict of results here—but a key need to bridge the gap between these various studies in order to establish whether or not inhibition is outcome-specific in these kinds of tasks.

Another point worth mentioning is the use of a brief instrumental retraining before the PIT test, the purpose of which was to serve as a reminder of the instrumental associations learned at the beginning of the task (see Alarcón & Delamater, 2019, for the same procedure in rodent research). It might be argued that this is similar to an “overtraining” procedure, resulting in habitual responding rather than goal-directed behaviour, which is less dependent on the outcome representation. However, we do not consider this training to be sufficient to produce habits. There is one example in which habits have been achieved by overtraining instrumental responses in humans, but it required several separate sessions of instrumental conditioning (Tricomi, Balleine, & O’Doherty, 2009), and recently a series of failed attempts to reproduce those findings have been published (de Wit et al., 2018). Related to overtraining, Garofalo and Robbins (2017) provided evidence that extended training in humans enhances the outcome-specific PIT effect and it would be interesting to assess if the ability of the CIs to reduce selective PIT remains after instrumental overtraining.

A final consideration is that, although it was designed to be similar to the tasks used in animal research, we did not obtain a direct measure of Pavlovian conditioning but rather asked participants about explicit expectations of the outcomes, as in contingency learning tasks (e.g., Shanks, 2007). Some researchers have argued that the processes involved in this type of task depend on the generation of propositions about the relationships between the events, rather than the traditional associative view, and there have been recent efforts to explain outcome-specific PIT with a propositional account (e.g., Hogarth et al., 2014; Seabrooke, Hogarth, & Mitchell, 2016). However, the outcome-specific PIT effect is consistently found in both animal and human studies, suggesting that the measure of Pavlovian conditioning might not be critical. For instance, the inhibition of PIT results found by Quail et al. (2017) was remarkably similar to previous results found in rodents by the same group of researchers (Laurent et al., 2015). Nevertheless, adding direct measures of Pavlovian conditioning might provide further information about the underlying processes in this type of task.

In summary, we have demonstrated that simple preexposure can make a stimulus behave like a CI in a summation test, relative to a novel control; however, this was not evident in a PIT test of inhibition. Of several possible interpretations, one possibility is that summation tests are sensitive to effects other than Pavlovian inhibition and that the PIT test is a better measure of a stimulus’ inhibitory properties. We also failed to produce evidence of outcome-specific inhibition in a summation test, despite training two inhibitors, one for each outcome. Although this ostensibly contradicts findings from other studies (e.g., Quail et al., 2017), a more likely explanation is that this discrepancy arises from procedural differences. More generally, we believe that these findings expand our understanding of the mechanisms underlying PIT and that this task might be helpful to study the effect of conditioned inhibition in maladaptive behaviour, such as drug-seeking behaviours and overeating.