Following operational second-order conditioning (SOC; i.e., S1–US pairings followed by S2–S1 pairings, where S1 is a first-order conditioned stimulus, S2 a second-order conditioned stimulus, and US is an unconditioned stimulus), the second-order (S2) stimulus often acquires the potential to produce a conditioned response (CR) although it has never directly been paired with the US (Pavlov, 1927). Two alternative explanations of SOC in terms of associative linkage can be found in the literature (e.g., Barnet, Arnold, & Miller, 1991; Rizley & Rescorla, 1972). The first of these is that at test S2 elicits the CR through an associative chain that includes S1 (i.e., S2→S1→US→CR, or S2→S1→CR).Footnote 1 Both of these associative structures are variants of the associative-chain account, which assumes a role for the representation of S1 at test of S2 (e.g., Hall, 1996). The second type of account of SOC assumes that activation of the US representation by S1 during the Phase 2 S2→S1 pairings effectively creates direct pairings between S2 and the reactivated representation of the US, resulting a direct association between S2 and the US (i.e., S2→US; Konorski, 1967). Thus, the direct association account of SOC differs from the associative-chain account in its omission of any role for the representation of S1 at the time at which S2 is tested.

The conventional test to differentiate between the associative-chain account and the direct association account of SOC involves assessment of the consequences of extinction of S1 (following the S2–S1 pairings) on responding to S2. A reduction in responding to S2 as a result of extinction of S1 is conventionally viewed as consistent with the associative-chain account of SOC in which presentation of S2 at test activates the representation of extinguished S1 (see Fig. 1). However, most published studies in which S2 and S1 were sequentially presented during S2–S1 pairing failed to report a decrease in responding to S2 following subsequent extinction of S1 (e.g., Rizley & Rescorla, 1972; but see Molet, Miguez, Cham, & Miller, 2012, which found evidence for the associative-chain view). Results similar to those of Rizley and Rescorla were observed in a human causal learning study (Jara, Vila, & Maldonado, 2006). Consequently, the prevailing view is that, given sequential S2–S1 pairings in Phase 2, responding to S2 depends on a direct association that was established during Phase 2 of SOC treatment between S2 and either the response evoked by the US or the representation of the US (i.e., an S2→CR or S2→US→CR associative sequence, or both as was suggested by Polack, Molet, Miguez, & Miller, 2013).

Fig. 1
figure 1

Hypothetical associative structures underlying responding to S2 at test, following SOC (1a and 2a) and following post-SOC extinction of S1 (1b and 2b), according to the associative-chain account (1a and 1b) and the direct association account (2a and 2c), respectively

Full size image

In contrast to the conventional absence in reduction of responding to S2 following post-SOC extinction of S1, Rescorla (1982) reported that when the S2–S1 pairings had been simultaneous during Phase 2, subsequent extinction of S1 did reduce responding to S2. He concluded that with simultaneous S2–S1 pairings, the representation of S1 played a role at test of S2. Thus, Rescorla concluded that simultaneous and sequential S2–S1 pairings resulted in different associative structures for the resultant SOC. Thus, Rescorla suggests that in Fig. 1a represents the structure underlying simultaneous SOC, whereas 2a represents the associative structure of sequential SOC.

There is, however, an alternative account of why, following simultaneous S2–S1 pairings, extinction of S1 decreases responding to S2. That is, with simultaneous S2–S1 pairings, S1 during Phase 2 may mediate formation of a direct S2–US association based on S2 being present and S1 activating a representation of the US with which it had been paired during Phase 1. Then, during Phase 3 (i.e., extinction of S1), presentation of S1 may evoke the representation of S2 in the absence of US, which could result in extinction of the evoked representation of S2 and reduce responding to S2 at test. Note that within this account of why extinguishing S1 reduces responding to S2 is the assumption that extinction of an invoked representation also reduces responding to the actual cue at test. We call this explanation of why extinction of S1 decreases responding to S2 the mediated-extinction hypothesis.

The mediated-extinction hypothesis is consistent with the direct association account of SOC in that the direct association view assumes that responding to S2 depends on a direct association between S2 and the evoked representation of the US. Furthermore, it contrasts with the associative-chain account of SOC, in which S2 at test activates a representation of S1 and consequently S1’s current response potential. Extinction of the evoked representation of S2 may have occurred either in addition to or in the absence of S1 playing a contributing role in an associative chain at test of S2. That is, direct associations may contribute to SOC in parallel with an associative chain or be the sole basis of SOC.

Presumably, responding to S2 is not subject to appreciable extinction as a result of nonreinforced evocation of S2 in the case of prior sequential S2–S1 pairings because S1 does not strongly evoke the representation of S2 during S1’s extinction. This is because, in the sequential case, S1 would have to activate the representation of S2 through a backward association. The possibility of an absent cue entering into the same type of association as its companion cue when the companion cue is presented alone was first proposed by Holland (1981, 1983; Holland & Forbes, 1982) and elaborated on by Hall (1996).Footnote 2

The central point of what has been discussed above is that the presence or absence of decreased responding to S2 when S1 is extinguished does not necessarily reveal the associative structure underlying SOC (i.e., direct S2–US association or S2–S1–US associative chain). Some sequential SOC studies did observe decreased responding to S2 as a result of S1 being extinguished (e.g., Molet et al., 2012) while others did not (e.g., Rizley & Rescorla, 1972). In contrast, simultaneous S2–S1 pairings seem to consistently yield decreased responding to S2 (e.g., Rescorla, 1982). The presence or absence of decreased responding to S2 could be related to S1’s potential during its extinction to activate a representation of S2 and affect direct extinction of the evoked representation of S2. The present experiments sought to investigate the potential role of mediated extinction in reducing responding to S2. Experiment 1 used sequential S2–S1 pairings with the expectation that extinction of S1 would not reduce responding to S2. If sequential SOC is less sensitive to post-SOC extinction of S1 because of the greater difficulty for S1 to backward evoke the representation of S2 during extinction, then extinction of S2, which has a forward association with S1, should produce pronounced mediated extinction of S1. Hence, in the sequential case the mediated-extinction account anticipates extinction of the evoked representation of S1 (i.e., decreased responding of S1) as a result of extinguishing S2. However, little or no decrease of responding to S2 as a result of S1’s extinction is expected if S1’s backward association prevents effective mediated extinction of S2. Thus, the two theoretical accounts make different predictions with regard to responding to S1 when S2 is extinguished.

Experiment 1

Experiment 1 compared the influence of extinguishing S1 on responding to S2 (relative to its appropriate control condition) with the influence of extinguishing S2 on responding to S1 (relative to its appropriate control condition) in a SOC preparation in which sequential S2–S1 pairings occurred during Phase 2. Of central interest was whether the evoked representation of S1 (forward activated by S2 presentations during extinction of S2) along with the absence of the US would decrease responding to S1 (i.e., the first-order cue) as well as decrease responding to S2 (i.e., the extinguished second-order cue). If this did not occur, it would be highly improbable that extinction of S1 would result in pairings of the evoked representation of S2 (backward activated by S1 presentations during extinction of S1) with the absence of the US leading to extinction of S2.

Method

Participants

Twenty-four University of Lille (France) male and female students (N = 24), 18–29 years of age, volunteered.

Apparatus and materials

Individual cubicles were used to provide a quiet space for participants. The program was written using Inquisit. Seven white digits served as the S1 cues, and seven faces (512 × 383 pixels) portraying neutral affective states (Minear & Park, 2004) served as the S2 cues. Which digits and which faces served in each condition of the experiment were randomly determined for each participant. All cues were presented on a light gray background of a computer screen. The digits (Helvetica 150) were presented in the middle of the screen and the faces on the left side of the screen during the entire experiment. The US (hereafter referred to as an outcome, given its dubious biological significance) was the word WIN (Helvetica 80) in red letters that always appeared on the right side of the screen. All stimuli were presented equidistant from the top and bottom of the screen.

Procedure

Throughout the experiment, instructions were provided to the participants on the computer screen. Participants were first asked to learn which S1s (i.e., digits) were followed by the WIN outcome and which were not. They were also instructed to predict the outcome on each trial by pressing on a “WIN” or an “empty screen” key (left and right cursor keys, counterbalanced). The S1s remained on the screen until participants responded. Each condition started with Phase 1 in which the seven digits were randomly assigned to the seven S1 roles (S1SOC, S1S1ext, S1S2ext, S1Filler1, S1Filler2, S1Filler3, S1Novel; see Table 1). Participants were then presented with three successive blocks of the first six of these seven S1s (S1Novel was only presented at test). Each block contained one presentation of each of the six presented S1s with the order randomly determined anew for each block. Upon termination of each S1 presentation, that is, when the participant had made a prediction, there was a 1.5-s presentation of the WIN outcome on the right side of the screen following three of the S1s (S1SOC, S1S1ext, and S1S2ext), while the three others (i.e., fillers: S1Filler1, S1Filler2, and S1Filler3) were followed by the absence of the outcome (i.e., only the dark gray screen background was visible) for the same 1.5 s. The fillers were added to equate the number of reinforced cues and nonreinforced cues as well as to encourage discrimination learning as opposed to simply responding on all trials. The intertrial intervals (ITIs) were 2 s (so that the absent nonreinforcement outcome and the immediately subsequent ITI was simply a blank, 3.5-s, dark-gray screen).

Table 1 Design of Experiment 1
Full size table

Prior to Phase 2, participants were instructed on the screen that they would now see a face (i.e., S2) paired with each digit (i.e., S1), and that they should try to learn which face goes with which digit without making any prediction concerning the face that followed each digit. All participants were exposed once to each S2–S1 sequential pairing in random order. Each S2 appeared for 2 s on the left-hand side of the screen, and following termination of S2 its S1 associate was presented for 2 s in the center of the screen. After a 2-s ITI, the next S2 was presented.

Before Phase 3 began, all participants were instructed that they would have to learn new associations between a digit (i.e., S1) or a face (i.e., S2) and the outcome. Similar to Phase 1, they were instructed to predict, on each trial, what the outcome would be. Six within-subject conditions were constituted. In the SOC-S1 ext condition, operational extinction of S1S1ext occurred during Phase 3. In the SOC-S2 ext condition, operational extinction of S2S2ext occurred during Phase 3. In the SOC condition, there were no relevant trials during Phase 3. Of the three filler conditions in which S1 had been presented without an outcome in Phase 1, one of them (Filler 1) now had S1 paired with the outcome (S1Filler1–outcome). A second condition (Filler 3) had S2 paired with the outcome (S2Filler3–outcome). The remaining filler condition (Filler 2) had no new presentations of S1 or S2 during Phase 3. Fifteen blocks of each event occurred in Phase 3. Note that the fillers served to encourage learning about specific cues in Phases 1 and 3 and discourage the learning of nonspecific rules.

Finally, participants were instructed that, for each cue (digits and faces), they would have to respond by pressing the “WIN” key if they thought that the cue predicted the outcome, and the “empty screen” key if they thought that it predicted the absence of the outcome. They were instructed throughout the test to keep the index finger of their preferred hand positioned on the down arrow key except for when they were responding. Participants indicated their choice by pressing the left or right arrow keys (counterbalanced as “WIN” or “empty screen” across participants within conditions in each of the two groups). Because responding to a cue may influence subsequent responding to its associate, test order was recorded for all cues. Participants were not informed that their reaction times (RTs) were being recorded, but they were encouraged in the instructions on the screen to respond as soon as they knew their answer. Also, participants were not informed whether their response was correct or not (i.e., no feedback; specifically, they were told on the screen “This time there will be no feedback (i.e., no outcomes will be displayed even if they occurred.)”. Each of the six S1s and the six S2s seen by participants during training was tested plus S1Novel and S2Novel. Each test stimulus remained on the screen until the participant had responded; then “next trial” was displayed for 2 s. The 14 test cues were presented in random order.

The type of response (i.e., “WIN” or “empty screen”) on each test was recorded as well as the RT to respond. These two dependent variables were transformed into a single variable following Craddock, Molet, and Miller’s (2012) transformation rule. To normalize the RTs, each of the 14 test RTs was divided by the participant’s largest RT of the 14 (RTmax). Any “WIN” response RT was then transformed into [−ln (RT / RTmax)], and any “empty screen” response RT into [+ln (RT / RTmax)]. Hence, transformed data near zero (i.e., RT / RTmax near 1) denote hesitating responses, while positive or negative transformed data indicate that the participant answered “WIN” or “empty screen” without hesitation, respectively. Because a slight increase or decrease in a short RT is presumably more significant than an equivalent increase or decrease in a long RT, differences between short RTs were amplified relative to differences between long RTs through the use of a logarithmic transformation. This transformation also minimized the positive skew ordinarily seen in RTs and yielded data sets within each condition that were approximately normally distributed.

Our first test of the mediated-extinction hypothesis compared responding to S1S2ext and S1SOC. Mediated extinction would take the form of less responding to S1S2ext presumably because activation of the representation of S1S2ext without the outcome during extinction of S2S2ext during Phase 3 provides an opportunity for mediated extinction of S1S2ext. Importantly, activation of S1S2ext during extinction of S2S2ext in Phase 3 was expected to be strong due to the forward relationship of S2S2ext to S1S2ext during Phase 2. In contrast, extinction of S1S1ext should have had a smaller decremental effect on responding to S2S1ext, relative to S2SOC, because S1S1ext’s association with S2S1ext was backwards (i.e., in Phase 2, S1S1ext followed S2S1ext). In contrast to the mediated-extinction account, the associative-chain account anticipates decreased responding to S2 when S1 is extinguished because, at test, S2 evokes S1 and its extinguished response.

Results and discussion

Mean transformed RTs for each cue are depicted in Fig. 2. SOC was demonstrated by comparing scores for S2SOC to zero, t(23) = 2.18, p < .04, d = 0.40. Also, S2SOC yielded higher scores than novel S2Novel, t(23) = 3.75, p < .002, d = 0.76, and higher scores than S2Filler2, an S2 paired with an explicitly unpaired S1, t(23) = 2.69, p < .02, d = 0.55.

Fig. 2
figure 2

Mean transformed RT and standard error as a function of cue for Experiment 1. S1s were always digits. S2s were always faces. See Table 1 for the roles of the various S1s and S2s

Full size image

Contrary to what is often found in the literature (e.g., Rizley & Rescorla, 1972), the comparison of S2S1ext to S2SOC found that extinction of S1S1ext during Phase 3 decreased responding to S2S1ext , t(23) = 5.36, p < .001, d = 1.10, while extinction of S2S2ext had relatively little effect on responding to S1S2ext. S1S2ext yielded nonsignificantly lower scores than S1SOC, t(23) = 2.03, p > .05, d = 0.41. No ANOVA was performed to quantitatively compare these two differences because S2s were faces and second-order cues, whereas S1s were digits and first-order cues, which would confound any direct comparison between these two differences.

During Phases 2 and 3 for Condition Filler 1, S1Filler1 was paired with first S2Filler1 and then the outcome in a sequence similar to sensory preconditioning (SPC) of S2Filler1, which appears to have increased responding to S2Filler1 compared to S2Filler2, t(23) = 2.89, p < .01, d = 0.59. However, no evidence was found that reinforcing S2Filler3 in Phase 3 influenced responding to S1Filler3 in Condition Filler 3 (compared to S1Filler2), t(23) = 0.23, d = 0.05. Note that SPC in the former case was likely attenuated by latent inhibition arising from pre-exposing participants during Phase 1 to S1Filler1 alone prior to the pairings of S2Filler1 with S1Filler1.

Experiment 1 yielded results that differed from what was expected based on the majority of the existing literature concerning sequential SOC in that we observed reduced response strength of S2 as a result of post-SOC extinction of S1. However, we do note that similar results were obtained by Molet et al. (2012) with sequential S2–S1 pairings with rats. Interpretation of the effect of extinguishing S1S1ext on responding to S2S1ext in Condition SOC-S1 ext in terms of the formation of an association between the representation of S2 and the absence of outcome during Phase 3 (i.e., mediated extinction) is unlikely for two reasons. First, it would have to assume that S1 backward activated the representation of S2. Second, if backward activation of S1SOC is assumed, why was it stronger than forward activation of S1S2ext by S2S2ext in Condition SOC-S2 ext, given that forward pairings are widely seen to result in stronger behavioral control than backward pairings (e.g., Pavlov, 1927)? Modulations of responding to S2S2ext did not appreciably influence responding to S1S2ext. These observations cast doubt on the mediated-extinction explanation of how extinction of S1S1ext on responding to S2S1ext in Condition SOC-S1 ext.

The associative-chain account more readily explains the present results by assuming that responding to S1S2ext in Condition SOC-S2 ext at test reflected the level of S1S2ext–outcome association acquired during Phase 1 because backward activation of S2S2ext by S1S2ext at test was weak and hence did not engage S2S2ext’s association to “no outcome.” However, extinction of S1S1ext greatly influenced responding to S2S1ext at test, presumably because the test of S2S1ext forward activated S1S1ext’s representation.

Finally, the data provided no support for the direct association account of SOC according to which neither S2 nor S1 should be influenced by their respective companion’s extinction because what counts in that framework is only the association between the tested cue and the outcome. Importantly, in Experiment 1, assessment of the mediated-extinction account of S2’s dependency on extinction of S1 was indirect as it hinged on assumptions about directionality of associations being of differential effectiveness and on digits (i.e., S1s) and faces (i.e., S2s) having similar associabilities. Hence, a further test of the mediated-extinction hypothesis of how extinction of S1 might decrease responding to S2 seemed warranted.

Experiment 2

Experiment 2 further assessed whether SOC depends more on an S2–S1–US associative chain than any potential direct S2–US association with mediated extinction explaining the reduction in responding to S2 that is sometimes seen when S1 is extinguished. This was accomplished by presenting the S1-no outcome trials prior to the S2–S1 pairings, thus preventing any direct S2–outcome association from being modified post-SOC. In Phase 1, S1SOC1 was reinforced in Context A and nonreinforced in Context B, a distinctly different context (see Table 2). Training with S1SOC2 was identical to that of S1SOC1, but with Contexts A and B playing reversed roles. Critically, although excitation of these two S1s in Context A (or B) initially generalized to Context B (or A), over repeated training trials they were extinguished in Context B (or A) while their excitatory values were maintained in Context A (or B). Thus, all training of S1SOC1 and S1SOC2 occurred before S2SOC1 and S2SOC2 were ever paired with them. This precluded the evoked representations of S2SOC1 (S2SOC2) from being paired with the absence of the outcome during the nonreinforced S1SOC1 (S1SOC2) trials in Context B (A). Then, S2SOC1 and S2SOC2 were sequentially paired with S1SOC1 and S1SOC2, respectively, in a novel context (C) in which generalized conditioned expectation of the outcome was expected. Presenting S1SOC1 and S1SOC2 in Contexts A or B never evoked S2SOC1 or S2SOC2 because these S1s and S2s had not yet been paired with each other. Therefore, based on the mediated-extinction hypothesis, responses to S2SOC1 and S2SOC2 were expected to be equivalent in whichever context they were tested. However, if responses to S2SOC1 (S2SOC2) mimicked the expected context dependency of responses to S1SOC1 (S2SOC2), that is, renewal in Context A (C), the results would support an associative-chain account of SOC.

Table 2 Design of Experiment 2
Full size table

In Experiment 2, following conditioned discrimination training of S1SOC1 and S1SOC2 between Contexts A and B in Phase 1 and subsequent S2SOC1–S1SOC1 and S2SOC2–S1SOC2 pairings in Phase 2 in Context C, all four cues were tested in the contexts in which S1SOC1 and S1SOC2 had been differentially reinforced (A and B) and the context in which the S2–S1 pairings had subsequently occurred (C). Importantly, the S2s were paired with the S1s in a neutral context (C) just before testing; that is, there was no post-SOC training of S1SOC1, S1SOC2, S2SOC1, or S2SOC2. The mediated-extinction hypothesis, concerning the effects of post-SOC extinction of S1 on S2, as well as the direct association account of SOC upon which mediated extinction hypothesis is based, assumes that decrementing responding to S2 requires activation of the representation of S2 along with the absence of the outcome’s representation. As S2 could not have been evoked during the Phase 1 S1-no outcome presentations, in the direct S2–US framework the test context was expected to control responding to S1 (S1SOC1 and S1SOC2) but not to S2 (S2SOC1 and S2SOC2). However, if SOC was based on an S2–S1–US associative chain which depended on activation S1 at test, responses to S2 should mimic the context dependency of responding to that specific S1.

Method

Participants

Forty-eight university students (males and females; 18–29 years of age), from SUNY-Binghamton, volunteered. Their participation was approved by the local Institutional Review Board.

Apparatus

As in Experiment 1, S1s and S2s were represented by digits and faces, respectively, and the outcome was the same as in Experiment 1. The background screen was one of three color–texture dyads, different for each experimental context (i.e., A, B, and C.). The roles of the three color–texture dyads were randomly assigned across conditions independently for each participant.

Procedure

Conditioned discrimination training during Phase 1 consisting of eight trials/block of four S1 cues each, which were reinforced in one context and nonreinforced in a second context. Specifically, in Context A, S1SOC1 and S1 S2alone were reinforced while S1noS2 and S1SOC2 were nonreinforced. In Context B, S1noS2 and S1SOC2 were reinforced, whereas S1SOC1 and S1 S2alone were nonreinforced (see Table 2). Phase 1 was conducted as repeated cycles of the eight trial types depicted in Table 2, each type of trial occurred once in a random order for each subject, and the order was determined anew for each cycle. Phase 1 was continued for each participant until the participant’s response was correct on all eight trials for two consecutive cycles, at which time the participant was advanced to Phase 2. Any participant who failed to reach this retention criterion within 12 cycles of Phase 1 training was informed that the experiment was over and was dismissed.

Each trial of Phase 1 was preceded with 0.5 s of gray screen (ITI) that was devoid of any overt contextual features, 1.0 s of the current trial context (A or B), presentation of the cue (S1 or S2 with duration dependent on the participant’s response) in the current trial context, 1.5 s of presentation of the outcome or absence of the outcome in the current trial context, and 1.0 s further exposure to the current trial context. Text at the top of the screen reminded participants which key was to be used to predict either the presence or absence of the “WIN” outcome. A purple sticker on one of the keys represented the outcome-present key, while an orange sticker represented the outcome-absent key. When the “WIN” predicted response key was the left key (i.e., when the purple sticker was on the left key), S1 appeared in the middle and the “WIN” outcome appeared on the left. Conversely, when the “WIN” predicted response key is the right key, S1 appeared in the middle and “WIN” appeared on the right. The two keys were the horizontal cursor keys and their assignment was counterbalanced across subjects.

A Phase 2 cycle consisted of one each of three trial types: S2SOC1–S1SOC1 and S2SOC2–S1SOC2 (sequential pairings corresponding to, respectively, the SOC1 and SOC2 conditions), and S2S2alone presented by itself (i.e., the S2-alone condition), all of which occurred in a novel context (C). In order to minimize potential extinction for S1SOC1 and S1SOC2 during Phase 2, we limited Phase 2 to three cycles. If the “WIN” outcome appeared on the left side of the screen in Phase 1, S2 appeared on the opposite side (right side) for Phase 2. Conversely, I, if “WIN” appeared on the right side during Phase 1, S2 appeared on the left side in Phase 2. A Phase 2 trial in Context C included a 2.5-s ITI (an increase of the ITI for the S2–S1 pairings from Experiment 1, which was intended to minimize potential learning across adjacent trials), and 1.5-s presentation of each stimulus (sequentially when there was a pairing). A 2.5-s ITI preceded the first Phase 2 trial and followed the last Phase 2 trial. The order of the three trial types within a cycle was random and drawn without replacement.

Testing occurred once in each of the three contexts (A, B, and C) in one of the following six orders, ABC, BCA, CAB, ACB, BAC, and CBA, with the order of test context counterbalanced across subjects. All cues (including novel S1Novel and S2Novel cues) were tested once in each context with the test trials blocked by context and, within test block, the order of cues randomly assigned to each participant. After having completed the entire experiment, participants were presented with a debriefing screen and were excused.

Results and discussion

In order to have a consistent estimator of effect sizes, (\( d=\frac{T}{\sqrt{n}}\Big) \)was used to estimate both the effect size of comparisons between two means and the effect size of specific planned 2 × 2 interactions. In the latter case, d \( =\sqrt{\frac{F\ of interaction}{n}} \).

SOC was tested in Context C by comparing presumably neutral S2S2alone to S2SOC1 and S2SOC2 (see Fig. 3). S2SOC1 yielded higher scores than S2S2alone in Context C, t(47) = 4.57, p < .001, d = 0.66, and S2SOC2 yielded higher scores than S2S2alone in C, t(47) = 5.74, p < .001, d = 0.83. No significant difference between responses to S2SOC1 and S2 SOC2 was observed in Context C, t(47) = 1.51, p < .15, d = 0.22.

Fig. 3
figure 3

Mean transformed RT and standard error as a function of cue for Experiment 2. S1s were always digits. S2s were always faces. See Table 1 for the roles of the various S1s and S2s

Full size image

As previously mentioned, the prediction of the mediated-extinction hypothesis and the underlying S2–outcome direct association account of SOC is that there should be no difference in responding between S2SOC1 and S2SOC2 in either Context A or Context B. This is because, according to this hypothesis, second-order conditioned discrimination requires S2 or its evoked representation to undergo pairings with the absence of outcome in one context and not the other. Therefore, because no such post-SOC training occurred, S2SOC1 should have yielded responses similar to S2SOC2 regardless of the context in which testing occurred. Admittedly, when testing occurs in a different context from the one in which SOC was learned, inter-context generalization may yield weaker second-order CRs (Hall & Honey, 1990). However, in that case, both S2SOC1 and S2SOC2 should have been affected the same way when tested in Contexts A and B. That is, responding to S2SOC1 and S2SOC2 should generalize indifferently to Contexts A and B because the two cues to which they were paired in Context C (i.e., S1SOC1 and S1SOC2, respectively) had received the same amount of reinforcement and nonreinforcement during Phase 1. The results showed that responses to S2SOC1 and S2SOC2 changed in opposite directions between Context C and Contexts A and B. That is, there was a significant interaction between cues (S2SOC1 vs. S2SOC2) and contexts (C vs. A), F(1, 47) = 35.05, p < .001, d = 0.85. Additionally, there was an interaction between cues (S2 SOC1 vs. S2SOC2) and context (C vs. B), F(1, 47) = 5.41, p < .025, d = 0.35. Moreover, S2SOC1 yielded higher scores than S2SOC2 in Context A, t(47) = 7.06, p < .001, d = 1.02, whereas the reverse was observed in Context B in which S2SOC2 yielded higher scores than S2SOC1, t(47) = 4.69, p < .001, d = 0.68. Consistently, the interaction between context (A and B) and cues (S2SOC1 and S2SOC2) was significant, F(1, 47) = 47.08, p < .001, d = 0.99. As predicted, no appreciable differential effect of testing in Contexts A and B on responding to S2S2alone, which was never paired with any S1, was observed, t(47) = 0.10, p < .95, BF01 = 6.94 (where BF01 is the Bayesian factor is in favor of the null hypothesis). Because the same S2, whether it was S2SOC1 or S2SOC2, yielded high responding or weak responding dependent upon the test context, it appears that responding to S2SOC1 and S2SOC2 was controlled by the contextually modulated associative status of S1SOC1 and S1SOC2, respectively, at the time of testing. This is contrary to what the mediated-extinction hypothesis predicted but fully concordant with the associative-chain account of SOC.

Finally, as expected from training, S1SOC1 yielded higher scores in Context A than in Context B, t(47) = 8.48, p < .001, d = 1.20. The results were inverted for S1SOC2, which yielded higher scores in Context B than in Context A, t(47) = 9.68, p < .001, d = 1.39. Moreover, S2SOC1 mimicked S1SOC1 by yielding higher scores in Context A than in Context B, t(47) = 4.10, p < .001, d = 0.59, and S2SOC2 mimicked S1SOC2 by yielding higher scores in Context B than in Context A, t(47) = 8.58, p < .001, d = 1.24. These latter results, together with those presented earlier, lend support to the associative-chain account of SOC because S2 elicited responses reflecting those to S1.

General discussion

Experiment 1 was expected to demonstrate that SOC with sequential S2–S1 pairings depends on a direct S2–outcome (or S2–CR) association that is impervious to extinction of S1 following the S2–S1 SOC pairings. However, Experiment 1 found that sequential SOC was attenuated by extinction of S1. Consequently, we examined the associative structure of sequential SOC using the data from Experiment 1 and then designed Experiment 2 as a more direct test of the mechanism by which extinction of S1 decreased (sequential) SOC.

Specifically, Experiment 1 tested the respective influence of extinguishing S1 or S2 (after they had been sequentially paired) on responding to the companion cue, S2 or S1, respectively. As we mentioned in the introduction, the learning of a direct association between S2 and the outcome (or CR) had previously been suggested to account for the absence of a decrease in the response to S2 when S1 was extinguished following sequential SOC. Assuming a direct S2–outcome actually underlies SOC when S2 and S1 are paired sequentially, our mediated-extinction hypothesis asserts that this hypothetical S2–outcome direct association survives extinction of S1 due to the backward S2←S1 association not inducing extinction of S2 during extinction of S1. Thus, we expected little decrement in responding to S2 when S1 was extinguished. In contrast, we anticipated a strong influence of S2’s extinction on responding to S1 because of the forward association from S2 to S1. That is, the forward S2→S1 association was expected to result in an effective pairing of the evoked representation of S1 and the absence of the outcome during extinction of S2, which could result in reduced responding to S1. The results did not support either of these expectations. Rather, extinction of S1 greatly reduced responding to S2 despite the assumed weak potential of S1 to activate S2’s representation during Phase 3. Additionally, extinction of S2 failed to appreciably reduce responding to S1. Given the forward relationship of S2 to S1, it is here that we had the highest expectation of observing mediated extinction, but it did not materialize. Thus, the results of Experiment 1 appear to support the associative-chain account of SOC.

Although the failure of extinction of S2 to reduce responding to S1 in Experiment 1 argues against the mediated-extinction hypothesis, the argument here is indirect because S2 and S1 were distinctly different cues functionally and physically (i.e., second-order faces and first-order digits, respectively). In contrast, Experiment 2 provided a more direct test of the mediated-extinction hypothesis and the direct association account of SOC more generally. The procedures used in Experiment 2 categorically precluded the possibility of an evoked S2 during extinction of S1 because the S2–S1 pairings occurred after extinction of S1. Moreover, Experiment 2 assessed responding to S2 as a function of the associative status of S1, which was manipulated by test Contexts A and B serving as discriminative stimuli for different response potentials of S1. Because the S2–S1 pairings were presented after S1’s discriminative training (i.e., excitatory conditioning in Context A and extinction in Context B), any decrease in responding to S2 in Context B relative to Context A (and Context C) would have to be attributed to activation of S1 at test. The results here, too, were consistent with the associative-chain account of SOC. Importantly, they showed that responses to the representation of S1 evoked by S2 were determined by the effective relationship that S1 had with the outcome in each test context. Alternatively stated, the S1 representation evoked by S2 predicted the outcome if and only if the actual S1 predicted the outcome in the test context. Notably, the reverse does not seem to be true. That is, Experiment 1 showed that potentially extinguishing the evoked S1 by extinguishing S2 had little to no effect on responding to S1.

It should be pointed out that preservation of responding to S2 following extinction of S1 that has been observed by some researchers in sequential SOC preparations (in contrast to what we observed in the present experiments) does not necessarily preclude SOC depending on an S2–S1–US associative chain at test. Because an extinguished response is often renewed when the cue is tested in a different context from the one in which it was extinguished (Bouton & Bolles, 1979; Bouton & King, 1983), it is possible that responding to S2 in those experiments was enabled by activation of a renewed S1–outcome association. In this situation, responding to S2 at test may have been the result of an evoked S1 representation, the response potential of which was renewed because the context of S1’s evocation during a test of S2 (i.e., physical S2) differed from the context in which S1 was extinguished (i.e., absence of physical S2). Hence, the context of S1’s extinction is the absence of S2, while the context of S1’s evocation at test of S2 is S2. Therefore, the sometimes observed strong response to S2 outside the context of nonreinforcement of S1 may reflect a mediated renewed response (i.e., S2→S1→renewed CR). This alternative explanation of why extinction of S1 sometimes fails to reduce responding to S2 (i.e., an associative chain involving a renewed S1 response cannot account for the results we observed in Experiment 1 because the test of S2S1ext yielded low scores (i.e., no mediated renewal was observed). But it can possibly explain the absence of modulation of S2’s response reported by several authors, and it highlights the potential role of the context in controlling acquired responses.

In the associative-chain framework, presentations of S2 without S1 during extinction of S2 weakened the S2–S1 association necessary to observed responding to S2 that depended on activation of S1 at test. Thus, in Experiment 1, when S2 was extinguished, the test of S2 was expected to elicit a distinctly weaker response than the one S1 elicited (consistent with S2S2ext yielding negative scores and S1S2ext yielding positive scores) because of weakened evocation of S1 by the extinguished S2. In Experiment 2, learning that S1 was followed by no outcome in one context did not eradicate S1’s potential to evoke the outcome representation in a different context. Thus, in the associative-chain framework, strong responding was expected to S1 and consequently S2 outside of the context in which S1 was nonreinforced. The present data support both of these expectations.

More speculatively, the current results are consistent with an optimal predictive-learning framework in that they suggest economy in the rules governing associations. Organisms adapt by learning to anticipate events that are likely to happen. Therefore, the more they take advantage of prior learning, the less they waste time and energy. The evoked representation of S1 in the presence of S2, following sequential learning, signals the organism of the likely appearance of S1. Experiment 1 shows that what was learned to follow S1 is now expected given the presence of S2, allowing the organism without direct S2–outcome experience to be prepared to receive or avoid S1’s outcome. More importantly, if S1 stops predicting the outcome, all events which preceded S1 inherit extinction together with S1. Had the mediated-extinction hypothesis been validated, each cue having established a second-order association with the representation of the outcome would have had to undergo extinction itself through its evoked representation. Therefore, our data from Experiment 1 might be viewed as reflecting adaptive processes. The same principle characterizes what was observed in Experiment 2. Responding to S2 reflected context-dependent associations between S1 and what followed. Here again, once S2 predicted S1, it was the associative status of S1 that controlled responding to S2, sparing the organism having to learn individual meanings of second-order cues dependent upon the context.

In conclusion, the mediated-extinction hypothesis was proposed as a novel means of explaining the decremental effect of extinguishing S1 on responding to S2 in SOC. As it was expected to apply to S2–S1 simultaneous pairings and not S2–S1 sequential pairings because of the backward relationship of S2 to S1, its success could have reduced both types of SOC (i.e., simultaneous and sequential) to a single direct S2–US associative structure, which would have been pleasing in its parsimony. However, Experiment 1 and, more directly, Experiment 2 provided evidence against the mediated-extinction account of why post-SOC extinction of S1 only sometimes reduces responding to S2. Thus, although the mediated-extinction account in principle provided a potentially unifying account of the associative structure of all SOC, the present data argue strongly against it. As both of the present experiments examined sequential SOC, one might ask whether mediated extinction might at least contribute to the commonly observed reduction in responding to S2 that results from extinguishing S1. However, even if such extinction of the evoked representation of S2 contributes to this effect in simultaneous SOC, the present data demonstrate that the absence of mediated extinction in sequential SOC renders implausible the kind of direct associative structure of all SOC that we initially had entertained.