BackwaRD veRSuS fORwaRD BlOckINg: evIDeNce ...

Psychological Reports, 2011, 109, 3, 1001-1016. © Psychological Reports 2011

Backward versus forward blocking: evidence for performance-based models of human contingency learning1, 2 David Luque

Miguel A. Vadillo

Departamento de Psicología Básica Universidad de Málaga

Departamento de Fundamentos y Métodos de la Psicología Universidad de Deusto

Summary.—Two types of theories are usually invoked to account for cue-interaction effects in human-contingency learning, performance-based theories, such as the comparator hypothesis and statistical models, and learning-based theories, such as associative models. Interestingly, the former models predict two important cueinteraction effects, forward and backward blocking, should affect responding in a similar manner, whereas learning-based models predict the effect of forward blocking should be larger than the effect of backward blocking. Previous experiments involved important methodological problems, and results have been contradictory. The present experiment was designed to explore potential asymmetries between forward and backward blocking. Analyses yielded similar effect sizes, thereby favoring the explanation by performance-based models.

Learning to predict outcomes from currently available cues is a key to the survival of organisms. The study of how animals and humans learn the cue-outcome associations necessary to accomplish this task has been a major topic of research for decades, and it has become the object of important theoretical debates given recent research on human contingency learning. In a typical contingency-learning experiment, participants are first exposed to a series of trials on which certain cues are paired with certain outcomes, and then they are asked to rate the extent to which the cues are related to the outcomes (for recent reviews, see De Houwer, 2009; Mitchell, De Houwer, & Lovibond, 2009; Shanks, 2010). One of the most consistent results found in these experiments is that not all the cues paired with an outcome become equally associated: to become a useful signal, the cue not only has to be systematically paired with the outcome, but it must also be a good predictor, or at least better than other available cues. The study of cue-selection phenomena, such as blocking, is among the most relevant topics in such current research and theory (Shanks, 2010). The traditional design used to study blocking is shown in Address correspondence to David Luque, Departamento de Psicología Básica, Facultad de Psicología, Universidad de Málaga, Campus de Teatinos s/n, 29072 Málaga, Spain or e-mail ([email protected]). 2 Support for this research was provided by Dirección General de Investigación, Tecnología y Empresa of the Junta de Andalucía (Grant P08-SEJ-03586) and Departamento de Educación, Universidades e Investigación of the Basque Government (Grant IT363-10). 1

DOI 10.2466/22.23.PR0.109.6.1001-1016

ISSN 0033-2941

1002

D. Luque & M. A. Vadillo

Table 1. The usual result of experiments conducted with this design is that pairing a cue, B, with an outcome, O1, in the presence of another cue, A, i.e., a better predictor of the outcome, usually results in participants failing to learn the association between Cue B and O1 (or, alternatively, failing to express their knowledge of this association). As shown in Table 1, two alternative designs can be used in blocking experiments, namely, forward and backward blocking. In the forward blocking design, a cue is paired with an outcome in the first phase (i.e., Cue A → O1). Then, in the second phase, the previous cue and a new one are presented simultaneously with the same outcome (i.e., Cues AB → O1). In the backward-blocking design, the order of the phases is inverse (i.e., AB → O1 followed by A → O1 trials). Blocking of the target B → O1 association, as compared to a standard control condition, has been observed in both forward and backward blocking designs. In some associative learning models, similar mechanisms are proposed to explain both forward and backward blocking. In general, this is a common feature of so-called performance-based models. According to this family of models, cue-interaction effects, such as blocking, arise not because processes occur while encoding the cue-outcome associations, but because processes are related to the way in which this information is used to produce a response. For example, according to one of the best known performance-based models, named the comparator hypothesis (Miller & Matzel, 1988), responding to Cue B at test would be directly proportional to the strength of the B–O association and inversely proportional to the associative strength of other cues which have been paired with the target cue, B, and are also alternative predictors of the outcome (e.g., Cue A, in a blocking experiment). Given that this comparison of associations is assumed to occur at test, the order in which the associations are learned is irrelevant within this framework. In a similar vein, algorithmic statistical models of contingency learning, which also belong to this category of performance-based models, assume that the critical cognitive processes TABLE 1 Design Summary of the Experiment Group

Phase 1

Phase 2

Test

Forward

A → O1

Backward

AB → O1 CD → O2

AB → O1 CD → O2 A → O1

B? C? / D?

Note.—Letters from A to D stand for predictive cues. In the present learning task, coloured rectangles were used as cues. O1 and O2 stand for the outcomes. The outcomes used in the present experiment were pictures of plants. Question marks in the test mean that in these trials, no feedback was provided.

BACKWARD VS FORWARD BLOCKING

1003

underlying blocking occur at test (Melz, Cheng, Holyoak, & Waldmann, 1993; Cheng & Holyoak, 1995). Note that we are referring here only to the statistical models that propose an algorithmic-level explanation of human contingency learning. Statistical models labeled as computational are not concerned with the algorithmic-level processes and, hence, it is difficult to draw clear predictions about the effect of procedure manipulations, such as the order of the presentation of the learning phases. Hereafter, the term “statistical models” refers to the algorithmic-level statistical models. Thus, comparator theory and statistical models predict backward and forward blocking of the same magnitude. Supporting this prediction, some experiments comparing backward and forward blocking have shown similar effect sizes for both effects (Aitken, Larkin, & Dickinson, 2001, Exp. 1; Beckers, Van den Broeck, Renne, Vandorpe, De Houwer, & Eelen, 2005; Vandorpe & De Houwer, 2006; McCormack, Butterfill, Hoerl, & Burns, 2009). At odds with these results, other studies in which both effects were compared within the same experiment have yielded larger forward blocking than backward blocking effects (e.g., Chapman, 1991; Lovibond, Been, Mitchell, Bouton, & Frohardt, 2003; Melchers, Lachnit, & Shanks, 2004, 2006; Mitchell, Lovibond, Minard, & Lavis, 2006). The smaller effect size of backward relative to forward blocking is nicely explained by an alternative family of theories, namely, learning-based models, in which cueinteraction effects are explained as the result of processes which occur during the encoding of cue-outcome associations. For instance, most associative models account for forward and backward blocking by changes in associative strength of the target Cue B during learning phases. For example, asymmetries between forward and backward blocking have been explained by Dickinson and Burke (1996), as well as Van Hamme and Wasserman (1994) models. According to these authors, more conditions must be met for backward than for forward blocking to occur, which would explain why the latter is sometimes larger than the former. Specifically, they proposed that a strong association between the two competing cues, B and A, is key to obtain backward but not forward blocking: Participants can only re-evaluate the status of Cue B during the subsequent training of Cue A, to the extent that they remember that Cue A and Cue B were presented together. However, this process does not occur in forward blocking, because in that design, participants are assumed to evaluate the status of Cue B at the same time that Cue A is present (i.e., it is not necessary to rely on memory). From this point of view, a forward blocking effect would be larger than a backward blocking effect simply because in the standard preparations for backward blocking, some participants do not correctly learn the within-compound association between Cues A and

1004


B. Hence, these participants would show no backward blocking, but an intact forward blocking. Indeed, in some experiments the former was obtained only when participants with the poorest scores on a memory test were removed from statistical analysis (Wasserman & Berglan, 1998; Wasserman & Castro, 2005). There is mixed evidence about the relative size of forward and backward blocking when they are compared within the same experiment. On the one hand, some experiments supporting performance-based models have shown both blockings of the same magnitude. But others have shown weaker backward than forward blocking, as in learning-based models. Unfortunately, it is difficult to draw clear conclusions from such results because many researchers did not use the correct control to assess the magnitude of the blocking effects. Most experiments used a control condition in which one element of the compound was paired in the explicit absence of the outcome (i.e., CD → O2; C → NoO2). The main drawback of this procedure as the control condition in a blocking experiment is that any statistically significant difference between the experimental and control conditions can be attributable either to a genuine blocking effect (i.e., a decrease in responding to Cue D following C → O2 trials in the experimental condition) or to another cue interaction effect called recovery from overshadowing (i.e., an increase in responding to Cue D following C-NoO2 trials in the control condition). Recent experiments have shown that recovery from overshadowing is a cue-selection phenomenon different in many senses from blocking and tends to be somewhat stronger and more easily observed in standard tasks (e.g., Vandorpe & De Houwer, 2006). Thus, to draw strong conclusions about blocking, it is very important not to confound blocking and recovery from overshadowing phenomena by using one as a control of the other. Appropriate control conditions for blocking should be used instead. Although there are several alternative control conditions for blocking, it is commonly assumed that a condition in which participants are exposed only to CD-O2 trials and no cues of the compound are presented in isolation, either paired with the outcome or with its absence (see Table 1), provides an appropriate comparison (see Wasserman & Berglan, 1998). Quite surprisingly, there are only a few experiments comparing the relative magnitudes of backward and forward blocking using the correct control. For example, Lovibond, et al. (2003), Mitchell, et al. (2006), and Vandorpe and De Houwer (2006) reported results on series of wellcontrolled blocking experiments conducted with adults. These studies showed mixed evidence. While Lovibond, et al. (2003) and Mitchell, et al. (2006) obtained larger forward than backward blocking, Vandorpe and De Houwer (2006) obtained effects of the same magnitude. But Beckers, et al. (2005) and McCormack, et al. (2009) conducted blocking experiments and


1005

included appropriate control conditions for children as participants. In these two studies, the order in which children received information did not affect the magnitude of blocking. Moreover, the tasks used in these experiments are remarkably similar, which might favor a specific pattern of results over the alternative. Specifically, all previous experiments with adults in which correct control conditions were included used exactly the same task (Lovibond, et al., 2003; Mitchell, et al., 2006; Vandorpe & De Houwer, 2006). In this task, called the allergy task, participants are asked to imagine that they are allergists who have to learn the causal relations between a number of foods (cues) and one or several allergic reactions (outcomes). From a procedural point of view, the participants’ task can be broadly divided into two different types of phases. In the learning phases, participants learn trial-bytrial that some foods cause an allergy in a fictitious patient (and others not). In a learning trial, participants first see the food or foods eaten by the patient on a particular occasion. Then, they have to guess whether the patient will have an allergic reaction by pressing the keys “Y”es or “N”o. Then, feedback is provided, informing them whether the patient actually had the allergy. Thus, error-driven learning is expected to occur in base of this feedback, and participants adjust their responses to the actual contingencies between food(s) and allergy (cues and outcome). The knowledge acquired in learning phases is evaluated in test phases by means of subjective judgments. In these judgments, participants have to rate the strength of the causal relations on Likert-type rating scales (with different anchors in each experiment). The specific wording of the questions usually stresses that participants have to rate the strength of the causal relationship between the cues and the outcome. For example, in Lovibond, et al. (2003), the participants were asked to rate the “likelihood that each food would cause an allergic reaction in Mr. X.” Then, all cues (foods) were displayed with a Likert-type scale below each one. These scales had nine points with anchors of 0: Definitely not, 4: Possibly, and 8: Definitely. On the other hand, the two studies conducted with children used adapted experimental tasks, noticeably different from the allergy task. For instance, McCormack, et al. (2009) used the “blicket detector” task. In this task, children are asked to discover whether a number of objects (in the role of cues) are blickets. To do this, they can use a special blicket detector that goes off only when a blicket is placed on top of it. Thereby, in two of three experiments in which the allergy task was used (Lovibond, et al., 2003; Mitchell, et al., 2006), larger forward than backward blocking was found. In contrast, in two experiments in which an alternative task was used (Beckers, et al., 2005; McCormack, et al., 2009), similar forward and backward blocking was obtained. Thus, evidence shows that the asymmetry between forward and backward blocking is found only in some experi-

1006


ments and only when the allergy task was used. This raises serious doubts about the generalizability of this result. In view of these puzzling results, the present aim was to assess whether forward blocking is a stronger effect than backward blocking. In this study, forward and backward blocking were tested with adults. In contrast with most of the previous studies, correct control conditions were included in the design (see Table 1). Also, the allergy judgment task was not used. To get more general results, a non-causal task was used. The participants were required to learn arbitrary associations without any causal meaning. Also, subjective judgments were not the main dependent variable. Instead, the number of operant responses on test trials was measured (see Luque, Morís, Cobos, & López, 2009). By doing so, contrasting the generalizability of previous studies was possible, and assessment of backward and forward blocking was done. In sum, to know whether forward blocking is stronger than backward blocking is an important issue for discriminating between human-contingency learning models. Experiments including the correct control condition are sparse (Lovibond, et al., 2003; Beckers, et al., 2005; Mitchell, et al., 2006; Vandorpe & De Houwer, 2006; McCormack, et al., 2009). In addition, the evidence from these experiments is difficult to interpret. All the experiments with adults were conducted with the allergy task and tend to yield stronger forward than backward blocking (Lovibond, et al., 2003; Mitchell, et al., 2006). On the other hand, experiments with children were conducted with the blicket detector task and yielded forward and backward blocking effects of similar size. These divergent results between experiments could be explained in at least two ways. On the one hand, it is possible that the asymmetry between forward and backward blocking has an ontogenic origin, and it arises with age (see McCormack, et al., 2009). On the other hand, this asymmetry might depend on the use of the allergy task and, therefore, it would disappear when a different task is used. To discriminate between these possible accounts, a forward/backward blocking experiment with adults was conducted, using an experimental procedure that is very different from the standard allergy task. If the differences between forward and backward blocking depend on the age of the participants, larger forward than backward blocking would still be obtained, as in previous experiments with adult participants. On the other hand, if the forward/backward asymmetry depends on the task, effects of similar size would be obtained, as observed in previous studies that have not relied on the allergy task. Method Participants and Apparatus Forty-nine psychology students from the University of Deusto volun-


1007

teered for this study. The experiment was run in a room in which participants sat 1 m apart from each other. The task was programmed using Visual Basic (Microsoft, USA). Design and Procedure The design is shown in Table 1. To gather more data of participants’ behavior during the experiment, two different cues were used for Cue A (A’ and A”), two different cues were used for Cue B (B’ and B”), and two different outcomes for O1 (O1’ and O1”).3 In both groups, each type of compound-cue trial was presented 20 times in pseudo-random order, without giving more than two consecutive trials of the same type. Singlecue trials were presented 15 times, each following the same randomized criteria. In the test phase, Cues B’, B”, C, and D were presented once per participant in a counterbalanced order. The procedure was similar to that used by Luque, et al. (2009). The participants could earn points by betting on each trial. To do this, they had to learn the relations among some coloured rectangles and some fictitious plants; these were used as cues and outcomes, respectively. (An English translation of the task instructions is given in the Appendix, pp. 10151016.) Rectangles were blue, brown, yellow, red, green, and pink, and their assignment to Cues A–D was randomized across participants. The plants (pictures of fictitious plants labelled as Kollin, Dobe, and Yamma) were used as outcomes O1’, O1”, and O2, and were counterbalanced. To avoid participants learning associations between cues and specific motor responses, the response key assigned to each outcome varied randomly trial by trial. For this reason, the first thing which occurred immediately before each trial was a display of the three possible random response options (labelled plant photos), one per outcome. This was done so participants could know, before the trial started, which key to press during that trial if they wanted to bet for any of the three possible outcomes. Under each response option, a scroll bar, together with a text box, was displayed to indicate the number of points a participant was betting for the corresponding option on each trial. Then, the cue or cues (i.e., one or two of the coloured rectangles) appeared at the middle top of the screen for 2.5 sec., during which participants had to bet which of the three plants they thought was associated with the cue (or cues) present on the screen for that trial. When two cues were displayed on one trial, they were placed one beside the other, with the specific location for each one being counterbalanced trial by trial. To respond, participants placed their bets by pressing either Key “1,” The control condition was not doubled, as in the experimental one, because the standard control condition already provided two sources of data, participants’ behavior in response to C and to D.

3

1008


or Key “2,” or Key “3” for the plant (i.e., the response option) located at the left, middle, or right bottom of the screen, respectively. While a given response key was kept pressed, points bet for the corresponding option increased continuously, and were indicated analogically, by the movement of the scroll-bar from left to right, as well as with numbers ranging from 0 to 100 in the corresponding text box. Once the 2.5 sec. time had passed, the cue (or cues) disappeared, as indicated by the rectangle (or rectangles) taking on the grey colour, and pressing any response key no longer affected the amount of points bet. On each trial, participants earned as many points as those bet for the correct outcome, and lost as many points as those bet for an incorrect outcome. After each bet, feedback was given as (a) the correct plant, which was visible as the remaining ones were removed from the screen, (b) number of points earned on that trial, as indicated in a text box at the centre of the screen over the labelled plant photographs, and (c) total points gained throughout the training trials, which was indicated in a text box at the right top of the screen (see Fig. 1). The test phase were four additional trials which differed from the previous training trials in only two respects. Cues were presented for 5 sec., and participants received no feedback after their bets. (A)

(B)

(C)

Fig. 1. Experimental task. (A) In the beginning of each trial, the three outcomes (the plants at the bottom of the screen) were visible. Participants’ responses were not allowed until the cue was present. (B) Then the cue was present, and participants could start making their bets. They could either bet points on the right, center, or left plants (i.e., Keys 1–3 on the keyboard, respectively). Bets were allowed only for the 2.5 sec. that the cue was present. (C) Once bets had been made, the correct outcome was shown. Moreover, participants could check the number of points won or lost in the current trial and the total score accumulated across trials.

Results Data from participants whose total scores at the end of the two learning phases were less than zero points were removed. Using this criterion, data from five participants (one from the Forward condition) were exclud-

1009


ed from subsequent analyses. The dependent measure for the following analyses was the number of points bet on the correct outcome on test trials (e.g., in the test trials for backward blocking, the dependent measure was the number of points bet on Outcome O1; see Table 1). Mean correct responses at test by the remaining participants are shown in Table 2. As can be seen, they tended to respond less to the blocked cues than to the control cues. A 2 (group: Forward vs Backward) × 2 (cue: Blocked vs Control) mixed analysis of variance yielded a statistically significant main effect of cue, but the interaction for cue × group was not statistically significant (Table 3), as the size of this blocking effect was comparable across groups. Interestingly, the main effect of group was also statistically significant, as participants in the Backward group responded during the test phase more than participants in the Forward group. Visual inspection of the data suggests that differences in base-rate responding in Forward and Backward groups are the basis of this unexpected result. This difference in base-rate responding is also observable in an analysis of the total number of responses to all the responses options (regardless of whether they were correct or not). That is, overall participants responded more in the Backward than in the Forward condition. Importantly, this effect of group did not interact with the cue factor (Table 3). The present experimental design does not allow the cause of this difference in base-rate responding to be ascertained. However, this issue is not related to the main goal of this experiment. Regardless of the mechanisms responsible for this effect, a potential shortcoming of previous analyses is that their focus was only on the number of correct responses and so the number of incorrect responses at test were neglected. Therefore, the seemingly high number of correct responses given by participants in Table 2 Responses at Test: Means and Standard Errors Measure Correct R Correct–Incorrect % Correct R

Group FB BB FB BB FB BB

Blocked (B’/B’’)

Control (C/D)

M

SEM

M

SEM

9.9 25.8 1.8 2.1 44.3 45.2

3.3 6.6 7.9 8.7 8.5 9.0

16.0 38.5 7.8 25.2 50.8 74.2

3.5 6.9 7.6 8.3 8.2 8.6

Note.—B’/B” is the mean response to the correct outcome (O1’ or O1”) when the blocked cues (B’ and B”) were presented at test. C/D is the mean response to control cues (C/D) at test and also to the correct outcome, O2. Lower responding to B’/B” than to C/D is indicative of blocking. Both FB and BB are observed with three alternative measures of responding: absolute number of correct responses, difference between the number of correct and incorrect responses, and percent of correct responses.

1010

D. Luque & M. A. Vadillo Table 3 Statistical Analyses

Analysis

Dependent Variable

ANOVA: Correct 2 (group) × responses 2 (cue)

Effect

Cue Group Cue* Group ANOVA: All responses Cue 2 (group) × Group 2 (cue) Cue* Group ANOVA: Correct minus Cue 2 (group) × incorrect Group 2 (cue) Cue* Group ANOVA: % Correct Cue 2 (group) × responses Group 2 (cue) Cue* Group

F1,42

MSE

p

η2

β

5.08 10.94

382.6 738.7

.03* > .01*

0.11 0.21

0.60 0.90

0.62 1.63 22.57

382.6 990.6 738.7

.44 .21 > .01*

0.02 0.11 0.35

0.12 0.24 0.99

0.33 4.57 0.91

990.6 1,008.9 1,877.5

.57 .04* .35

0.02 0.10 0.02

0.08 0.55 0.15

1.57 7.58 0.91

1,008.9 1,097.6 4,275.1

.22 > .01* .16

0.04 0.15 0.05

0.23 0.77 0.29

1.75

1,097.6

.19

0.04

0.25

*p