The hidden structure of overimitation

Lyons et al. 10.1073/pnas.0704452104.

Supporting Information

Files in this Data Supplement:

SI Figure 6
SI Movie 1
SI Figure 7
SI Figure 8
SI Figure 9
SI Figure 10
SI Text
SI Movie 2
SI Movie 3
SI Movie 4
SI Movie 5
SI Figure 11
SI Table 3

Fig. 6. Objects used to train participants on the distinction between necessary and unnecessary actions. The experimenter retrieved the toy dinosaur from each object as follows: (A) Tapping on the side of the jar with the feather, followed by unscrewing the lid. (B) Jingling the bell, followed by unscrewing and removing the top half of the cylinder. (C) Putting on the wristband, followed by lifting the lid. (D) Unfastening and lifting the outer lid, followed by removing the lid from the inner soap dish (this item served as a control in which both of the experimenter's actions were necessary). (E) Pulling out the drawer, followed by removing a Velcro star (not visible) from the back of the drawer. (F) Pressing a small piece of silly putty onto the glass top, followed by removing the top from its base. (G) Making a small dot on the aquarium with the pen, followed by lifting the aquarium's lid. (H) Unzipping the purse, followed by tapping the side of the open purse with the wooden mallet.

Fig. 7. The Puzzle Box, modeled after a stimulus used in ref. 1. Vertical columns of images show the two different action sequences that the experimenter used to open the Puzzle Box; horizontal panels group equivalent stages within each sequence. In the left-hand column, the Puzzle Box is opened by using the wand to push out the red wooden bolt and tap in the empty upper compartment (irrelevant), and then pulling out the round plug in the center of the door assembly (relevant); in the right-hand column, the box is opened by using the wand to pull out the bolt and tap in the upper compartment (irrelevant), and then sliding the door assembly aside (relevant). In both cases, the final step is using the wand to retrieve the toy turtle from the box's lower compartment.

1. Horner V, Whiten A (2005) Anim Cognit 8:164-181.

Fig. 8. Puzzle Box detail, highlighting the horizontal Plexiglas divider that separates the box into upper and lower compartments. The upper compartment, accessed through an opening on top of the box beneath the removable red bolt (see SI Fig. 7), is visibly empty. The turtle is located in the opaque blue tube in the lower compartment, and is accessed through an opening on the front of the box behind the circular red door.

Fig. 9. The Cage. Vertical columns of images show the two different action sequences that the experimenter used to open the Cage; horizontal panels group equivalent stages within each sequence. In the left-hand column, the Cage is opened by rotating the metal basket using its side handle (irrelevant) and then unscrewing the cap on top of the central spindle and removing the basket (relevant); in the right-hand column, the Cage is opened by rotating the metal basket using its top handle (irrelevant) and then pulling out the central spindle and removing the basket (relevant). In both cases, the final step is lifting the blue-and-white plastic lid covering the toy turtle.

Fig. 10. The Dome. Vertical columns of images show the two different action sequences that the experimenter used to open the Dome; horizontal panels group equivalent stages within each sequence. In the left-hand column, the Dome is opened by rotating the white wooden locking arm aside and opening the clear plastic box (relevant), and then pulling the metal rod out of the base of the plastic box using its handle (irrelevant); in the right-hand column, the Dome is opened by pulling out the yellow locking pin, flipping the white wooden arm up, and opening the clear plastic box (relevant), and then pulling the metal rod out of the base of the plastic box by using the ball and ribbon (irrelevant). In both cases, the final step is lifting the red wooden lid covering the turtle.

Fig. 11. The Igloo. The uppermost panel shows the two forms of the object, connected and disconnected, varying in whether the two halves are presented as a continuous whole or as discrete objects. The experimenter's action sequences were identical for both forms. The vertical columns of images in A-D show the two different action sequences that the experimenter used when opening the Igloo; horizontal groupings show equivalent stages within each sequence. In the left-hand column, the Igloo is opened by rotating the white wooden wand and lowering it into the right half of the object (irrelevant), and then removing the lid from the left half of the object (relevant); in the right-hand column, the Igloo is opened by pulling out the black support pin and lowering the white wand into the right half of the object (irrelevant), and then lifting the front door on the left half of the object and pulling out the white platform (relevant). In both cases, the final step is lifting the blue wooden lid covering the turtle.

SI Movie 1

Movie 1. An excerpt from the training task used in each of the described experiments. The participant correctly answers that tapping on the jar with the feather is unnecessary for retrieving the dinosaur and receives positive reinforcement from the experimenter.

SI Movie 2

Movie 2. A participant overimitating on the Puzzle Box by removing the irrelevant red bolt and then tapping on the floor of the box's empty upper compartment. The participant also reproduces the causally irrelevant "control tap" elements (described in the SI Methods) used to equate the two versions of the experimenter action sequence for this object.

SI Movie 3

Movie 3. A participant overimitates on the Cage by unnecessarily rotating the wire basket 180° around its central axis. The experimenter's control taps are also reproduced.

SI Movie 4

Movie 4. A participant overimitates on the Dome by pulling the irrelevant metal bolt out of the base of the plastic box. The participant also reproduces both of the experimenter's control taps on the white wooden locking arm and on the handle of the irrelevant bolt.

SI Movie 5

Movie 5. An example of the surreptitious procedure used in Experiment 1B. This participant has just completed Experiment 1A and is picking out her prize. The busy experimenter then requests a simple favor, asking the participant to check whether the assistant remembered to put the toy turtles back into the puzzle objects. As shown, despite time pressure and the fact that the study has apparently ended, overimitation persists. Rather than simply opening the lid of the Dome directly, the participant first pauses to pull the irrelevant metal rod out of the base of the object.

Table 3. Summary of Inter-rater Reliability Statistics

^a Cohen's Kappa statistic could not be meaningfully computed in this case, as both raters assigned a single identical code to all trials.

**, P < 0.001; *, P < 0.01. All values are two-tailed.

SI Text

Methods: Test Stimuli and Experimenter Action Sequences. Test stimuli for Experiments 1A, 1B, and 2A were three novel puzzle objects: the Puzzle Box (based on a stimulus from ref. 1), the Cage, and the Dome (Fig. 1, SI Figs. 7-10); a fourth puzzle object, the Igloo, was used in Experiment 2B (Fig. 5, SI Fig. 11). Each puzzle object was predominantly transparent such that the causal significance of actions performed on it could be easily observed.

The experimenter used two different sequences of relevant and irrelevant actions to open each puzzle object, with presentation counterbalanced across participants. The two action sequences for a given object were largely identical, each calling for the operation of the object's relevant and irrelevant mechanisms in a fixed order. Sequences differed in terms of the specific means that were used to perform these operations (see SI Figs. 7 and 9-11). In one of the Puzzle Box action sequences, for example, the experimenter pushed the red bolt out and removed the central plug from the door (SI Fig. 7, left column), whereas in the other he pulled the bolt out and slid the door aside (SI Fig. 7, right column).

To make the two sequences for each puzzle object as equivalent as possible, the experimenter also performed a few small tapping actions (not shown in the figures; see SI Movies 2-4 for examples) to equalize the attention drawn to each part of the object. For example, when the experimenter removed the Puzzle Box's red bolt by pulling from the left, he would first lightly tap on the bolt's right end; similarly, when the bolt was removed by pushing from the right, the experimenter first tapped on the left end. In this way, attention was drawn to both of a given mechanism's affordances in both action sequence variants. Although participants frequently reproduced these irrelevant "control taps," such tapping was not scored as overimitation. Our approach was more conservative, requiring children to actually operate the puzzle objects' irrelevant mechanisms to be counted as overimitating.

Discussion: Experiment 1A. Children's willingness to optimize the adult's style of action not only clarifies the level of detail at which causal encoding occurs, it also gives us a clearer window onto the relationship between the overimitation reported here and the seemingly opposing phenomenon of selective (or rational) imitation that has been observed in other contexts (2-9). Children have an impressive capacity for selective imitation, weighing factors such as efficiency (2, 3, 5-7), intentionality (4, 8), and external constraints (7) to rationally infer which parts of an adult's behavior to reproduce and which parts to ignore. On its face, such selectivity appears to be distinctly at odds with overimitation, and prior accounts have implicitly endorsed this oppositional view. Overimitation has been framed as an outlier in the social learning landscape-a socially compliant mode of imitation that differs significantly from children's more rational modus operandi (1, 10-12). The style optimization results of the present study, however, argue that overimitation and selective imitation are not as mutually exclusive as they first appear.

Our core argument in this paper, of course, is that overimitation is a kind of rational imitation, reflecting an effective means of causal learning whose logic even adults employ. One way of supporting this account-and differentiating it from prior views of overimitation as irrational (or only socially rational)-would be to show that overimitation bears some signature features that have also been observed in children's selective imitation. Such commonality would give us greater confidence that even when children are overimitating, they are still operating within the same fundamentally rational framework that guides selective imitation. In fact, the reported style optimizations provide exactly this kind of connection.

More specifically, the observed style optimizations replicate a robust trend from the selective imitation literature: When children understand the overall goal of an adult's behavior, the fidelity of their imitation is often modulated by an efficiency bias (2, 3, 5-7, 13). For example, when children infer that an adult's goal is to touch her right ear, they will copy that action using a simple right-handed (ipsilateral) reach, even when they observe the adult reach in a less efficient left-handed (contralateral) manner (3, 6). The roots of this efficiency bias appear to be quite deep, with even 9-month-olds showing surprise when an agent chooses an inefficient means of accomplishing a simple goal (14).

The tendency of our participants to "correct" stylistic inefficiencies in the adult's behavior is thus highly consistent with the prior literature on selective imitation. The fact that these optimizations occurred while children were overimitating the adult's global strategy demonstrates that overimitation does not occur "instead of" selective imitation; the two phenomena can in fact co-occur within the same episode of imitative learning. Therefore, rather than saying that overimitation represents an unexpected or contradictory departure from children's usually rational imitative tendencies, it seems more accurate to say that overimitators are still operating within a rational framework-their rationality is simply being misled by observationally-induced distortions in their causal understanding of the puzzle objects.

Discussion: Experiment 2B. In Experiment 1A we found that violations of the "efficiency principle" [also known as the principle of rational action, i.e., children's expectation that agents will prefer the most direct means of accomplishing their goals (15)] did not block overimitation; children continued to overimitate the adult's operation of the irrelevant mechanisms, even when the specific style in which the adult operated them was objectively suboptimal (see Experiment 1A: Procedure, Results, and Discussion, paragraph 7). In the present experiment, contrastingly, we found that overimitation was blocked by implied violations of the contact principle. Children no longer overimitated when encoding the adult's irrelevant action as causally necessary would require them to infer that the contact principle had been violated.

This contrast between the contact and efficiency principles' effectiveness in blocking overimitation merits further consideration. That is, if the contact principle's ability to inhibit overimitation is explained by its foundational status, then we need to ask why the efficiency principle would not be similarly effective. Although the efficiency principle is not a part of core knowledge per se, it is certainly a core aspect of action perception from as early as 9 months of age (14). Given that the efficiency principle plays such a prominent role in infants' early interpretation of events, one might reasonably have predicted that violations of this principle would have the same effect on overimitation as violations of contact causality.

Yet there is an important distinction between the apparent violations of the contact principle that children observed in Experiment 2B, and the violations of the efficiency principle that they observed in Experiment 1A. Specifically, implied violations of the contact principle have the effect of undermining the physical plausibility of the puzzle object itself. Violations of the efficiency principle, contrastingly, do not imply that there is anything amiss with the object-only perhaps with the actor who is manipulating it. The fact that the error signals presented by these two types of violations center on different components of the display-one on the object and one on the actor-helps to explain why they influence overimitation in different ways. That is, when the contact principle appears to be violated, automatic causal encoding is blocked because the state transformations applied to the target object cannot be mapped onto any physically plausible causal mechanism. Violations of the efficiency principle do not present such a difficulty. While the child might flag the means by which a particular state transformation was accomplished as suboptimal, there is nothing to prevent the transformation itself from being encoded as causally necessary. As a result, despite the fact that both the contact principle and the efficiency principle are "core" elements of action perception, only contact violations succeed in blocking children's tendency to overimitate.

General Discussion: The Limits of Overimitation. What kinds of factors constrain overimitation? In Experiment 2B we began to address this question by showing that children will not causally encode irrelevant actions that imply a violation of the contact principle. It is likely that other factors-particularly of the social and developmental variety-also impose boundaries on the effect. In this section we will use evidence from the broader imitation literature to examine several plausible constraints on overimitation.

Novelty and Complexity. What role does the nature of the target object itself play in overimitation? As we have already described, the novelty of the object is likely to be important. Our hypothesis has been that causal encoding is triggered specifically by the presence of novel objects, and the contrast in children's responses to the familiar training objects versus the unfamiliar puzzle objects argues in favor of this view. That is, as we describe in the Experiment 1A discussion, even children who identified the adult's irrelevant actions on the familiar training objects with 100% accuracy were highly prone to overimitating on the novel puzzle objects (in fact, training score was independent of later overimitation, see Experiment 1A: Procedure, Results, and Discussion, paragraph 5). A more subtle question relates to whether the likelihood of overimitation would change as a child accumulated additional experience with a formerly unfamiliar object. It would be interesting to examine whether a suitable quantity of direct experience with an object might eventually begin to counteract the observationally induced changes in causal beliefs that support overimitation.

A second object-related variable that could potentially constrain overimitation is complexity. Of course, the studies reported here have shown that overimitation and complexity are surprisingly independent in the sense that children will readily overimitate on even causally transparent objects (recall that age-matched baseline participants who did not observe the experimenter readily understood all of the puzzle objects from a cursory visual inspection). Nonetheless, it is possible that there is an extremely small but "non-zero" threshold of complexity that is necessary for overimitation to occur. An interesting question for future work will be to determine whether such a threshold indeed exists, and if so how it changes developmentally as children's causal reasoning becomes more sophisticated.

Free Versus Constrained Action. The logical premise of children's causal encoding is that an adult's purposeful object-directed actions are likely to correspond closely to the causal structure of the object in question. This heuristic is a powerful one, but it is not infallible; it is possible to imagine reasonable sets of circumstances in which its validity would be greatly diminished. Imagine for example a situation in which an adult model's actions are not selected completely freely, but are instead constrained or determined by external factors. An adult with a cast on her hand provides a good exemplar of this scenario. Such an adult might operate a tool in an unconventional way not out of deep causal insight, but simply because of the externally imposed limitations on her movement. Might children's causal encoding be sensitive to such situational constraints, becoming less likely to occur when an adult's object-directed actions are not freely selected?

A classic imitation study by Gergely and colleagues (7) argues that this may in fact be the case. In the experiment, two groups of 14-month-olds observed an adult causing a novel lightbox object to illuminate by pushing on it with her forehead. In the hands-free group, this unusual action appeared to be selected deliberately; the adult used her forehead to push on the lightbox despite the fact that her hands were free and could have been used for the task. Contrastingly, in the hands-occupied group, the adult's hands were engaged by a salient secondary goal, namely holding a blanket around her shoulders as though to stay warm.^* In this case the adult's use of the forehead-pushing action appeared to be imposed by the situation. The question of interest was how this relatively minor alteration in the display would effect children's later imitation of the observed behavior.

The hands-free versus hands-occupied contrast proved to have a powerful influence on children's imitative responses. Children in the hands-free condition showed a strong tendency to copy the adult exactly, activating the lightbox by pushing on it with their foreheads 69% of the time. Despite having seen a nearly identical display, however, children in the hands-occupied condition imitated the adult's use of her forehead only 21% of the time. In other words, children were far more likely to reproduce the adult's unusual and inefficient means of activating the lightbox when she appeared to select it deliberately than when she appeared to employ it for lack of a better option.

These data have clear implications for overimitation, arguing that children's causal encoding is likely be blocked when the adult's behavior appears to be externally determined. An adult's actions may need to be be unconstrained as well as intentional to elicit overimitation.

Pedagogical Cueing. A further possible constraint on overimitation relates to the communicative quality of the observed behavior. Specifically, it is possible that children's causal encoding may be influenced by the presence of natural pedagogical cues (16-19), i.e., cues indicating that the adult's actions are intended to communicate new knowledge.

Gergely and Csibra (16-19) have argued that children are keenly sensitivity to natural pedagogical cues such as eye contact, joint attention, pointing, and referential speech. These cues are thought to effectively constrain the size of children's learning space, functioning like attentional beacons that help children to hone in on information that would be too difficult or abstract to learn in unsupervised manner (see ref. 18 for a more detailed exposition). A convincing body of empirical evidence supports this theoretical account, demonstrating that the presence or absence of natural pedagogical cues strongly influences the manner in which even very young children process observed behavior. A good example of this influence derives from a follow-up to the aforementioned lightbox experiment (17, 20). Following a procedure largely identical to that of the initial study, a new sample of 14-month-olds again observed an adult activating a lightbox object by pushing on it with her forehead. In this version, however, the presence of pedagogical cues was systematically manipulated. Some participants saw the experimenter present the same set of natural pedagogical cues that had been used in the original experiment while for other participants these cues were omitted.

This difference in pedagogical cueing led to a large difference in children's later imitation. While children who received the cues replicated the original result, showing a strong tendency to copy the adult's use of her forehead in the hands-free condition but not in the hands-occupied condition, children in the no-cues condition showed a very different pattern. These participants tended to ignore the adult's use of her forehead regardless of whether her hands had been free or occupied. That is, in the absence of any pedagogical cues, children failed to encode the adult's use of her forehead as a significant component of the display even when she appeared to choose this unusual action deliberately. The same general pattern has been replicated more recently by Brugger and colleagues (2), who found that 14- to 16-month-olds were less likely to reproduce the causally unnecessary components of an adult's action sequence when the sequence was presented without pedagogical cues such as initial eye contact and infant-directed speech.

Overall, results such as these suggest that pedagogical cueing may play a role in facilitating children's causal encoding. In the studies presented here, for example, the simple fact that the adult sat next to the child before acting on the puzzle objects (thus adopting a common visual perspective) may have helped to implicitly frame the interaction as one that was intended to convey new knowledge. It is possible that the same intentional actions presented without the communicative framing might not be as effective in triggering overimitation. Alternatively, it is also possible that the pattern of data observed by other investigators-with the reproduction of irrelevant actions occurring less frequently in the absence of pedagogical cues-may rest on developmental rather than pedagogical factors. That is, as we argue in the next section, overimitation may become much more robust in the period elapsing between the age of the 14-month-olds tested in these prior studies (2, 17, 20) and the 3- to 5-year-olds tested in our experiments. It is thus possible that the overimitation of preschool-aged children may be less sensitive to pedagogical cues, occurring in the presence of intentional action regardless of whether such cues are present. Experiments specifically designed to answer these questions, disentangling the roles that simple intention and pedagogical ostension play in children's causal encoding, will be an important constituent of future overimitation research.

Developmental Factors. The previously mentioned study by Brugger and colleagues (2) suggests that overimitation may interact with development in an interesting way. In that study, 14- to 16-month-olds observed an adult performing a simple two-step action sequence to attain a goal, for example opening a Velcro latch and then lifting a lid to access a set of colorful plastic flowers. The first step in these sequences was manipulated so as to be either (i) causally necessary for the second step to occur (e.g., the Velcro latch initially held the lid shut) or (ii) irrelevant to the second step (e.g., the lid could be opened without operating the latch). Brugger et al. found that infants were less likely to reproduce the first step in the observed sequence when it was unnecessary than when it was causally important. In other words, 14- to 16-month-olds did not overimitate in the same manner as the 3- to 5-year-olds studied here. The much younger subjects seemed better able to apply their own causal knowledge to the adult's action sequence, successfully parsing out components of the observed behavior that were not actually necessary.

While this pattern initially appears somewhat counterintuitive, it actually accords well with our account of overimitation. That is, if overimitation is driven by the way in which children use socially derived data to support their causal learning, then it makes sense that the effect might not be observable until children's social cognition has reached some minimum threshold of maturity. Age may thus constrain overimitation in an inverse manner, with the effect becoming less likely at younger ages. It is worth noting that this pattern integrates nicely with data at the phylogenetic level of analysis. The finding that 14-month-olds are less likely to overimitate than preschoolers may be related to the larger scale finding that chimpanzees are much less likely to overimitate than human children. Again, the overall suggestion is that overimitation may be a later developmental occurrence, one that comes online only as children's social cognition begins to mature in species-specific ways.

1. Horner V, Whiten A (2005) Anim Cognit 8:164-181.

2. Brugger A, Lariviere LA, Mumme DL, Bushnell EW (2007) Child Dev 78:806-824.

3. Bekkering H, Wohlschläger A, Gattis M (2000) Q J Exp Psychol 53A:153-164.

4. Carpenter M, Akhtar N, Tomasello M (1998) Infant Behav Dev 21:315-330.

5. Carpenter M, Call J, Tomasello M (2005) Dev Sci 8:F13-F20.

6. Gleissner B, Meltzoff AN, Bekkering H (2000) Dev Sci 3:405-414.

7. Gergely G, Bekkering H, Király I (2002) Nature 415:755.

8. Meltzoff AN (1995) Dev Psychol 31:838-850.

9. Schwier C, Van Maanen C, Carpenter M, Tomasello M (2006) Infancy 10:303-311.

10. McGuigan N, Whiten A, Flynn E, Horner V (2007) Cognit Dev, in press.

11. Nielsen M (2006) Dev Psychol 42:555-565.

12. Uzgiris IC (1981) Int J Behav Dev 4:1-12.

13. Carpenter M, Call J, Tomasello M (2002) Child Dev 73:1431-1441.

14. Gergely C (2003) Phil Trans R Soc Lond B 358:447-458.

15. Gergely G, Csibra G (2003) Trends Cognit Sci 8:396-403.

16. Gergely G, Csibra G (2005) Interact Stud 6:463-481.

17. Gergely G, Csibra G (2006) in Roots of Human Sociality: Culture, Cognition, and Human Interaction, eds Enfield NJ, Levenson SC (Berg Publishers, Oxford), pp 229-255.

18. Csibra G, Gergely G (2006) in Processes of Change in Brain and Cognitive Development. Attention and Performance, XXI, eds Munakata Y, Johnson MH (Oxford Univ Press, Oxford), pp 249-274.

19. Gergely G, Egyed K, Király I (2007) Dev Sci 10:139-146.

20. Király I, Csibra G, Gergely G (2004) Poster, presented at the 14th Biennial International Conference on Infant Studies, Chicago, IL.

• Early Edition
• Archives
• Online Submission

• Feature Articles

  (Expanded research articles of exceptional breadth)

• Commentaries

  (Expert commentary on leading research)

• Letters

  (Online-only letters to the editor)

• Inaugural Articles

  (Articles by recently elected NAS members)

• PNAS Profiles

  (Biographical profiles of Academy members)

• Sustainability Science

  (Interactions of human and environmental systems)

• Collected Papers

  (Editorials, Perspectives, Reviews, Colloquia, etc.)

• Special Features

  (Solicited articles on exceptional research topics)

• Sackler Colloquia

  (Scientific reports held under NAS auspices)