Perception and action as viewed from the Theory of Event Coding: a multi-lab replication and effect size estimation of common experimental designs

Perception and action are the cornerstones of psychological inquiry, and perceptual processing as well as action-related decision making have enjoyed widespread interest throughout most of psychology’s history. In contrast, the mechanics of how decisions and action intentions are actually translated into overt motor behavior has traditionally received only little attention in psychological research, and this research area has, therefore, been dubbed as the “Cinderella of Psychology” (Rosenbaum, 2005). In fact, motor processes have been often viewed as the mere endpoint of a processing stream, echoing basic ideas of early philosophical accounts of perception and action (Descartes, 1664; Dewey, 1896; see also Hommel & Elsner, 2000).² This view of motor processes as the final, operative end of psychological processing is particularly explicit in stage theories of human information processing (Donders, 1868; see also Pashler, 1994; Sanders, 1990). These theories assume a processing stream beginning with perception of a stimulus. In a subsequent stage, often referred to as response selection, the resulting perceptual representation is translated into a motor representation which is used to emit the respective motor behavior, rendering the motor part a vicarious, subordinate process. The question of how cognitive processes actually interface with the agent’s motor repertoire, however, is commonly not considered by these models.

The theoretical stance adopted by information processing accounts mainly derived from a common focus on inferring processing characteristics—such as potential capacity limitations—rather than tackling how action plans are represented in the human cognitive system (but see Janczyk & Kunde, 2020, for a specification related to TEC). This state of affairs changed quite drastically with the rise of research that was stimulated by the hypothesis of common coding of perception and action (Prinz, 1984, 1992; see also Prinz et al., 2013).

Research that implemented the common coding framework was driven by the ambitious goal of determining the representational foundations of human action control (Prinz, 1984). Common coding breaks with traditional views that see perception and action as two distinct compartments of cognitive function—an assumption that psychological research had inherited from its roots in Western philosophy and its Cartesian view of the mind (Descartes, 1664). Neurophysiological and neuroanatomical approaches had followed the same route by portioning the brain in “sensory” and “motor” compartments (e.g., Harleß, 1861; Laycock, 1845; see also Pfister & Janczyk, 2012). Accepting a qualitative divide between perception and action requires additional theorizing to explain how actions are brought about, for example, by introducing concepts such as stimulus–response translation or response selection as mentioned above. As an elegant and powerful alternative, common coding suggests that perception and action planning share the property to internally represent events in a commensurable representational format, that is, in a common code. The notion of common coding therefore explains a range of direct effects of perception on action planning and vice versa, and it allows for modelling action planning and initiation as anticipated perception (Hommel, 2009), a notion borrowed from the ideomotor principle of human action control (Harleß, 1861; James, 1890). These ideas were summarized in a landmark article that introduced TEC (Hommel et al., 2001). TEC has had a considerable impact on psychological research in this field ever since, and we describe its major tenets in the following section.

Perception and action control involve many layers of physiological and cognitive processes. Perception for instance, will always involve basic receptors, such as the photoreceptors in the retina that transmit signals for further processing. Crucially, these sensory signals encode information about entities in the world (objects, people, the agent’s own body), and representations of these entities are constructed by processing certain aspects of the sensory data (Brunswik, 1944; Heider, 1930/1959; Prinz, 1992). TEC is concerned with these latter representations, that is, representations in terms of distal codes. Proximal events, such as activity of sensory receptors or muscle physiology, are only relevant in that they interface with distal codes and form the basis of such distal representations (Hommel, 2009; Hommel et al., 2001).³ It is because actions rely on anticipations of to-be produced perceptions that the level of distal representations prepares the ground for a common representational format in which perception and action control can interact (Prinz, 1992).

TEC further follows previous work on (visual) perception and motor planning in assuming that perceptual events and action plans alike are composed of individual features that have to be bound together to form integrated representations (Allport, 1993; Rosenbaum, 1980; Treisman, 1996). In TEC’s terminology, such feature bundles are referred to as event files (Hommel, 2004; Hommel & Wiers, 2017). If a feature becomes activated, it spreads activation to the event that it is associated with. Once an event is fully retrieved or established by integrating the corresponding features into an event file, this representation will inhibit competing event files that would draw on features that are now less available. This dynamic interplay of activation and integration is a core mechanism to explain why several perceptual representations and action representations sometimes boost each other while they interfere at other times (see also Thomaschke et al., 2012).

But why would an organism form such integrated event files at all? A major consequence of the proposed cognitive architecture is that activating features allows for selecting, planning, and initiating overt bodily movements. Here, TEC imports ideomotor theorizing in assuming that actions are coded in terms of the sensory consequences they produce (Hommel, 2009). These sensory consequences can relate to perceived movements of the agent’s body and perceived changes in the agent’s environment alike, as long as they are consistently related to certain motor actions (Pfister, 2019). Crucially, TEC assumes that activating sufficiently many features of such an action is the cognitive antecedent of overt bodily movements in that these representations are bi-directionally related to activity of the muscular system (Hommel, 2009; Hommel & Wiers, 2017). This assumption thus closes the gap between perception and action and allows TEC to capture the cognitive underpinnings of perceptual processing and action planning alike. It, therefore, offers a theoretical connection between several areas of psychological theorizing, ranging from multisensory integration to learning and motor control. Each of these fields, of course, comes with its own theories tailored to understanding relevant key findings (e.g., Ernst & Bülthoff, 2004; Jeannerod, 2006). What makes TEC unique, however, is that it provides a framework for distilling commonalities between these seemingly distinct aspects of human cognition. Beyond this basic research interests, aspects of TEC have further been used as a framework in applied research as well, for example, in clinical research (Kleimaker et al., 2019; Petruo et al., 2019).

TEC draws on a large array of experimental approaches that address interesting questions on the interplay of perception and action. As such, it has stimulated a remarkable range of novel discoveries. Most of the original studies, however, originate from a period of psychological research in which power calculations had not arrived in the methodological mainstream and small sample sizes were thus the norm (n = 8–18 in the main studies considered below). Certainly, this does not compare favorably with current requirements and standards (Brysbaert, 2019), and this empirical limitation has several direct consequences for current research.

As a first consequence, small sample sizes provide little power for detecting effects of interest if corresponding effect sizes are small, and they provide little grounds for judging the absence of an effect as compared to the mere absence of evidence in favor of this effect. Several relevant theoretical claims do rest on such null effects, however. For example, Kunde (2001, Exp. 1) took a non-significant effect to infer that different mappings of responses to response-contingent effects can be acquired flexibly without being overshadowed by previous associations including the same events. Another example relates to the study by Hommel (1998), who bases a key argument on the absence of higher-order interactions in the corresponding experimental design. To evaluate the empirical foundations of such claims, it is instructive to ask whether such studies would have had sufficient power to detect small effect sizes that would be theoretically relevant nevertheless. Here, the study by Kunde (2001) came with a power of about 1 − β = .05 to detect an effect size of Cohen’s d_s = 0.2 in the corresponding between-participants comparison (n = 10 overall, distributed equally across both sequence conditions), and the study by Hommel (1998) came with a similar power of 1 − β = .07 to detect relevant within-participant effects of Cohen’s d_z = 0.2 (n = 8). Conversely, the former study would have to assume Cohen’s d_s = 2.0 for a satisfactory power of 1 − β = .80 and the latter study would have to assume Cohen’s d_z = 1.2 for the same power, so that sufficient power is only available for considerably large effects. Whether or not the corresponding theoretical claims hold is thus difficult to assess based on these original studies.

As a second consequence, small sample sizes tend to yield effect size estimates that are heavily biased and often inflated (see Ulrich et al., 2018). In this context, it should be noted that the presence versus absence of an effect is critical for TEC as described in the preceding paragraph, whereas the exact size of an empirical effect does not have theoretical implications per se. As TEC continues to stimulate novel research, however, it would be highly useful for the community to have reliable estimates of how big certain effects are when planning sample sizes via power calculations. If such power calculations rely on inflated estimates of the population effect sizes, the resulting studies will be underpowered and thus are easily subject to not detecting an effect that is actually present. Such null effects might either end up as unpublished file-drawer results; worse, they might be misunderstood as another example of non-reproducible phenomena or they might even motivate incorrect theorizing when taken to indicate absence of an effect. In this regard, it seems notable that there are no direct replications of the original experiments in the literature, despite a striking number of conceptual replications and extensions. Within these conceptual replications, effect size estimates also vary considerably at times (e.g., ranging from d_z = 0.37 to d_z = 0.99 in a re-analysis of the data reported by Pfister et al., 2010), suggesting that some of the findings might relate to underpowered designs. Considering the precision of effect sizes is thus both of practical and theoretical relevance. For instance, moderation and mediation hypotheses require knowledge about the heterogeneity of effect sizes, because they ask questions about relative changes in the size of effects. Similarly, fields such as personality and clinical psychology ask questions about the interindividual variation of effect sizes.

The present study aims to solve these limitations by re-assessing the empirical basis of TEC with experiments using sufficiently large samples. We will describe the corresponding methodological considerations in the following section.

A challenge for determining the most relevant empirical phenomena for TEC is that the theoretical relevance of different phenomena certainly comes with a subjective component so that it may be gauged differently across researchers. Moreover, several phenomena have been picked up by distinct communities, depending on their preferred theoretical focus. We therefore chose to base our selection not on our own preferences, but rather opted to involve a large number of experts in the field to strive for a maximally representative selection. We reached out to 114 researchers who had published studies with related observations and asked them to rate which empirical phenomena and corresponding approaches they perceived as closely tied to TEC, either by lending considerable support to its theoretical notions, or by being immediately stimulated by TEC. More precisely, we asked for a structured response in a brief questionnaire (see Appendix A) and received a total of 49 responses (including our own assessment). The questionnaire was online for one month after its dissemination via e-mail in September, 2018.

This questionnaire included the following ten items in alphabetical order: action-induced blindness (e.g., Müsseler & Hommel, 1997a), code occupation (e.g., Stoet & Hommel, 1999), dimension weighting (e.g., Fagioli et al., 2007), distractor-response binding (e.g., Frings et al., 2007), feature weighting (e.g., Memelink & Hommel, 2013), response-effect (R-E) compatibility (e.g., Kunde, 2001), R-E learning (e.g., Elsner & Hommel, 2001), sensory attenuation (e.g., Blakemore et al., 2000), short-term R-E binding (e.g., Dutzi & Hommel, 2009), and stimulus–response (S-R) binding (e.g., Hommel, 1998).

The questionnaire further provided the opportunity to comment on our selection and to suggest additional phenomena that were not included in the original list. In addition to elaborate theoretical remarks, several such phenomena were brought up by the experts, including the impact of action effects on S-R compatibility phenomena (Hommel, 1993), referential coding (Dolk et al., 2013), longer-term associations between stimuli and responses such as instances (Logan, 1988), levels of S-R representations (Horner & Henson, 2011), and stimulus-task bindings (Waszak et al., 2003). Their clear relation to TEC notwithstanding, none of these phenomena was brought up by more than one expert so that we chose to retain our original selection of ten phenomena.

For each of the ten empirical phenomena, we assessed how many of the respondents had marked them as highly related to TEC. Figure 1 summarizes the resulting opinions. Six of the ten phenomena were suggested by more than 50% of the experts, so that we settled on these phenomena as a final selection: R-E compatibility (82%), action-induced blindness (80%), R-E learning (76%), S-R binding (67%), code occupation (59%), and short-term R-E binding (59%).

The selection of the exact experimental design we used in the present study was based on three criteria. First, we considered always the first study on the particular effect of interest as judged by the date of the corresponding journal publication. Second, we opted for taking the first reported experiment in a study that used a within-subject design for the effect of interest. Third, we took into account the possibility to run the experiment on standard computer hardware without specific display or input devices, to be able to run the same experiment in different labs. From these considerations, Experiment 1 was chosen for the studies by Dutzi and Hommel (2009), Kunde (2001), Müsseler and Hommel (1997a), and Stoet and Hommel (1999). For the study by Elsner and Hommel (2001), we chose Experiment 3a for the following two reasons: Experiment 1 would have involved a between-subject comparison, and Experiment 2—although using a within-subject comparison—did not employ catch trials in a free-choice task, a point we considered a shortcoming of this experiment. Finally, from the study by Hommel (1998), we focused on Experiment 1b, thus omitting the single-response condition used in Experiment 1a.

The goal of our study was to improve the precision of effect size estimates for a range of experimental effects. We, therefore, could not rely on traditional power calculations, because the resulting sample sizes are optimized for deciding whether or not a non-zero effect is present in the population, given an effect of particular size. Instead, we planned our sample size according to how precisely effects can be estimated, operationalized as the width of the 95% confidence interval for standardized effect sizes.

A detailed account of this approach is provided in Appendix B. This approach suggests a sample size of n = 100 participants to yield such confidence intervals with a width of 0.4, thus allowing to characterize the resulting effect size at a precision of roughly ± 0.2 (note that confidence intervals for standardized means are not perfectly symmetrical).

With this in mind, we decided to recruit a total sample of n = 120 participants to be able to compensate potential data loss. As per our pre-registration (https://​aspredicted.​org/​ad6xq.​pdf), data from n = 30 participants were collected in each of four different laboratories in Freiburg, Jena, Trier, and Würzburg (all Germany).⁴ Table 1 summarizes the demographics for the entire sample and the subsamples. An Analysis of Variance (ANOVA) on the participants’ age with the between-participants factor laboratory site was significant, F(3, 116) = 14.33, p < .001, \(\eta_{{\text{p}}}^{2}\) = .27. The gender distribution did not differ between laboratories, χ²(6) = 8.30, p = .217, and the same was true for handedness, χ²(3) = 3.02, p = .389.

If participants were excluded from analyses (due to specific criteria or because of incomplete data), this will be mentioned in the respective methods parts of the particular experiment.

Based on the expected duration of the experiments (judged by pretests), the six experiments were grouped into two clusters. Cluster 1 comprised short-term R-E binding (Dutzi & Hommel, 2009, Exp. 1), action-induced blindness (Müsseler & Hommel, 1997a, 1997b, Exp. 1), and code occupation (Stoet & Hommel, 1999, Exp. 1); Cluster 2 comprised R-E compatibility (Kunde, 2001, Exp. 1), R-E learning (Elsner & Hommel, 2001, Exp. 3a), and S-R binding (Hommel, 1998, Exp. 1b). All participants took part in all six experiments and the two clusters were applied in two 2-h sessions, scheduled on two different days within one week. The order of experiments within a session was counterbalanced, while the order of clusters was determined randomly for each participant.

Analyses and presentation of results were kept closely to the original studies. In several cases, we supplement these results with additional analyses, for example, including additional factors that were not analyzed originally or applying different criteria for participant exclusion. Crucially, we also extracted effect size estimates for each main comparison of interest, and we computed the corresponding 95% confidence interval for standardized means around this effect size estimate. Figure 2 provides an overview of the main results relative to the estimates of the original studies. Details on how we computed the original effect sizes are reported in Appendix C. The six selected studies will be described in the following, ordered according to the degree of agreement among the experts. Data and analyses scripts can be found at https://​osf.​io/​hgy5q/​.

Mean correct RTs and PEs are summarized in Table 2, and Fig. 3 shows detailed descriptive statistics for the quintile analysis (Fig. 3B) and the distribution of individual R-E compatibility effects (Fig. 3C). For RTs, the main effect of R-E mapping was significant, F(1, 118) = 13.03, p < .001, \(\eta_{{\text{p}}}^{2}\) = .10. In addition, participants responded faster with the order ‘incompatible-compatible’ than with the opposite order, F(1, 118) = 5.28, p = .023, η_p² = .04. The interaction was not significant, F(1, 118) = 2.67, p = .105, \(\eta_{{\text{p}}}^{2}\) = .02. Averaged across both orders, the R-E compatibility effect was 15 ms (see Fig. 3C), t(119) = 3.58, p < .001, d = 0.33, 95% CI_SM = [0.14, 0.51].

Participants made overall more errors with the order ‘compatible-incompatible’, although the main effect was not significant, F(1, 118) = 3.69, p = .057, \(\eta_{{\text{p}}}^{2}\) = .03. The main effect of R-E mapping was not significant, F(1, 118) = 0.25, p = .619, \(\eta_{{\text{p}}}^{2}\) < .01, and the same was true for the interaction, F(1, 118) = 0.32, p = .570, \(\eta_{{\text{p}}}^{2}\) < .01.

Mean correct RTs are visualized in Fig. 3B separately for the five RT quintiles. The corresponding ANOVA revealed a significant main effect of R-E mapping, F(1, 119) = 13.03, p < .001, \(\eta_{{\text{p}}}^{2}\) = .10, and of bin, F(4, 476) = 1740.33, p < .001, \(\eta_{{\text{p}}}^{2}\)= .94, ε = .295. In addition, the interaction was significant, indicating an increase of the R-E compatibility effect with increasing RTs, F(4, 476) = 13.07, p < .001, \(\eta_{{\text{p}}}^{2}\) = .10, ε = .361. The R-E compatibility effect was significant from the second quintile on, Quintile 1: t(119) = 0.71, p = .478, d = 0.06; Quintile 2: t(119) = 2.61, p = .010, d = 0.24; Quintile 3: t(119) = 3.52, p = .001, d = 0.32; Quintile 4: t(119) = 3.85, p < .001, d = 0.35; Quintile 5: t(119) = 3.84, p < .001, d = 0.35.

As in the original demonstration of an R-E compatibility effect (Kunde, 2001, Exp. 1), participants responded slower when producing incompatible rather than compatible action effects. This observation is consistent with the idea that effect anticipations are functionally relevant for human action control.

The effect size observed in the present sample was clearly smaller than the originally published effect size (d_z = 0.33 as compared to d_z = 0.96).⁶ Observing a small effect for this design also seems to be plausible, because action effects were a mere by-product of the response and thus entirely task-irrelevant. We would, therefore, assume the small sample size of the original study to have yielded an inflated effect size estimate. This reading of the data is also in line with the observation that the confidence interval around the original effect size estimate spans a considerable range, including the present estimate. It thus seems to be useful to conceptualize R-E compatibility effects as being of small size if the action-effect mapping is not relevant to the task at hand. Follow-up research with this approach has further suggested that task-relevance boosts R-E compatibility effects (Ansorge, 2002; Janczyk et al., 2015; Zwosta et al., 2013). The present observation of a small effect size should thus be seen as applying specifically to settings in which action effects are not relevant to the task at hand. Settings with task-irrelevant effects, however, might be regarded as providing particularly strong evidence for the functional role of effect anticipations if they still yield a reliable impact of R-E compatibility. We, therefore, believe that the present estimate provides useful information for future studies that aim to delineate when human action control draws on environment-related action effects in addition to or even instead of body-related action effects such as proprioceptive and kinaesthetic reafferences triggered by the moving body (Pfister, 2019; Pfister et al., 2014c; Thébault et al., 2018; Wirth et al., 2016).

Owing to the higher power of the present study relative to the original one, the data also suggest a reliable impact of the order of R-E mappings that had not been significant in the original study (Kunde, 2001). Responses were faster when participants began with the incompatible R-E mapping. This observation might suggest that participants did not include the experimentally induced action effects into their action representations when starting with an incompatible mapping so that they could establish efficient stimulus–response associations between target colors and corresponding body-related action effects. When transitioning to the compatible condition, they might have consolidated these associations sufficiently strong so that they established stimulus-effect associations independently of the existing stimulus–response associations. When starting with the compatible mapping, in contrast, they might have been inclined to integrate stimuli, responses, and effects in joint associations instead. This explanation is of course speculative at this point. Its speculative nature highlights that the interplay of pre-existing and newly learned R-E mappings is poorly understood at present which requires additional work to establish when which kinds of action effects are preferably integrated into action representations, and whether different R-E associations can be built up and retrieved in parallel.

A final consideration pertains to different types of R-E compatibility. Even though the present design with spatial compatibility between keypress responses and discrete visual effects likely is the most common version used (e.g., Ansorge, 2002; Janczyk & Lerche, 2019; Janczyk et al., 2017; Pfister & Kunde, 2013; Pfister et al., 2010; Shin & Proctor, 2012), there are numerous other viable possibilities to implement dimensional overlap between responses and effects (Kornblum et al., 1990). R-E compatibility effects were also reported and discussed, for example, with overlap on the dimensions intensity (Kunde, 2001; Kunde et al., 2004), duration (Kunde, 2003; Pfister et al., 2013), as well as for semantic compatibility (Földes et al., 2018; Hubbard et al., 2011; Koch & Kunde, 2002; Koch et al., 2021). Different response modalities included tool-transformed movements such as rotations via steering wheels and flight yokes (Janczyk et al., 2012c, 2015; Yamaguchi & Proctor, 2011), and the operation of one-pivot levers (Janczyk et al., 2012b; Kunde et al., 2007, 2012), as well as continuous mouse movements (Pfister et al., 2014b; Hommel et al., 2017; Wirth et al., 2015; but see Schonard et al., 2021). One case where no R-E compatibility effect was observed concerned touchless gestures (Janczyk et al., 2019). In general though, each of these studies set out to test specific theoretical predictions or applied scenarios that go beyond the scope of this investigation. They also yielded highly different effect size estimates, with many studies exceeding the effect size reported here. When approaching a novel setting without direct precursors in the literature, it seems useful to plan for the possibility of R-E compatibility effects being relatively small, thus requiring sufficiently many participants for meaningful investigations.

Targets were indeed identified less accurately when they faced in the direction of a planned key press (i.e., when the prepared key press and arrow direction were compatible) than when they faced in the opposite direction (i.e., when prepared key press and arrow direction were incompatible). However, the effect was considerably smaller (d = 0.16) than the medium-sized effect in the original study (d = 0.61).

Similar results were reported by several other studies using spatial compatibility (Müsseler & Hommel, 1997b; Müsseler et al., 2001; Thomaschke et al., 2018) as well as compatibility between pronounced color words and visually presented color patches (Kunde & Wühr, 2004). Going beyond intrinsic overlap of action features and events in the environment, action-induced blindness has also been observed for newly induced action effects such as specific onset and offset events in a visual display (albeit using the somewhat different measure of identification speed rather than accuracy; Pfister et al., 2012).

It is an open question, however, how these blindness effects relate to attenuated sensory processing after a response has actually been carried out. Findings of sensory attenuation following own actions were reported for neurophysiological measures such as event-related potentials (Horváth, 2015; Klaffehn et al., 2019; Timm et al., 2014; see also Schafer & Marcus, 1973). Corresponding behavioral findings have been reported only rarely in the literature, and if such effects were observed, they also tended to be of small size for body-external action effects (Weiss et al., 2011a, 2011b; for attenuated perception of body-related effects, see Bays et al., 2005, 2006; Shergill et al., 2003). It is tempting to attribute action-induced blindness during action planning and sensory attenuation after action execution to a single mechanism, that is, integration of feature codes into an event file (but see Thomaschke, 2012; Thomaschke et al., 2012). Additional work at the intersection of both phenomena is thus well-advised to recruit sufficiently large sample sizes, especially when addressing behavioral proxies of sensory processing.

A first relevant distinction between action-induced blindness and sensory attenuation is the question of how specific both effects are relative to different perceptual events. Here, action-induced blindness operates on precisely those feature codes that are represented as part of an action plan (Müsseler & Wühr, 2002). The blindness effect, therefore, applies only to a clearly defined subset of incoming features, whereas most empirical demonstrations of sensory attenuation can also be explained by nonspecific mechanisms (Horváth, 2015). For instance, focusing on planning and controlling a movement could be assumed to interfere with the processing of task-irrelevant sensory information, such as action-triggered changes in the environment (Horváth et al., 2012). Only attenuation effects that apply to specific features as represented in an action plan can, therefore, be assumed to derive from a similar mechanism as action-induced blindness. Because control conditions to assess such specificity have rarely been included in experiments on sensory attenuation, however, answering this question will require a larger empirical database. A second potentially relevant distinction between action-induced blindness and sensory attenuation pertains to the modality of to-be-perceived events. Whereas blindness effects were obviously probed in the visual modality, sensory attenuation has not yet been shown for behavioral identification of visual stimuli (Schwarz et al., 2018a). These circumstances might also relate to small population effects. TEC’s mechanism of feature integration is not tied to a specific modality, however, so that it would be interesting to investigate conceptually similar effects to action-induced blindness also in other modalities that have previously yielded robust effects of sensory attenuation (i.e., auditory and tactile events).

One participant was excluded, because the experiment was cancelled after the acquisition block, and nine participants were excluded for responding in more than 10% of the no-go trials (in the original study, such participants were replaced with new ones).

The dependent variable of interest in the test phase was choice consistency in go trials. A choice in the test phase was consistent if the response and the auditory stimulus matched the combination from the acquisition phase; otherwise, it was inconsistent. For example, consider the R-E mapping in the acquisition phase where left response → low tone and right response → high tone. If a low tone was presented in the test phase, a left response would count as consistent, while a right response would count as inconsistent. The percentage of consistent choices was calculated per participant and tested against chance with a one-sample t-test. The percentage of consistent responses was used as the individual consistency effect. RTs of consistent and inconsistent choices were further compared with a paired t-test.

In the acquisition phase, participants committed 3.94% anticipations and 0.93% response omissions. Participants pressed the right key in 50.51% (SD = 2.26) of the trials, a slight but significant preference for this response option, t(109) = 2.39, p = .019, d = 0.23 (μ₀ = 50).

In the test phase, anticipations did not occur, but 1.02% response omissions occurred, and participants responded in 2.43% of the no-go trials. Considering only valid (go) trials, participants chose a consistent response in 57.54% of the cases (see Fig. 5B), t(109) = 5.16, p < .001, d = 0.49, 95% CI_SM = [0.29, 0.69] (μ₀ = 50). RTs were 671 ms and 675 ms for consistent and inconsistent choice, respectively, t(109) = 1.09, p = .276, d = 0.10.

This experiment replicated the results obtained by Elsner and Hommel (2001, Exp. 3a). Participants showed a bias toward consistent response choices, as would be expected if (bidirectional) associations between responses and effects have been established during the acquisition phase. The effect size (d = 0.49) was about half of the original effect size (d = 0.94).

While this experiment (as well as Elsner & Hommel, 2001, Exp. 2–4) employed a free-choice test phase and focused on choice rates, other experiments focused on RTs and used a forced-choice test phase in which a particular response was demanded by the stimulus (e.g., Eder & Dignath, 2017; Elsner & Hommel, 2001, Exp. 1; Hoffmann et al., 2009, Exp. 1; Hommel, 1996; Hommel et al., 2003). Responses in such experiments were faster when the stimulus was previously presented as the effect of the required response. Similar results were even obtained with subliminally presented stimuli (Kunde, 2004). In addition, Wolfensteller and Ruge (2011) investigated the time-course of R-E learning and systematically varied the amount of acquisition trials (200 in Elsner & Hommel, 2001). Reliable effects in RTs were obtained after only eight repetitions of R-E episodes already. Finally, a discussion, albeit one going beyond the present paper, has evolved around whether R-E learning occurs only with free-choice acquisition phases (Herwig & Waszak, 2009; Herwig et al., 2007) or also with forced-choice acquisition phases (Pfister et al., 2011).

A particularly relevant question for future work is whether R-E learning generalizes to settings that are more complex than the typical focus on two simple and distinct responses with perfectly contingent effects. Such situations therefore comprise multiple action opportunities (Watson et al., 2015), imperfect R-E contingencies (Elsner & Hommel, 2004), as well as actions that unfold over extended timescales (Land, 2018). Extending the available database on this question might also prove useful for implementing computational approaches to ideomotor learning (Herbort & Butz, 2012).

R1s were incorrect, missing, or anticipated (RT1 < 100 ms) in 1.13%, 1.93%, and 3.52% of the trials. Mean correct RT1 was 413 ms. R2s were missing or anticipated in 0.90% and 0.13% of the trials, and these trials were excluded from further analyses.

The integration of stimulus and response features into event files is at the heart of TEC, and Experiment 4 mirrors earlier reports in showing that re-encountering previous stimulus features retrieves event files, and thus corresponding responses (Hommel, 1998, 2004). Albeit smaller than in the original study, this setup still resulted in a considerably large effect size for task-relevant information (d_z = 1.56 as compared to d_z = 2.70 in the original). The present observations further mirror the original results in showing largest partial repetition costs for task-relevant stimulus features (here: form), and smaller partial repetition costs for task-irrelevant features (here: location and color; d_z ≤ 0.76). Notably, this pattern should indeed be seen as reflecting task-relevance rather than intrinsic properties of different features, because the original publication had already included a control experiment that showed a reversed pattern when color rather than location was implemented as task-relevant (Hommel, 1998, Exp. 2).

In contrast to the original data, however, the present results also yielded small partial repetition costs for the task-irrelevant feature of color. Because the corresponding interaction of color sequence and response sequence had not been significant in the original study, this study concluded (Hommel, 1998, p. 200): “Yet, the indication that colour information is not integrated with information about stimulus or response location suggests that feature binding is selective." The present data suggest that at least the design of the S1R1-S2R2 task rather supports the assumption of unselective binding, a notion that is reinforced by numerous replications of binding between task-irrelevant distractors and responses (Frings et al., 2007; Hommel, 2005; Moeller et al., 2016). Whereas binding seems to be effective by default, the amount of attention devoted to task-irrelevant information certainly affects retrieval of event files, however (Hommel et al., 2014; Moeller & Frings, 2014).

Another noteworthy observation was the fact that several three-way interactions were significant in our study (Form × Color × Location only for RTs, and Form × Color × Response only for PEs). This was not the case in the original study by Hommel (1998), and was interpreted as evidence for local and independent bindings between two features within the event file (see also Hommel, 2007). Similarly, results reported by Giesen and Rothermund (2014, 2016) also argue for independent, local, and binary bindings between relevant and irrelevant stimulus as well as response features within an event file. The present results challenge these conclusions though, albeit it is fair to say that the obtained three-way interaction effects range in the realm of small effect sizes only. This could certainly explain why higher-order interactions were not observed in previous studies with smaller sample sizes. Interestingly, the alternative view, that is, a “uniform-event file hypothesis” (Hommel, 1998, p.188) is also not fully supported by the present data, since the four-way interaction was not significant. Future research should systematically address issues relating to the structure of event files in more detail. Yet, since the higher order interactions seem to come with smaller effect sizes, such research would be well advised to plan sample sizes accordingly.

Mean RT_Bs, MT_Bs, and PE_Bs as a function of overlap are summarized in Table 5. RT_Bs and MT_Bs were longer with overlap compared to without overlap, RT_Bs: t(117) = 7.33, p < .001, d = 0.68, 95% CI_SM = [0.47, 0.87] (see Fig. 7B, C); MT_Bs: t(117) = 6.02, p < .001, d = 0.55. The difference for PE_Bs was not significant, t(117) = 0.74, p = .458, d = 0.07.

Mean IRTs, MT_As, and PE_As as a function of overlap are summarized in Table 5. IRTs and MT_As were shorter with overlap compared to without overlap, IRTs: t(117) = 4.55, p < .001, d = 0.42; MT_As: t(117) = 8.61, p < .001, d = 0.79. The difference for PE_As was not significant, t(117) = – 0.16, p = .875, d = – 0.01.

Mean RT_Bs, MT_Bs, and PE_Bs as a function of overlap and complexity of R_A are summarized in Table 6. The ANOVA on RT_B revealed only a significant main effect of overlap, F(1, 117) = 54.59, p < .001, \(\eta_{{\text{p}}}^{2}\) = .32. Neither the main effect of complexity, F(1, 117) = 0.51, p = .478, \(\eta_{{\text{p}}}^{2}\) < .01, nor the interaction were significant, F(1, 117) = 0.07, p = .797, \(\eta_{{\text{p}}}^{2}\) < .01. For MT_Bs, the main effect of overlap was significant, F(1, 117) = 36.16, p < .001, \(\eta_{{\text{p}}}^{2}\) = .24. MT_Bs were also longer with complex compared to easy conditions, F(1, 117) = 6.71, p = .011, \(\eta_{{\text{p}}}^{2}\) = .05. In addition, the effect of overlap was larger with complex R_A (16 ms) than with easy R_A (12 ms), and the interaction was significant, F(1, 117) = 7.23, p = .008, \(\eta_{{\text{p}}}^{2}\) = .06. For PE_Bs, no effect was significant, overlap: F(1, 117) = 0.56, p = .457, \(\eta_{{\text{p}}}^{2}\) < .01; complexity: F(1, 117) = 0.38, p = .541, \(\eta_{{\text{p}}}^{2}\) < .01; interaction: F(1, 117) = 0.76, p = .385, \(\eta_{{\text{p}}}^{2}\) = .01.

Mean IRTs, MT_As, and PE_As as a function of overlap and complexity of R_A are summarized in Table 6. For IRTs, only the main effect of overlap was significant, F(1, 117) = 20.68, p < .001, \(\eta_{{\text{p}}}^{2}\) = .15. Neither the main effect of complexity, F(1, 117) = 2.07, p = .153, \(\eta_{{\text{p}}}^{2}\) = .02, nor the interaction were significant, F(1, 117) = 3.08, p = .082, \(\eta_{{\text{p}}}^{2}\) = .03. For MT_As, the main effect of overlap was significant, F(1, 117) = 75.72, p < .001, \(\eta_{{\text{p}}}^{2}\) = .39, and, trivially, MT_As were longer in the complex than in the easy condition, F(1, 117) = 2427.78, p < .001, \(\eta_{{\text{p}}}^{2}\) = .95. The interaction was not significant though, F(1, 117) = 0.01, p = .938, \(\eta_{{\text{p}}}^{2}\) < .01. For PE_As, only the main effect of complexity was significant with more errors in easy than in complex trials, F(1, 117) = 12.07, p = .001, \(\eta_{{\text{p}}}^{2}\) = .09. Neither the main effect of overlap, F(1, 117) = 0.03, p = .872, η_p² < .01, nor the interaction were significant, F(1, 117) = 2.70, p = .103, \(\eta_{{\text{p}}}^{2}\) = .02.

As in the original study, responses were initiated more quickly when they did not overlap with an action plan held in memory as compared to situations with overlapping response sides (Stoet & Hommel, 1999). The effect size was, again, smaller for the present results (d = 0.68) compared with the original experiment (d = 1.01). In addition to significant effects for RT_B, but in contrast to the original experiment, MT_Bs were also significantly different as were IRTs and MT_As. Thus, preparing a response appears to bind features that are subsequently less available for other responses. We also additionally included the factor complexity into analyses; yet, these analyses offered no further relevant insights.

Of particular theoretical interest in this kind of experiments is that features of an event file compete between multiple action plans as compared to a single action plan and unrelated perceptual processing. Feature binding can therefore affect the interplay of different actions, and can also bias the choice for one action alternative over another (Richardson et al., 2020). This interplay extends to situations that include even more complex, multi-step actions (e.g., Fournier & Gallimore, 2013). For example, it has been shown that partial overlap with the first response in a sequence A hinders response B execution more than partial overlap with a later part of the sequence (Fournier et al., 2014). Hence, different parts of complex actions seem to be bound into the event file to different degrees. Notably, this is not the case, if the bound responses were individually planned, rather than being part of one more complex action (Moeller & Frings, 2019a; Moeller & Frings, 2019b). The question how longer action sequences are cognitively represented and what the structure of such events looks like is an important one, as it might move theoretical implications from TEC to more natural everyday tasks (e.g., Vallacher & Wegner, 1987; Yamaguchi & Logan, 2014; Zacks & Tversky, 2001; Zacks et al., 2001) and vice versa (see Moeller & Frings, 2021). Yet, since this is still a relatively new area in the literature, much more research is needed, to achieve a clearer understanding of structures within such events and solve questions such as under what circumstances response order becomes part of the event representation. The effect size estimation of our Experiment 5 suggests that researchers can plan with a medium-sized effect in this area.

Based on RT2s, 0.96% of the trials were excluded as anticipations (RT2 < 100 ms) and 0.79% were excluded as response omissions (RT2 > 1500 ms).

Participants pressed the right key as R1 in 55.00% of the trials, a value significantly deviating from 50%, t(104) = 5.82, p < .001, d = 0.57. Response repetitions occurred in 49.72% of the effect-stimulus repetition trials and in 34.51% of the effect-stimulus switch trials. This difference was significant, t(104) = 8.76, p < .001, d = 0.86, 95% CI_SM = [0.63, 1.08] (see Fig. 8B, C).

RT2s in response alternations were shorter (561 ms) than in response repetitions (578 ms), t(104) = 5.02, p < .001, d = 0.49.

The expected higher proportion of response repetitions in congruent compared with incongruent trials was observed and thus this experiment replicated Experiment 1 of Dutzi and Hommel (2009). In contrast to the previous experiments, the effect size (d = 0.86) was closer to the one reported in the original study, albeit still smaller (original d = 1.15). Similar results were obtained in several studies (e.g., Janczyk et al., 2012a; Schwarz et al., 2018b). Previous work further sought to investigate whether short-term action-effect binding is related to S-R binding on the one hand (Moeller et al., 2016, 2019), and to longer term R-E learning on the other hand (Herwig & Waszak, 2012). Especially the relation of short-term binding to stable long-term associations is still poorly understood, however, so that future studies can build on the present effect size estimate.

The bottom line of the following results is straightforward: (1) All phenomena were replicated as in the original study. (2) The resulting effect sizes were (much) smaller though when compared with that reported in the original publications (see Table 7 and also Fig. 2 for a graphical summary).⁸

More precisely, Experiment 1 demonstrated an R-E compatibility effect (Kunde, 2001), usually interpreted as evidence for ideomotor effect anticipation. Experiments 3 and 6 were concerned with learning R-E associations over a longer time course (Exp. 3; Elsner & Hommel, 2001) and binding of R-E events from trial to trial (Exp. 6; Dutzi & Hommel, 2009). Experiments 2, 4, and 5 were concerned with event-file bindings of stimuli and responses. When a (spatial) feature was bound into an event file representing an action plan, concurrent perception was impaired for stimuli requiring the same (spatial) feature (Exp. 2; Müsseler & Hommel, 1997a, 1997b) and the same was true for concurrently executed motor actions (Exp. 5; Stoet & Hommel, 1999). Finally, Experiment 4 confirmed that stimuli and responses are bound into an event file once encountered, impairing subsequent performance when only parts of the implicated features were repeated in a subsequent event (Hommel, 1998).

These results are useful in at least two ways. First, we expect TEC to continue stimulating research in basic and applied areas. For researchers embracing TEC for one or the other reason, the present effect sizes can be used as informed starting points underlying their choice of sample size when conducting power analyses. Table 7 provides a summary of sample sizes required for 1 − β = .80 and 1 − β = .90 when using the simplest analysis tool, that is, a paired t-test. As a cautionary note, these numbers can change when the design is more complex and involves further independent variables (see, e.g., Brysbaert, 2019). Also, the calculated effect size was obtained with one particular experimental setting suited for data collection in the laboratory in this present study. Other settings may well yield larger effect sizes and thus allow a smaller sample size as well (see, e.g., the discussion of Exp. 1). When performing power analyses, it is thus useful to choose effect sizes from experimental settings that resemble the planned study as closely as possible. The values provided in Table 7 provide rules of thumb when an informed intuition about the size of the expected effect cannot draw on similar previous work.

The second contribution is that all six experiments replicated the main results reported in the original studies. Against the background of attempts to replicate certain psychological phenomena in the past (e.g., Open Science Collaboration, 2015), this outcome is certainly positive and encouraging news for the field. Considering recent calls for integrative frameworks going beyond phenomena-specific explanations (Eronen & Bringmann, 2021; Muthukrishna & Henrich, 2019; Oberauer & Lewandowsky, 2019) and cumulative theory building, TEC appears well-suited as being such a framework, backed up by solid evidence for core phenomena derived from it.

In our view, future research on TEC can extend the state of the art in two complementary directions that we label as vertical or horizontal.

With vertical extensions, we refer to research that aims at a deeper understanding of particular TEC-related phenomena with regard to potential underlying mechanisms. While TEC provides a coherent framework for understanding diverse empirical phenomena, many of these phenomena can also be captured by alternative accounts that do not necessarily accord with TEC’s perspective. For instance, findings on R-E learning (Exp. 3; Elsner & Hommel, 2001) have been suggested to reflect strategic choices (Vogel et al., 2018; Weller et al., 2017), possibly relying on propositional representations rather than direct action-effect associations (Sun et al., 2022). Having a precise grasp on the size of individual effects will allow for more stringent tests of potential alternative explanations, thus contributing to theoretical progress.

Horizontal extensions, by contrast, refer to the interrelation of different empirical effects. A major strength of TEC is its broad applicability across diverse phenomena in perception and action. With the present combined assessment of several empirical phenomena, we hope to contribute to research in this tradition. That is, the present re-assessment of selected phenomena should be especially valuable for building bridges between design-specific effects, possibly towards areas that are not commonly discussed in relation to TEC.

One example that would benefit from such cross-design approaches concerns the relation of action-induced blindness (Exp. 2; see Müsseler & Hommel, 1997a, 1997b) and sensory attenuation (Blakemore et al., 2000) as argued in the discussion of Experiment 2. Addressing how impaired perceptual processing either immediately before or immediately after performing an action builds on potentially similar mechanisms might allow for a more parsimonious explanation, while also highlighting shared mechanisms of TEC and common models that are invoked to explain sensory attenuation, especially forward and comparator models (Dogge et al., 2019; Horváth, 2015). Such research would also inform communities that have mainly relied on specific perceptual and response modalities, as is apparent in the decidedly visual focus of action-induced blindness, while at the same time drawing on modality-unspecific mechanisms to explain central observations. Finally, this strategy will also include the assessment of specific populations and corresponding inter-individual variation of certain phenomena in both clinical and non-clinical settings.

A second example concerns the relation of code occupation effects (Exp. 5, see Stoet & Hommel, 1999) and experiments that address the role of anticipated action effects for switching between different responses (Kunde et al., 2002; Mocke et al., 2020). On the surface, both strands of research seem to employ highly similar setups. Code occupation is commonly studied by having participants plan an action while performing an intermediate action that either shares features with the planned action or does not share any relevant features, with shared features yielding overlap costs. The alternative strand of research is identical in the sense that participants plan a certain movement, but have to perform a separate response while holding the initial action plan active. These studies also implement overlap conditions, but they do so by having responses produce either compatible or incompatible action effects in the actor’s environment, such as distinctive effect tones rather than in terms of body-related features as in typical setups that probe for code occupation. Strikingly, despite the highly similar setup, these latter studies have consistently found overlap benefits rather than overlap costs (see also Janczyk & Kunde, 2014). This change in direction of the effect appears puzzling and, therefore, requires future, cumulative evidence to arrive at a satisfactory account of both phenomena.

A third example of TEC’s broad applicability across different empirical phenomena is illustrated by recent trends to study well-established cognitive effects in social settings, that is, when two or more individuals engage together in a cognitive task. Interestingly, many experiments now come with a social variant (e.g., the joint Simon task, Sebanz et al., 2003; joint flanker task, Atmaca et al., 2011; social inhibition of return, Janczyk et al., 2016; observational R-E associations, Paulus et al., 2011; Pfister et al., 2014a; observational S-R binding, Giesen et al., 2014, to name just a few). On the one hand, these developments illustrate that these well-established effects are not immune to social influences, an insight that is only recently recognized by cognitive psychologists. On the other hand, the findings from these new approaches set the stage to introduce the terminology of TEC to account for social effects. This allows for re-assessing prominent effects known from social psychology and explaining them at least partly from TEC’s perspective (for recent examples, see Giesen et al., 2014, 2017, 2018; Hommel & Colzato, 2015; Hommel & Stevenson, 2021).

In sum, the present re-assessment of selected phenomena should be especially valuable for building bridges between design-specific effects, possibly towards areas that are not commonly discussed in relation to TEC. Not surprisingly, in the two decades after their introduction, the basic ideas of TEC have been further developed and extended beyond those paradigms closely associated with TEC. For example, the recent Binding and Retrieval in Action Control (BRAC) framework explicitly separates integration (i.e., binding) and retrieval processes theoretically and proposes that these two processes affect action control also in many standard experiments, typically not closely associated with event coding (e.g., negative priming, task switching, Pavlovian conditioning, visual search; Frings et al., 2020). Such broad applicability of the processes involved in the current experiments, is another reason to expect continued research interest in TEC-related phenomena in the future.