Agency effects on the binding of event elements in episodic memory

Representing events in episodic memory in a coherent manner requires that their constituent elements are bound together. So far, only few moderators of these binding processes have been identified. Here we investigate whether the presence of an agentic element in an event facilitates binding. The results from six experiments provided no evidence for a facilitating effect of agency on the binding of event elements. In addition, binding effects were only found when event elements were presented simultaneously, but not when they were presented sequentially pairwise, contrary to previous findings. The results suggest that the presence of an agentic element in an event does not, or only to a very limited extent, contribute to the formation of coherent memory representations and that additional processes may be required when binding event elements across temporarily divided encoding episodes. These findings add to a growing body of research regarding moderators and processes relevant for the binding of event elements in episodic memory. Explanations of these findings and directions for future research are discussed.


Design
The experiment employed a one-factorial (agency condition: agency vs. non-agency) within-subjects design.In the agency condition one event element served as the agent and was placed as the grammatical subject in active sentences.In the non-agency condition there was no agent and only passive sentences were used.
An additional 24 nouns, 8 of each type, and 8 verbs were used as primacy buffers.Stimuli were partly taken from Schreiner et al. (2023).We used three types of objects to avoid confounding with animacy (cf.Schreiner et al., 2023).Using the stimuli, we randomly created 48 events for each participant, each consisting of a means of transportation, a tool, a food, and a verb.Events were randomly assigned to the 2 experimental conditions, resulting in 24 events per condition and 4 primacy buffer events per condition, which were presented first.

Procedure
The procedure was based on the separated encoding paradigm (Horner et al., 2015;Horner & Burgess, 2014).In the learning phase of the experiment, participants were presented a sentence containing two event elements and the verb associated with the event in each trial.There were three learning trials (i.e., sentences) for each event.Sentences referring to the same event were presented interleaved with sentences referring to other events.The presentation order was randomized with the constraint of a minimum of two other-event trials being presented between two same-event trials.The experimental conditions were randomly distributed across learning trials.In the agency condition active sentences were used if the sentence contained the agent and passive sentences were used if it did not (e.g., The bicycle grabs the hammer.,The bicycle grabs the apple., and The hammer and the apple are being grabbed.).Each stimulus type (i.e., means of transportation, tool, or food) served as the agent equally often across events.In the non-agency condition only passive sentences were used (e.g., The bicycle and the hammer are being grabbed.,The bicycle and the apple are being grabbed., and The hammer and the apple are being grabbed.).
In passive sentences it was randomized which of the event elements appeared in the first sentence position.Thus, all possible pairwise associations between event elements were shown across the three learning trials referring to the same event.Encoding episodes were thus coherent (using closed-loop structures, cf.Horner et al., 2015;Horner & Burgess, 2014).Event elements in the agency condition were defined in terms of whether they were the agent or one of the non-agents in an event, yielding the associations agentnon-agent 1 , agent -non-agent 2 , and non-agent 1 -non-agent 2 .Event elements in the non-agency condition were defined in terms of their stimulus type, yielding the associations means of transportation -tool, means of transportation -food, and tool -food.Each trial consisted of a 0.5-s fixation cross, a 6-s sentence presentation, and a 1.5-s blank screen.
The experiment included two attention checks.After 50% of learning trials (not counting primacy buffers) participants were asked to click on a continue button within 10 s and after 50% of test trials they were asked to select the top left response option.Other than that, the procedure was identical to the one of Experiment 4, described in the main article.

Data Analysis
Data analysis was identical to the one conducted in Experiment 4, described in the main article, with the following exceptions: For the exploratory analysis of memory performance, agency condition was used as a between-instead of a within-subjects factor in the Bayesian generalized linear mixed models.Consistent with our approach in Experiment 4 we coerced the associations agent -non-agent 1 and means of transportation -tool, agent -non-agent 2 and means of transportation -food, and non-agent 1non-agent 2 and tool -food into a common factor level, respectively.For the dependency analysis, we fit a joint IRT model to the data of both agency conditions, since agency was manipulated within-subjects.

Dependency
The dependency of the retrieval of event elements is shown in Figure S.2.The dependency in both the agency condition (D = 0.04, SE = 0.02, p = .05)and non-agency condition (D = 0.01, SE = 0.02, p = .61)was non-significant.The dependency in the agency condition was not significantly larger than the one in the non-agency condition (D diff = 0.03, SE = 0.03, p = .13).

Discussion
In Experiment 1 we tested Hypothesis 1, which states that there is a stronger stochastic dependency of the retrieval of event elements for events with an agentic element than for events without an agentic element.There was no significant dependency in both the agency and non-agency condition and thus, the results are not informative regarding the hypothesis.However, descriptively, there was a positive dependency in the agency condition that was larger than the dependency in the non-agency condition.Thus, the non-significant findings may be a Type II error or the effects may be rather small.In Experiment 2 we aimed for a larger sample size and thus for a higher power for detecting dependencies.

Experiment 2
In Experiment 2 we again investigated whether there is a stronger stochastic dependency of the retrieval of event elements for events with an agentic than for events without an agentic element (Hypothesis 1).In addition, we investigated the binding structure of event elements by testing an integrated against a hierachical binding structure.
To this end, we extended the experimental design to include non-coherent encoding episodes (open-loop structures, see Horner et al., 2015;Horner & Burgess, 2014)  suggests that dependency does not vary as a function of the excluded association, because event elements are stored in a unitary representation that is accessible in an all-or-none manner (e.g., Damasio, 1989;Marr, 1971;Tulving, 1983).Thus, the association that is not presented should either be readily retrieved with the other associations (and thus dependency should be the same no matter which association is excluded) or the non-coherence of the encoding episodes may prevent integration and there should be no dependency in all open-loop conditions (and thus also no variation across open-loop conditions).A hierarchical binding structure suggests a system of pairwise bindings (see e.g., Cohen & Eichenbaum, 1993;Eichenbaum, 1999;Healy & Caudell, 2019) with asymmetrical binding strengths, in which event elements are preferentially bound to particular types of elements, such that some bindings are systematically prioritized over others (see Schreiner et al., 2023).Thus, excluding a more critical association should more strongly diminish dependency than excluding a less critical association, and consequently dependency should vary as a function of the excluded association.We expected that event elements are preferentially bound to the agent of the event, and thus expected to find a stronger stochastic dependency of the retrieval of event elements when excluding associations not involving an agentic element than when excluding associations involving an agentic element in non-coherent encoding episodes (Hypothesis 2).For events without an agentic element we expected no differences in the stochastic dependency of the retrieval of event elements in non-coherent encoding episodes (Hypothesis 3), because in this case all associations should be equally important.The experiment's design, hypotheses, and analysis plan were preregistered at https://osf.io/kts8p.

Participants
Participants were again recruited from the Web, using the same channels as in Experiment 1.They could join a lottery for winning vouchers of a total value of 470€ or receive course credit.An a priori power analysis with simulated data for detecting the predicted results pattern with small to medium differences between conditions (difference in event-specific trait variances of 0.75 according to the statistical procedure, cf.Glas et al., 2000;Wang et al., 2002, assumed  Thus, in the agency condition there was one condition in which the association agentnon-agent 1 was excluded (OL-AgNAg 1 ), one condition in which the association agentnon-agent 2 was excluded (OL-AgNAg 2 ), and one condition in which the association non-agent 1 -non-agent 2 was excluded (OL-NAg 1 NAg 2 ).In the non-agency condition there was one condition in which the association means of transportation -tool was excluded (OL-TrTo), one condition in which the association means of transportation -food was excluded (OL-TrFo), and one condition in which the association tool -food was excluded (OL-ToFo).The open-loop conditions were equated to the closed-loop conditions regarding the number of event elements rather than the number of presented associations.Previous studies yielded similar results irrespective of whether conditions were equated regarding the number of event elements or the number of presented associations (Horner & Burgess, 2014;Joensen et al., 2020).

Material and Procedure
Stimuli were identical to the ones of Experiment 1. Events were randomly assigned to the eight experimental conditions, resulting in six events per condition and one primacy buffer event per condition.The experimental procedure was identical to the one of Experiment 1, except that open-loop conditions were included in addition to closed-loop conditions.While event presentation in the closed-loop conditions consisted of three learning trials, event presentation in the open-loop conditions consisted of two learning trials.For the open-loop conditions, test trials included one trial per event with a cue-target pair that was not presented jointly in the learning phase (inference trials).
However, the respective cue and target overlapped with a common event element and could thus be flexibly related to enable reconstruction of the association that was excluded from presentation in the learning phase.

Data Analysis
Data analysis was identical to the one of Experiment 1 with the following exceptions: For the exploratory analysis of memory performance, loop condition was included as an additional predictor.Thus, there were three possible two-way interactions and one possible three-way interaction.To investigate the three-way interaction we compared the full model including all main effects and interactions with a model containing all main effects and two-way interactions but no three-way interaction.Similar to the handling of the factor association we coerced loop conditions into common factor levels to jointly include them in the models for the analysis of memory performance.To further investigate interactions, we conducted post-hoc pairwise comparisons using the package emmeans (version 1.8.6, Lenth, 2022).We considered a difference to be substantial if the 95% credible interval (highest posterior density interval) does not include zero.

Dependency
The dependency of the retrieval of event elements in loop conditions CL is shown in

Discussion
In Experiment 2 we again tested Hypothesis 1.In addition, we tested an integrated against a hierarchical binding structure.Since we did not find a significant dependency in any of the experimental conditions, the results are uninformative regarding the hypotheses and cannot properly distinguish between an integrated and a hierarchical binding structure.It is noteworthy that memory performance was even lower than in Experiment 1 and memory performance in both Experiment 1 and 2 was lower than, for example, memory performance in the experiments by Schreiner et al. (2023).Since low memory performance is associated with lower power for detecting dependencies and differences in dependencies (Schreiner & Meiser, 2023), in Experiment 3 we made some changes to the experimental design and procedure intended to improve memory performance.

Experiment 3
In Experiment 3 we again investigated whether there is a stronger stochastic dependency of the retrieval of event elements for events with an agentic than for events without an agentic element (Hypothesis 1).We introduced a number of changes to the design and procedure of Experiment 1 (and the closed-loop conditions of Experiment 2) intended to improve memory performance to achieve higher power for detecting binding effects (cf.Schreiner & Meiser, 2023).We changed the experimental design from a withinto a between-subjects design while keeping the number of events per condition identical to model in the non-agency condition) with values randomly drawn from the empirical distribution of the remaining parameters.

Memory Performance
On average, the proportion of correct responses was M = 0.27 (SD = 0.44) in the

Discussion
In Experiment 3 we again tested Hypothesis 1.However, there was again no significant dependency in both the agency and non-agency condition and thus, the results are not informative regarding the hypothesis.In addition, the changes made to the experimental procedure and design in Experiment 3 did not lead to an increase in memory performance compared to Experiments 1 and 2. It is possible that the more diverse sample counteracted effects of the changes of the experimental design and procedure.Compared to Experiments 1 and 2 the sample in Experiment 3 was older and comprised more males and fewer students.

Figure
Figure S.1 Raincloud Plot Depicting the Proportion of Correct Responses per Participant by Agency Condition in Experiments 1-3

Figure S. 2
Figure S.2Dependency of the Retrieval of Event Elements by Agency Condition in Experiments 1-3 in addition to the coherent encoding episodes (closed-loop structures) that were used in Experiment 1.While in closed-loop structures all possible pairwise associations between event elements are shown, we consistently excluded specific associations from presentation in the open-loop structures (cf.Schreiner et al., 2023).Integrated and hierarchical binding structures make different predictions regarding the pattern of dependency across open-loop structures in which different associations are excluded.An integrated binding structure baseline event-specific trait variance of 1) with 80% power using one-tailed testing yielded a desired sample size of 240 participants.Due to the potential necessity of some data exclusion we increased the desired sample size by 20%, thus planning for a sample size of 288 participants.Because we needed to exclude more participants than anticipated we collected data from an additional 90 participants.Thus, we collected data of 378 participants.All participants provided online informed consent for their participation and publication of their data.We excluded three participants from the analyses because they indicated not speaking German fluently.Another 89 participants were excluded because they did not pass both attention checks.Another seven participants were excluded because they suggested their data should not be used for the study (e.g., due to distractions) 1 .Another 18 participants were excluded because they conducted the study on a smartphone on which a correct display of the experiment content could not be guaranteed.Another six participants were excluded because they indicated having recently participated in a similar study (i.e., Experiment 1).An additional 13 participants were excluded because their browsing behavior suggested they interrupted the experiment frequently and for a longer duration.Thus, the final sample consisted of 242 participants (181 [75%] female, 1 [0.4%] non-binary, 198 [82%] students) with an average age of 27.1 years (SD = 9.6, range = 18-68).DesignThe experiment employed a 2 (agency condition: agency vs. non-agency) × 4 (loop condition: closed-loop and three open-loops) within-subjects design.The closed-loop conditions (CL) were identical to the conditions in Experiment 1.In each of the three open-loop conditions we consistently excluded one pairwise association from presentation.