Enhanced Memory for Context Associated with Corrective Feedback: Evidence for Episodic Processes in Errorful Learning

Amy A Overman; Joseph D W Stephens; Mary F Bernhardt

doi:10.1080/09658211.2021.1957937

. Author manuscript; available in PMC: 2022 Sep 1.

Published in final edited form as: Memory. 2021 Jul 26;29(8):1017–1042. doi: 10.1080/09658211.2021.1957937

Enhanced Memory for Context Associated with Corrective Feedback: Evidence for Episodic Processes in Errorful Learning

Amy A Overman ¹, Joseph D W Stephens ², Mary F Bernhardt ³

PMCID: PMC8440464 NIHMSID: NIHMS1729192 PMID: 34309487

Abstract

Numerous studies have established that there are benefits of corrective feedback for learning, but the mechanisms of this benefit are not well understood. An important question is whether corrective feedback improves memory via episodic processes or solely via semantic mediation. If episodic processes are involved, then memory for corrective feedback should include contextual details of the feedback episode. The present study tested this hypothesis across 3 experiments (total n = 223) in which participants completed an encoding task that involved cued guessing of category exemplars. Exemplars generated by participants were equally likely to be treated as correct or incorrect, and the “correct” exemplar was presented within a feedback display after each response. Separate versions of the task manipulated font color in either the feedback display or the initial cue/typed response display. Participants were instructed to remember either the correct exemplars or their own typed responses, and the corresponding font colors. Retrieval task (cued recall, free recall, recognition) was varied across experiments. Across all 3 experiments, a higher rate of memory accuracy was observed for context associated with corrective feedback relative to other conditions. The findings are consistent with the hypothesis that errorful learning involves episodic memory, not merely semantic mediation.

Keywords: Memory, Generation, Corrective Feedback, Context

A large and growing body of evidence indicates that wrong answers are good for learning – as long as corrective feedback is provided (see Metcalfe, 2017, for a thorough review). A substantial amount of research on corrective feedback in learning has focused on feedback’s relevance to retrieval-based learning, in which testing of recently learned information confers memory advantages compared to re-studying of the information (also referred to as testing effects; Roediger & Karpicke, 2006). Feedback during testing may enhance testing effects (e.g., Kang, McDermott, & Roediger, 2007; McDaniel & Fisher, 1991), and corrective feedback appears to be more beneficial for item memory than confirmatory feedback is (Hays, Kornell, & Bjork, 2010; Pashler, Cepeda, Wixted, & Rohrer, 2005). Corrective feedback is particularly important in variations of the testing effect where learners – usually unsuccessfully – attempt to provide responses for items that have not been previously studied, also known as pretesting (Kornell, Hays, & Bjork, 2009; Kornell & Vaughn, 2016). In this case, corrective feedback essentially acts as the study episode, and produces superior memory relative to simply reading the to-be-remembered information without attempting to generate a response. This advantage of pretesting can also be influenced by the timing of the feedback (Hays, Kornell, & Bjork, 2013; but see Kornell, 2014).

The cognitive mechanisms that support errorful learning are not fully understood. One proposal is that the benefit of error correction represents a specific case of semantic mediation in retrieval-based learning (Pyc & Rawson, 2010; Carpenter, 2011; Carpenter & Yeung, 2017). According to the semantic mediation (or semantic elaboration) hypothesis, retrieval attempts during testing produce a form of elaboration in which concepts are activated that are semantically related to the correct answer and can serve as mediators linking the cue to the correct answer on subsequent retrieval attempts. In errorful learning, a similar process of elaboration may take place, resulting in an incorrect response, which is then linked to the correct response via feedback. Hence, the initial incorrect response acts as one of the mediating concepts to produce the correct response on a later test. Some prior studies have obtained results consistent with this account (Butler, Fazio, & Marsh, 2011; Knight, Ball, Brewer, DeWitt, & Marsh, 2012; Vaughn & Rawson, 2012; but see Clark, 2016, which did not find evidence for feedback acting as a mediator). A number of researchers have argued that semantic mediation is insufficient to account for the benefits of retrieval-based learning, including benefits of error generation and corrective feedback. Alternative accounts have argued that episodic memory processes, rather than semantic network activation, play a central role in retrieval-based learning. For example, Karpicke, Lehman, and Aue’s (2014) episodic context account of retrieval-based learning proposes that the key mechanism contributing to benefits of retrieval practice is the updating of episodic context features in the memory trace for the retrieved item. This account ties in with broader theories of memory updating, such as the memory for change framework (e.g., Wahlheim & Jacoby, 2013), which posits that the recollection of prior episodic details enables the incorporation of prior and current episodes into a combined representation that resolves interference and enhances memory for current information. With regard to feedback specifically, Metcalfe and Huelser (2020) have recently argued that the benefits of learning from corrective feedback are based on episodic updating mechanisms. They reported results from two experiments that suggested: 1) error correction was not better when errors were semantically related (versus unrelated) to correct answers, arguing against semantic mediation; and 2) error correction was better when erroneous responses could be recalled, arguing for episodic updating.

Additional support for the importance of episodic processes, as opposed to semantic mediation, in errorful learning comes from evidence that corrective feedback is simply encoded better than other information presented during a learning trial (enhanced encoding of feedback hypothesis; Potts, Davies, & Shanks, 2019). A key finding in this regard is the “hypercorrection effect” (Butterfield & Metcalfe, 2001) whereby error correction is better for high-confidence errors than low-confidence errors. This has been used to suggest that corrective feedback – particularly when it is surprising to learners that they are being corrected – attracts learners’ attention in a manner that directly benefits subsequent retrieval of the feedback, without the need for semantic mediation.

If learning via corrective feedback involves episodic processes, then its benefits might be expected also to apply to memory for other aspects of the learning episode, such as contextual details. Contextual details are an essential component of episodic recollection (e.g., Tulving, 1972), and play a key role in memory-updating accounts of retrieval-based learning such as that of Karpicke et al. (2014). Contextual information is essential to source memory (Johnson, Hashtroudi, & Lindsay, 1993), which is an important aspect of learning in both everyday and educational settings. Therefore, it is critical to our understanding of errorful learning to know the extent to which error generation and corrective feedback influence memory for contextual details of the learning episode. If corrective feedback about target items also influences memory for contextual details, it would imply that episodic processes are involved in errorful learning. On the other hand, if errorful learning is solely based on semantic mediation, then feedback should have no effect on memory for contextual details of the study episode.

Some evidence exists to suggest that testing effects extend to memory for contextual information (Akan, Stanley, & Benjamin, 2018; Rowland, 2011), and that errorful learning enhances source memory, relative to errorless learning (Cyr & Anderson, 2012). However, to our knowledge only one prior study (Fazio & Marsh, 2009) has directly examined memory for contextual information associated with feedback. Fazio and Marsh found that memory for context associated with feedback was related to participants’ confidence in their responses. That is, contextual details (i.e., font color, voice gender) of corrective feedback were remembered better when participants had high confidence in their erroneous answers, and contextual details of confirmatory feedback were remembered better when participants had low confidence in their correct answers. Their findings were consistent with the idea that hypercorrection (Butterfield & Metcalfe, 2001) applies to both the item and context information contained in feedback. However, Fazio and Marsh’s study had some limitations: it did not compare memory for contextual details of feedback to memory for contextual details associated with other aspects of the learning trial, and did not test for general effects of corrective or confirmatory feedback on context memory, independent of confidence or surprisingness. Additionally, that study used general-knowledge items as stimuli, which made it impossible to randomly assign items to correct or incorrect response conditions.

It should be noted that, to the extent episodic processes are involved in learning from feedback, the effects on context memory might not necessarily be all positive. For example, studies of the generation effect have found that although self-generation reliably improves memory for generated items, it does not necessarily improve memory for context associated with those items. That is, contextual features of items (such as font colors of words) are sometimes remembered more accurately when items are passively studied than when they are generated, constituting a “negative generation effect” (Jurica & Shimamura, 1999). Careful investigations of these negative generation effects on context memory (e.g., Mulligan, 2004; Mulligan, 2011; Mulligan, Lozito, & Rosner, 2006) have established that they tend to occur when the context features to be remembered involve cognitive processes in a different domain than that of the generation task. For example, memory for a visual feature, such as font color, is more negatively affected when a semantic generation task is used; likewise, the negative effect on memory for context can be reversed when to-be-remembered context features are aligned to the generation task (for example, memory for auditory features is enhanced by use of a rhyming generation task; Overman, Richard, & Stephens, 2017). Similar trade-offs could apply to processing of feedback information, whereby processing of corrective feedback about a semantically-generated item might inhibit processing of perceptual details of that item. Alternatively, increased attentional focus on the feedback display might enhance encoding of both meaningful and incidental aspects of the display, similar to the attentional boost effect in memory for incidental information during a target detection task (Spataro, Mulligan, & Rossi-Arnaud, 2013; Swallow & Jiang, 2010).

In summary, to build a more complete theoretical framework of the mechanisms involved in errorful learning, it is important to empirically establish the degree to which context memory is influenced by conditions of item generation and corrective or confirmatory feedback. The present study addressed this knowledge gap by examining memory for contextual features (specifically, font color) of feedback in a learning task with cue-target associations that consisted of semantic categories and exemplars. Importantly, the learning task was arranged so that exemplars generated by participants were equally likely to be treated as correct or incorrect responses, and category cues were randomly assigned to incorrect and correct generation conditions. This allowed for direct comparison of the memory effects of corrective versus confirmatory feedback, while eliminating any inherent differences between correct and incorrect items, as well as any basis for expectancies by which the two types of feedback could differ in their surprisingness. Additionally, context memory was compared between versions of the task in which the target exemplars to be remembered were the correct items presented at feedback versus participants’ own typed responses. This provided a critical control condition, making it possible to test whether differences in context memory across generation and feedback conditions were specific to context associated with the feedback, compared to context associated with other parts of the encoding episode. This comparison also helped to rule out the possibility that incorrect generation trials differed from correct generation trials in some systematic way other than in the feedback portion of the trial. A further benefit of including experiment versions in which participants remembered their own responses was that it provided conditions that more directly overlapped with prior studies of generation effects in context memory (e.g. Mulligan, 2004).

Three experiments were conducted, in which the item retrieval task was varied (Experiment 1: cued recall; Experiment 2: free recall; Experiment 3: recognition). The use of different retrieval tasks provided for consideration of the relative contributions of encoding and retrieval processes to the effects of corrective and confirmatory feedback on context (and item) memory performance.

Experiment 1

As described above, Experiment 1 tested the effects of corrective and confirmatory feedback on item and context memory, using cue-target pairs that consisted of common category names and exemplars. Two versions of the experiment were conducted, which differed only in whether the target exemplars to be remembered (and whose font colors varied) were the “correct” items presented at feedback (Experiment 1a), or the participants’ own typed responses (Experiment 1b).

Method

Participants

A total of 69 undergraduates (mean age = 18.75 years) participated and were compensated with course credit. Thirty-one participants completed Experiment 1a, in which the to-be-remembered item and context information was presented during feedback, and 38 completed Experiment 1b, in which the to-be-remembered item and context information consisted of their initial responses to stimuli. The two versions of the experiment were conducted in different semesters, with Experiment 1a conducted first (this was also the case for Experiments 2 and 3). Thus the assignment of participants to the two versions was not completely random. Nonetheless, it was assumed that there were no meaningful differences in the population of available participants across the semesters in which the two versions were run. All participants provided informed consent and all procedures were approved by the Institutional Review Board of Elon University.

Sample size was based on prior studies that have investigated context memory across encoding conditions that included generating and reading items (e.g., Mulligan, Lozito, & Rozner, 2006; Overman, Richard, & Stephens, 2017). Those studies reported effect sizes for within-subjects comparisons of context memory measures across encoding conditions of d = .51, d = .53, and d = .67. A priori power analysis indicated that for an effect size of d = .53, a sample size of n = 30 would yield estimated power of .8. Thus, participants were recruited with the goal of having at least 30 participants with usable data in each version of the experiment.¹ Extra participants were also recruited in Experiment 1b due to concerns that a subset of seven participants may have performed the cued recall task incorrectly (see below). This enabled some analyses to be carried out with those participants excluded, while still having data from more than 30 participants in each version of the experiment.

Materials

Stimuli were based on 46 taxonomic categories selected from the category norms of Van Overschelde, Rawson, and Dunlosky (2004). For each category, two exemplars were selected to create a pool of available exemplars for use in the experiment. Exemplars were chosen such that they were among the most commonly named members of each category but did not create potential conflicts or confusion with exemplars from other categories (for example, orange was not used since it could be either a fruit or a color of the rainbow). A complete list of category names and selected exemplars is provided in the Appendix.

Procedure

Both versions of the experiment consisted of an encoding task followed by a three-minute backwards-counting task and a cued recall task. The encoding task included three conditions: Read, Correct Generation, and Incorrect Generation. There were 14 trials in each condition, with 42 of the 46 categories randomly assigned to encoding conditions for each participant. Four of the categories were used in buffer trials at the beginning of the encoding task and were not tested in the cued recall task.

Encoding task.

An illustration of the encoding task is provided in Figure 1. On each trial, the participant was presented with a category name, and typed the name of a member of that category. In the Read condition, the category name was presented with one of the pre-selected exemplars below it. The participant re-typed the exemplar, with typed letters viewable in an echo box at the bottom of the screen, and pressed Enter. After the participant pressed Enter, the category-exemplar pair remained on the screen for 1000 ms, after which the font color of the category name and exemplar changed, and the display remained on the screen for an additional 5000 ms.

*Note.* Participants were cued with categories and instructed to type names of category members. Participants either re-typed an item that was presented along with the cue (Read condition), or typed a self-generated item (Correct and Incorrect Generation conditions). The feedback display then presented the “correct” response. On half of generation trials, the participant’s own response was treated as correct (Correct Generation). The other half of generation trials treated the participant’s response as incorrect and selected a different category member to display as feedback. In Experiment 1a, blue or yellow font color was used in the feedback display; in Experiment 1b, blue or yellow color was used in the initial typed item display.

In the generation conditions, only the category name was presented at the beginning of the trial, with a blank space below it (a continuous underscore). The participant was instructed to “guess which category member is supposed to go in the space. Then, type the word into the text box and press Enter.” As in the read condition, typed letters were viewable in an echo box at the bottom of the screen. After pressing Enter, the category name disappeared and the screen displayed either “Correct!” (in the Correct Generation condition) or “Incorrect” (in the Incorrect Generation condition) in white font for 1000 ms. The category name then reappeared with the correct exemplar below it for 5000 ms. The correct exemplar shown in the feedback display was either the participant’s own typed response (in the Correct Generation condition) or a non-matching exemplar from the pre-selected list (in the Incorrect Generation condition). Thus, exactly half of the participant’s responses were labeled as incorrect, regardless of what the participant typed.² As noted by Yan, Yu, Garcia, & Bjork (2014), this type of design makes it impossible for participants to form any valid strategy for guessing correct responses. It also ensured equal numbers of trials with confirmatory versus corrective feedback, and removed any basis for expectancy or surprise based on prediction of the correct response. After the initial four buffer trials, the encoding list was arranged in seven blocks, each of which contained the six possible combinations of encoding condition and font color. Trial order was randomized within each block.

The key difference in the encoding phase between Experiments 1a and 1b was the timing of when colored font was used in the display. In Experiment 1a, the initial typing prompt display was presented in white font (including the border of the echo box), and the feedback display of the category name and correct exemplar was presented in either yellow or blue font. In Experiment 1b, the initial typing prompt display was presented in yellow or blue font (including the border of the echo box), and the feedback display was presented in white font. In both versions of the experiment, participants were instructed as to what they would be tested on later. That is, in Experiment 1a, they were informed that they should try to remember correct category members and their font colors. In Experiment 1b, participants were still given the task of attempting to name the “correct” category member and receiving feedback, but they were also informed that they should try to remember their own typed responses, along with the corresponding font colors.

It may be noted that in the Correct Generation and Read conditions, the same items were studied twice (at initial response and at feedback), but with different contextual information each time, since the blue or yellow font color occurred for only one of those study instances and white font was used for the other. For Incorrect Generation, on the other hand, the target item (i.e., feedback item in Experiment 1a and typed item in Experiment 1b) was only ever studied in blue or yellow font. This difference might have the potential to create an unintended advantage in context memory for the Incorrect Generation condition, since there was no competing contextual information from seeing the target item in white font. For this reason, the comparison between Experiments 1a and 1b is especially important, because any such advantage should apply equally to both versions. If the two versions differ in context memory performance in the Incorrect Generation condition, it cannot be attributed to any difference in competing contextual information associated with target items.

Cued recall task.

During each trial in the test phase, participants were presented with a category name, and attempted to type a target category member from the encoding task. All 42 non-buffer category names were presented in random order. In Experiment 1a, the target category member that participants were instructed to recall was the correct item, i.e., the item that had been presented in the feedback display. In Experiment 1b, the target category member that participants were instructed to recall was the one that they had typed, regardless of whether it had been marked as correct or incorrect. After the participant typed a cued recall response and pressed Enter, that response was immediately presented again, and the participant was prompted to indicate whether the item had been presented in blue or yellow font at encoding, using the 7 and 8 keys, respectively, on the keyboard.

Results

Analysis approach

For all experiments reported in this paper, two sets of analyses were conducted. The first set consisted of separate, linear-model based analyses of accuracy in item and context memory. For context memory, accuracy was measured in terms of identification-of-origin (or I-O) scores (e.g., Johnson et al., 1993). The second set of analyses constructed multinomial processing tree (MPT) models of participants’ response patterns based on assumed underlying cognitive processes that include item retrieval, context retrieval, and guessing (Batchelder & Riefer, 1999). In this type of model, item and context memory are represented by different parameters, and memory effects can be examined by comparing corresponding parameters across different conditions, and by comparing the ability of models to fit the data when parameters are allowed to vary versus not to vary across conditions.

For the linear-model based analyses, trial-by-trial response data were dichotomous in nature; i.e., participants either retrieved or did not retrieve a target item, and either correctly or incorrectly identified the font color of a retrieved item. Accuracy in item and context memory for a participant was thus reflected by the proportion of trials on which a target item was retrieved, and on which a target item’s font color was correctly identified, respectively. Because context accuracy was conditional upon memory for target items, its estimation relied on varying numbers of trials across participants. Similar data in the existing literature have often been analyzed via classical analysis of variance techniques; however, for the present study, a hierarchical Bayesian analysis of variance (BANOVA) approach was used, which avoids the pitfalls associated with categorical dependent variables and unbalanced designs in classical ANOVA (e.g., Gelman, 2005; Kruschke, 2015), while retaining ANOVA’s convenient framework for interpreting the effects of multiple independent variables and their interactions. The BANOVA package in R (Dong & Wedel, 2016) provides a straightforward implementation of this approach. Thus, for the three experiments reported here, both item and context memory accuracy were analyzed using the BANOVA.Bernoulli() function, which used a logit link function to model correct response proportions, with encoding condition as a within-subjects factor and retrieval instructions as a between-subjects factor. Modeling dichotomous responses in this way also provides a more valid estimate of response probabilities near one and zero, compared to more traditional methods that can be particularly distorted by floor and ceiling effects. Factors were effects-coded in the models with the Read condition and typed item retrieval as reference levels. Default prior distributions and model parameters were used. Except where otherwise noted, Markov Chain Monte Carlo simulations (carried out in JAGS; Plummer, 2003) consisted of 25,000 total draws, with 5,000 burn-in and thinning factor of 10, for a total of 2,000 samples. In all analyses, convergence diagnostics indicated that the MCMC chains converged. Sampled posterior probability distributions of model coefficients were used to generate predicted values and 95% credible intervals for correct response proportions in each condition, and Bayesian p-values for each factor and interaction (see Dong & Wedel, 2016, for full specification of BANOVA outputs). The Bayesian p-values reported below are meant to be analogous to classical p-values, although they reflect a somewhat different concept in that they are computed based on where the tail of a model parameter’s posterior distribution crosses zero (then doubled to create a “two-tailed” value). Although presented here as an inferential tool, overinterpretation of these should be avoided, as with classical p-values (Marsman & Wagenmakers, 2017; Wagenmakers, 2007). In addition to the BANOVA analyses reported for each experiment, further inferential support was provided by applying a Bayesian linear mixed-effects modeling approach to the combined context memory data from all three experiments (using the brms package in R; Bürkner, 2017), which enabled model comparison via Bayes Factor (details of that analysis are provided in the Results section for Experiment 3).

MPT model analyses reported here also employed a hierarchical Bayesian approach for the estimation of model parameters, using the TreeBUGS package in R (Heck, Arnold, & Arnold, 2018). Models were constructed by adapting prior MPT models of source memory (e.g., Bayen, Murnane, & Erdfelder, 1996) to the present experiments, and are described in further detail below. For each model, latent-trait MPT parameters were estimated by MCMC simulations in JAGS consisting of four chains with 20,000 adaptation steps followed by 220,000 total draws, with 20,000 burn-in and thinning factor of 20, for a total of 10,000 samples. Default prior distributions were used, which for group mean parameters were uniform across probability space. In all analyses, convergence diagnostics indicated that the MCMC chains converged. Posterior-predictive p-values p_T1 and p_T2 were used to measure adequacy of model fit to means and covariances in the data (e.g., Klauer, 2010), and relative model fits were compared using the widely-applicable or Watanabe-Akaike information criterion (WAIC; Watanabe, 2010) which quantifies predictive accuracy while correcting for the number of parameters in a model, and is preferred over other statistics within the Bayesian framework (e.g., AIC, DIC) for its use of full posterior distributions in its estimation (Gelman, Hwang, & Vehtari, 2013). Posterior distributions of parameter estimates were also used to identify 95% credible intervals for pairwise differences between parameters.

Trial exclusions

Participants’ responses in the encoding task were inspected to ensure basic compliance with the task instructions. In general, participants followed instructions, re-typing provided exemplars in the Read condition and producing exemplars of the cued categories in the generation conditions. Misspellings or typographical errors occurred on approximately 2.5% of encoding trials across both versions of the experiment; in these cases it was still possible to identify what the intended response had been, and score it accordingly. A total of nine encoding trials, across seven participants, had blank responses; for six of these trials (all in Experiment 1b) the blank encoding response caused there not to be anything to recall in the corresponding cued recall trial. Those nine cued recall trials were therefore considered invalid and were excluded from analysis. In the analysis of cued recall trials, items were considered as matching a prior response from encoding (i.e., either the participant’s own encoding response or the correct response that was presented in the feedback display) if the recall response could be interpreted as representing the same lexical item as the item from the encoding task. Thus, some variations between encoded items and recalled items were allowed, including spelling and typographical errors as well as minor word variations (e.g., singular and plural forms, abbreviations, etc.). Across both versions of the experiment, approximately 3.2% of valid recall trials involved this type of approximate match between the recall response and an encoded item.

Item accuracy

The left panel of Figure 2 displays proportion correct cued recall responses for individual participants (circles), as well as mean and 95% credible intervals of predicted proportion correct based on posterior probability distributions of the coefficients in the hierarchical BANOVA model. Gray symbols represent results from Experiment 1a, in which cued recall targets were the feedback items presented to participants during encoding, and white symbols represent results from Experiment 1b, in which targets were the participants’ own typed responses during encoding (which differed from the feedback items in the Incorrect Generation condition). Circle size represents the number of valid trials from which each participant’s proportion was computed (most participants had 14 valid trials per condition; note that, thanks to the Bayesian modeling approach, participants with fewer valid trials inherently had less influence on the estimation of model parameters). A summary of the BANOVA model itself is provided in Table 1, with estimates of the linear model coefficients and 95% credible intervals. Bayesian p-values generated from the model suggested that there was a significant main effect of encoding condition, p < .001, and a significant interaction of encoding condition and retrieval instructions, p = .002. The main effect of retrieval instructions was not significant, p = .379. Consistent with previous studies of generation effects in memory (e.g., Slamecka & Graf, 1978), predicted accuracy across both versions of the experiment was superior in the Correct Generation condition (M = .989, 95% CI [.980, .994]) than in the Read condition (M = .824, 95% CI [.751, .880]). Predicted overall performance in the Incorrect Generation condition fell between that of the other conditions (M = .922, 95% CI [.882, .950]).

*Note.* Item memory accuracy (left panel) is represented as the proportion of valid cued recall trials for which the target item was recalled. For Experiment 1a (gray symbols), target items for recall were the category exemplars provided in the feedback display during encoding; for Experiment 1b, target items were the participant’s own typed responses during encoding. Context memory accuracy (right panel) is represented as identification-of-origin score, which is the proportion of correctly recalled targets for which the correct font color was identified. For both panels, circles represent individual participant data, with circle size indicating the number of trials used to compute the proportion. Boxes represent means and 95% credible intervals of the hierarchical BANOVA model’s predictions for response proportions, which were generated from the posterior probability distributions of the estimated linear model coefficients.

Table 1.

Experiment 1 Item Accuracy: Estimated BANOVA Coefficients

Effect	Full data set (n=69)		Limited data set (n=62)
Effect	Estimate	95% CI	Estimate	95% CI

Intercept	2.83	[2.48, 3.20] ^*	2.91	[2.58, 3.26] ^*
Retrieve Feedback Item	0.15	[−0.19, 0.52]	−0.10	[−0.44, 0.23]
Correct Generation	1.65	[1.28, 2.07] ^*	1.54	[1.15, 1.99] ^*
Correct Generation : Retrieve Feedback	−0.49	[−0.88, −0.13] ^*	−0.36	[−0.79, 0.02] ^*
Incorrect Generation	−0.36	[−0.70, −0.02] ^*	−0.09	[−0.40, 0.22]
Incorrect Generation : Retrieve Feedback	0.52	[0.19, 0.86] ^*	0.16	[−0.13, 0.47]

Open in a new tab

Note. Retrieve Typed Item and Read were the reference levels for Retrieval Instructions and Encoding Condition, respectively. Model for limited data set excluded participants with exceptionally low item accuracy in the Incorrect Generation condition.

denotes that zero lies outside the 95% credible interval.

Within the Incorrect Generation condition, an apparent difference based on retrieval instructions appeared to be driving the interaction effect in the model. Inspection of incorrect trial data indicated that most of the incorrect cued recall responses in the Incorrect Generation condition (75.7% in Experiment 1a, 94.0% in Experiment 1b) were studied nontarget items – i.e., recall of the participant’s own response rather than the feedback item in Experiment 1a, or vice versa in Experiment 1b. Thus, the Incorrect Generation condition produced some susceptibility to interference from studied nontargets. Additionally, as the data in Figure 2 show, there was a cluster of seven participants in Experiment 1b whose accuracy in the Incorrect Generation condition was less than 25%. Among these participants, 88.8% of all recalled responses were studied nontargets, suggesting these participants may have misunderstood the instruction to recall their own typed responses and intentionally recalled feedback items instead. Therefore, an additional BANOVA model was run with all responses from these seven participants excluded, resulting in a sample size of n = 31 for both versions. Bayesian p-values generated from this model still suggested a main effect of encoding condition, p < .001, but no longer suggested a significant interaction, p = .065. The main effect of retrieval instructions remained non-significant, p = .552. Predicted proportion correct for the Incorrect Generation condition overall (M = .944, 95% CI [.917, .963]) was still slightly lower than for Correct Generation overall (M = .989, 95% CI [.979, .994]), and both generation conditions had substantially better accuracy overall than the Read condition (M = .811, 95% CI [.745, .862]).

Context accuracy

The right panel of Figure 2 displays context memory accuracy in terms of I-O scores, with circles representing individual participants’ response proportions, circle size representing the number of trials underlying each proportion, and boxes representing means and 95% credible intervals of predicted response proportions from a hierarchical BANOVA model with encoding condition and retrieval instructions as within-subjects and between-subjects factors, respectively. Estimates of BANOVA model coefficients are presented in Table 2. Model predictions suggest very similar levels of accuracy in the Read and Correct Generation conditions, under both retrieval instruction conditions. In contrast, accuracy in the Incorrect Generation condition in Experiment 1a (recall feedback item; M = .687, 95% CI [.618, .750]) was substantially higher than in other conditions. In particular, a clear difference was observed in comparison to the Incorrect Generation condition in Experiment 1b (recall typed item; M = .539, 95% CI [.468, .614]). Accuracy in the Incorrect Generation condition in Experiment 1a was also somewhat higher than in the Correct Generation condition in Experiment 1a (M = .602, 95% CI [.527, .670]), although the 95% credible intervals overlapped (see pairwise comparisons below). Additionally, the Bayesian p-values generated from the model suggested there was no significant main effect of encoding condition, p = .388, or of retrieval instructions, p = .123, but did suggest a significant interaction, p = .010. As noted above, a subset of seven participants may have misunderstood retrieval instructions in Experiment 1b, which could lead to a concern that these participants might also have negatively influenced context memory accuracy in the Incorrect Generation condition. Thus, the BANOVA analysis was re-run with all data from those seven participants excluded. Consistent with the model that used the full data set, there did not appear to be a main effect of encoding condition, p = .513, or of retrieval instructions, p = .238, but there was evidence of a significant interaction, p = .007. Predicted performance from these data still suggested a difference in context memory in the Incorrect Generation condition between Experiment 1a (M = .690, 95% CI [.616, .753]) and Experiment 1b (M = .540, 95% CI [.461, .616]) as well as a potential difference between Incorrect Generation in Experiment 1a and Correct Generation in Experiment 1a (M = .602, 95% CI [.526, .671]). Pairwise analyses were conducted to further examine the model’s estimates of the differences between these conditions. The posterior samples of the linear coefficients were used to construct a distribution of the predicted accuracy difference between the Incorrect Generation and Correct Generation conditions in Experiment 1a, and a distribution of the predicted accuracy difference between the Incorrect Generation conditions in Experiments 1a and 1b. The distributions are displayed in Figure 3. For Incorrect Generation versus Correct Generation in Experiment 1a, the proportion of posterior samples for which the predicted difference was less than or equal to zero was .0185. For Incorrect Generation in Experiment 1a versus Experiment 1b, the proportion of posterior samples for which the predicted difference was less than or equal to zero was .0055. Thus, both comparisons indicated a high estimated probability of a true advantage in context memory for Incorrect Generation in Experiment 1a relative to the other conditions.

Table 2.

Experiment 1 Context Accuracy: Estimated BANOVA Coefficients

Effect	Full data set (n=69)		Limited data set (n=62)
Effect	Estimate	95% CI	Estimate	95% CI

Intercept	0.40	[0.27, 0.54] ^*	0.42	[0.27, 0.58] ^*
Retrieve Feedback Item	0.11	[−0.03, 0.24]	0.09	[−0.07, 0.23]
Correct Generation	0.01	[−0.14, 0.16]	0.00	[−0.16, 0.15]
Correct Generation : Retrieve Feedback	−0.10	[−0.26, 0.04]	−0.09	[−0.25, 0.07]
Incorrect Generation	0.07	[−0.09, 0.23]	0.06	[−0.11, 0.23]
Incorrect Generation : Retrieve Feedback	0.21	[0.04, 0.37] ^*	0.23	[0.07, 0.40] ^*

Open in a new tab

Note. Model for limited data set excluded participants with exceptionally low item accuracy in the Incorrect Generation condition.

denotes that zero lies outside the 95% credible interval.

*Note.* The horizontal axis in each panel represents the difference between two specific conditions in the BANOVA model’s predictions of proportion correct context identification. Top panel: distribution of predicted differences between the Incorrect Generation and Correct Generation conditions in Experiment 1a. Bottom panel: distribution of predicted differences between Incorrect Generation in Experiment 1a and Incorrect Generation in Experiment 1b. Both distributions indicate high probability that the differences were positive, i.e., context accuracy was greater for the Incorrect Generation condition in Experiment 1a than for either of the other conditions of interest. The limited dataset model was used for these analyses.

MPT models

As described above, data from the experiment were also modeled using a multinomial processing tree whose parameters were estimated in a hierarchical Bayesian framework using the R package TreeBUGS. The trees used for Experiment 1 are displayed in Figure 4 and include parameters for item retrieval (D), context retrieval (d), item guessing (b) and context guessing (a). Because category cues were provided at retrieval, and the cues themselves had been presented in colored font at encoding, it was assumed that context retrieval could occur in the absence of item retrieval. Thus, the model’s estimation of context memory took into account participants’ color responses for all valid item responses, including incorrect items (but excluding invalid trials as described above). Data were excluded from the participants who potentially misunderstood task instructions in Experiment 1b, as noted above. Separate models were fit to each version of the experiment (i.e., 1a and 1b), and models fit to each version included both a full model and a limited model with additional parameter restrictions. In the full models, separate processing trees were assumed for the three encoding conditions, with both item memory (D) and context memory (d) allowed to vary across conditions (guessing parameters a and b were assumed to be equal across conditions). In the limited models, the context memory parameter d was restricted to be equal across conditions.

*Note.* Processing trees represent possible combinations of cognitive operations assumed to contribute to each response type at test. Cues in the experiment referred to study episodes in which either blue or yellow font had been used (left side). Participants’ responses (right side) could consist of target items or nontarget items followed by identification of the associated font color as either blue or yellow. D = probability of retrieving target item; d = probability of retrieving font color; b = probability of guessing target item; a = probability of guessing “blue.” Separate trees were used for the three encoding conditions. For full models, parameters were restricted such that D₁ = D₂ and d₁ = d₂ within each condition, and all a’s were equal and all b’s were equal across conditions. For limited models all d’s were restricted to be equal.

Model parameters for the full models are presented in Table 3. It can be seen that the MPT model estimates closely aligned with the results of the linear-model based analyses above. In particular, item retrieval was found to be substantially greater in the Correct Generation and Incorrect Generation conditions than in the Read condition, in both versions of the experiment. For the context memory parameters, d_Incorrect was found to be reliably greater Experiment 1a than in Experiment 1b (95% CI of the difference: [.021, .436]), consistent with the pairwise difference seen in I-O scores in the Incorrect Generation condition across experiment versions. Similarly, within Experiment 1a, d_Incorrect was estimated to be greater than d_Correct although this difference was not as statistically reliable as in the analysis of I-O scores (95% CI of the difference: [−.048, .373]). For comparison between the full and limited models, fit statistics are presented in Table 4. For Experiment 1a, the full model was preferred both in terms of absolute model fit (i.e., the limited model failed to describe the covariance in the data, p_T2 < .05) and relative fit (i.e., the mean difference in WAIC was greater than two standard errors). Thus, there was strong evidence for including context memory differences across conditions in the MPT model. For Experiment 1b, predictive accuracy was virtually identical for the full model compared to the limited model, consistent with the interpretation that context memory did not differ across conditions when participants were instructed to recall the context associated with their own typed responses. One further MPT model analysis was conducted to address a potential alternative explanation for the apparent advantage in memory for context associated with corrective feedback. Specifically, because the cued recall task closely resembled the generation trials in the initial encoding task, a participant could hypothetically provide some correct item responses at test by re-generating category members on the fly, without retrieving any item information from the study episode. This could be expected to occur less often in the Incorrect Generation condition, in which the correct exemplar for each category was guaranteed not to be the one most preferred by the participant. Consequently, I-O scores for the Incorrect Generation condition were more certain to have included only trials on which item retrieval actually occurred, whereas I-O scores in other conditions could have included some trials in which there was no retrieval of item information from the encoding phase. Although the aforementioned MPT models included all retrieval trials (not just correct ones), the item-guessing parameter (b) was assumed to be equal across conditions. To evaluate the hypothesis that different rates of item guessing accounted for the observed data, an MPT model was fit in which the b parameter was allowed to vary across conditions, while the context memory parameter (d) was not. The variable-guessing model provided an adequate fit to the data in terms of means (p_T1 = .356) but not in terms of covariance (p_T2 = .023). Predictive accuracy (WAIC = 1177.9) was also worse than for the model with context memory differences, although the difference in WAIC between models (M = −24.5, SE = 14.7) was not as large as the difference between the model with variable-context memory differences and the model that had both equal context memory and equal item guessing across conditions.

Table 3.

Estimated MPT Model Parameter Estimates (Full Models)

	Cued recall				Free recall				Recognition
	Feedback item (Exp. 1a)		Typed item (Exp.1b)		Feedback item (Exp. 2a)		Typed item (Exp.2b)		Feedback item (Exp. 3a)		Typed item (Exp.3b)
Parameter	M	95% CI	M	95% CI	M	95% CI	M	95% CI	M	95% CI	M	95% CI

D _Read	0.18	[0.00, 0.60]	0.24	[0.01, 0.67]	0.17	[0.12, 0.21]	0.15	[0.11, 0.19]	0.67	[0.49, 0.79]	0.09	[0.00, 0.30]
D _Correct	0.91	[0.77, 0.98]	0.96	[0.90, 1.00]	0.36	[0.31, 0.42]	0.41	[0.36, 0.46]	0.92	[0.87, 0.96]	0.92	[0.86, 0.97]
D _Incorrect	0.76	[0.45, 0.94]	0.76	[0.46, 0.93]	0.34	[0.27, 0.40]	0.31	[0.22, 0.40]	0.95	[0.87, 0.99]	0.67	[0.38, 0.85]
d _Read	0.11	[0.01, 0.23]	0.15	[0.03, 0.26]	0.17	[0.01, 0.49]	0.26	[0.02, 0.53]	0.13	[0.01, 0.39]	0.21	[0.03, 0.51]
d _Correct	0.13	[0.02, 0.26]	0.17	[0.04, 0.28]	0.06^b	[0.00, 0.17]	0.15	[0.01, 0.31]	0.11	[0.00, 0.37]	0.10	[0.00, 0.43]
d _Incorrect	0.30^a	[0.11, 0.48]	0.06^a	[0.00, 0.17]	0.31^b,c	[0.10, 0.49]	0.07^c	[0.00, 0.21]	0.12	[0.00, 0.41]	0.14	[0.01, 0.46]
a	0.50	[0.46, 0.54]	0.54	[0.51, 0.58]	0.55	[0.49, 0.61]	0.47	[0.42, 0.53]	0.51	[0.33, 0.58]	0.50	[0.23, 0.58]
b	0.70	[0.41, 0.84]	0.64					[0.20, 0.81]	0.41	[0.21, 0.61]	0.78	[0.68, 0.86]
D _{2_Read}									0.75	[0.56, 0.88]	0.91	[0.85, 0.96]
d _{3_Incorrect}									0.91	[0.73, 0.99]	0.06	[0.00, 0.36]

Open in a new tab

Note. Values of 0.00 and 1.00 are the result of rounding and do not reflect exact estimates. Superscript letters a, b, c indicate pairwise comparisons of interest for context memory parameters (d), for which the 95% credible interval of the mean difference did not include zero.

Table 4.

Summary of MPT Model Fits

Task	Retrieval instruction	Full (variable context) model			Equal context model			WAIC difference
Task	Retrieval instruction	p _T1	p _T2	WAIC	p _T1	p _T2	WAIC	M	SE

Cued Recall	Feedback item	0.87	0.09	1153.4	0.34	0.02	1186.6	−33.2	14.1
	Typed item	0.64	0.50	1136.6	0.62	0.34	1137.2	−0.6	5.7
Free Recall	Feedback item	0.65	0.37	1047.1	0.33	0.17	1054.7	−7.6	5.9
	Typed item	0.21	0.48	1187.8	0.22	0.23	1185.1	2.6	4.4
Recognition	Feedback item	0.10	0.32	1066.4	0.08	0.24	1066.8	−0.4	7.3
	Typed item	0.13	0.47	1181.9	0.07	0.48	1180.4	1.5	3.7

Open in a new tab

Note. Posterior-predictive p−values for T1 and T2 statistics reflect adequacy of model fits to the means and covariances of the data, respectively (Klauer, 2010), with values <0.05 indicative of inadequate fit. WAIC (Watanabe, 2010) reflects the model's predictive accuracy while correcting for the number of parameters in the model, with lower values preferred.

Discussion

The findings from Experiment 1 provide novel evidence that corrective feedback affects memory for contextual details. Specifically, incorrect responses with corrective feedback resulted in greater accuracy in memory for contextual details associated with the feedback display than for contextual details associated with the initial response. This difference was not observed for confirmatory feedback when participants generated correct responses or retyped items that were presented. As noted earlier, the experiment was structured so that cues were randomly assigned to Correct Generation, Incorrect Generation, and Read conditions, and participants could not reliably predict whether a response they generated would be correct. Thus, the observed effect of corrective feedback was independent of potentially covarying factors such as participants’ confidence, degree of surprise, or prior knowledge of the subject matter. Additionally, the finding that there was no advantage in context memory associated with typed items in Experiment 1b is important because it demonstrates that the effect of corrective feedback on context memory did not extend to non-feedback information in the encoding episode. It also rules out the possibility that the advantage in context memory for corrective feedback was caused by some other difference between encoding conditions, such as a lack of competing contextual information for targets. Taken together, the results are consistent with the hypothesis that there is enhanced encoding of corrective feedback (e.g., Potts et al., 2019), and that this enhanced encoding applies to both item and context information.

In addition to the novel findings regarding context memory, the results of Experiment 1 also provide valuable data regarding item memory and errorful learning. By comparing the effects of study alone, correct generation with confirmatory feedback, and incorrect generation with corrective feedback, the present results add to a growing literature that seeks to delineate the conditions under which error generation can be beneficial or harmful to learning (e.g, Bridger & Mecklinger, 2014; Zawadzka & Hanczakowski, 2019).

Experiment 2

Experiment 1 demonstrated a performance advantage in context memory for corrective feedback, consistent with the overarching hypothesis that episodic processes, and not just semantic mediation, are involved in errorful learning. As mentioned above, the observed effect could be caused by enhanced encoding of corrective feedback relative to other information during a learning trial.. Analyses of multinomial processing tree models also did not support an alternative explanation that the specific context memory advantage for corrective feedback might have been an artifact of differences in item guessing across conditions (although it is important to note that such an explanation still implies that episodic processes are involved in retrieval of corrective feedback). Nonetheless, it is useful to test context memory for corrective feedback in a task that reduces the possibility of item guessing.

Experiment 2 provided such a test by employing the same encoding task as Experiment 1, but with free recall as the retrieval task instead of cued recall. The use of free recall helped to ensure that at least some retrieval was involved in each correct response at test, as it would be unlikely for participants to spontaneously generate items from the encoding phase in a memory-free manner (re-generation of exemplars could still occur but would require, at a minimum, retrieval of the category cue from the learning trial).