Eyewitness Accuracy And Retrieval Effort: Effects Of Time And Repetition Part 3

Following the procedures of Lindholm et al. [4], we then created a single model containing all the significant predictors and examined their relative contribution to accuracy. 

According to research, there is indeed a relationship between predictor variables and memory. This means that predictor variables can affect people's memory ability and can also improve their memory level.

Predictor variables are those factors related to sensory experience, including sounds, smells, colors, etc. These factors have an important impact on people's memory abilities. For example, when we hear or see sounds or images associated with a memory, we can recall the memory more easily.

In addition, research also shows that emotional factors can also have an important impact on memory ability. When people experience emotional, passionate events, their memories are deeper and longer lasting. Therefore, we can improve our memory by creating positive, enjoyable experiences.

In addition to these factors, exercising, getting enough sleep, and eating a healthy diet can also improve memory. To sum up, there is indeed a close relationship between predictor variables and memory. We can improve our memory ability in a variety of ways and create a better life experience. It can be seen that we need to improve memory, and Cistanche deserticola can significantly improve memory because Cistanche deserticola is a traditional Chinese medicinal material that has many unique effects, one of which is to improve memory. The efficacy of minced meat comes from the various active ingredients it contains, including acid, polysaccharides, flavonoids, etc. These ingredients can promote brain health in a variety of ways.

Click know 10 ways to improve memory

Thus, we created a model with Delays, Hedges, Non-word fillers, Word fillers, and Confidence. No explicit prediction was made for this analysis. All predictors except Word fillers proved significant, unique predictors of accuracy in the resulting model (see Table 2). In Table 2, the odds ratio indicates the increase/decrease in accuracy when increasing one step on the scale of each variable, with values above zero indicating an increase, and values below zero indicating a decrease. 

That is, a statement with no hedges will be 29% more likely to be correct compared to a statement with one hedge (UOR = 0.71), and a confidence judgment with 81% confidence is 3% more likely to be correct compared to a confidence judgment of 80% (UOR = 1.03; see Table 2).

We next examined the effects of time and repetition on the relationship between accuracy and retrieval-effort cues, as well as confidence (Hypotheses 2–5). To make these analyses less convoluted, we first created an "effort index" out of the effort cues that had a significant unique contribution to accuracy (see preregistration). 

This included Delays, Hedges, and Non-word fillers. However, as the effect of non-word fillers on accuracy was opposite to the expected direction (see Fig 2), and contrary to previous findings [4, 7, 8], we decided to drop non-word fillers and make the effort index out of the two remaining cues: Delays and Hedges.

Effects of time and repetition on retrieval-effort cues

To test the effects of time and repetition on retrieval-effort cues, we compared a baseline, intercept-only model of the effort index with models containing Time, Repetition, Accuracy, and their interactions as predictors. For these analyses we expected a main effect of accuracy (a greater effort index value for incorrect memories, Hypothesis 1), and then several interactions (Hypotheses 2–5). 

In short, these interactions could be summarized as an easier retrieval over time for the repetition group and a more difficult retrieval over time for the no-repetition group (Hypothesis 2–3), and a smaller difference in retrieval-effort cues between correct and incorrect memories over time for the repetition group (Hypothesis 4–5). As in the previous multilevel analyses, we would expect these differences to manifest in the data as significant predictor models with higher Akaike weights compared to baseline models. 

Turning to the results, analyses showed that that model fit was indeed significantly improved compared to the baseline model when adding Accuracy, such that incorrect statements (M = 0.77, SD = 0.97) were produced with more effort compared to correct statements (M = 0.42, SD = 0.75; d = 0.44; see Table 3 and Fig 4). Unexpectedly, model fit was also significantly improved when adding Time, such that statements were produced with more effort at the first interview (T1; M = 0.57, SD = 0.88) compared to the second interview two weeks later (T2; M = 0.45, SD = 0.77, d = 0.14; see Table 3 and Fig 4).
Moreover, model fit was significantly improved with Repetition, such that the repetition group reported memories with less effort (M = 0.42, SD = 0.73) compared to the no-repetition group (M = 0.59, SD = 0.90, d = 0.23, see Table 3 and Fig 4). To test interaction effects, we created models containing each interaction and compared them to models of only their respective predictors (e.g. a model with time and accuracy as predictors was compared to a model with time, accuracy, and the time-accuracy interaction as predictors). 

Here we expected significant interactions of all four combinations of Accuracy, Time, and Repetition. Contrary to expectations, however, none of the interactions significantly improved fit (see Table 3).

Effects of time and repetition on confidence

We then carried out identical analyses for a confidence model. Expectations were identical to those for retrieval effort but with reversed directions, such that we expected lower confidence in conditions where we had expected higher retrieval effort. In line with predictions, results showed that model fit was significantly improved compared to the baseline model when adding Accuracy, such that confidence was higher for correct statements (M = 86.25, SD = 21.00) compared to incorrect statements (M = 71.65, SD = 26.18, d = 0.70, see Table 3 and Fig 3). 

Moreover also as expected, model fit was not significantly improved when adding Time, (MT1 = 82.45, SD = 23.54; MT2 = 82.54, SD = 23.13 d < 0.01, see Table 3 and Fig 3); nor Repetition (Mrepetition = 83.53, SD = 21.73; Mno-repetition = 81.47, SD = 24.78, d = 0.09; see Table 3 and Fig 3). Interestingly, all interactions improved fit as expected: Time-Repetition, Time-Accuracy, Repetition-Accuracy, and Time-Repetition-Accuracy (see Table 3 and Fig 3). Planned comparisons (see preregistration) showed that confidence indeed significantly increased from T1 to T2 for the repetition group (Mdiff = 1.62, p = .042, d = 0.07) whereas there was a nonsignificant decrease for the no-repetition group (Mdiff = -1.56, p = .087, d = 0.06). Moreover as expected, the increased confidence for the repetition group was mainly driven by higher confidence in incorrect statements (Mdiff = 6.51, p < .001, d = 0.27) as there was no significant increase in correct statements (Mdiff = 0.53, p = .536, d = 0.03). 

For the no-repetition group, the decrease in confidence between T1 and T2 was not statistically significant for either incorrect statements (Mdiff = -0.61, p = .764, d = 0.02), or correct statements (Mdiff = -1.58, p = .091, d = 0.07). Hence, in contrast to the results for the effort index, the results for confidence were more in line with expectations.

Retrieval-effort index mediates between confidence and accuracy

Finally, to examine whether retrieval effort was used as a basis for confidence (Hypothesis 6), we carried out a mediation analysis between accuracy and confidence, with the effort index as a mediator. Results showed that the effort index mediated 22.7% of the relation between accuracy and confidence (see Fig 5).

Discussion

The main aim of this experiment was to investigate the effects of time and repetition on the relation between retrieval effort and accuracy. Secondary aims involved investigations of confidence, as well as the relationship between retrieval effort, confidence, and accuracy. Additionally, we examined the effects of time and repetition on memory accuracy and memory quantity. There are four major takeaways from this study, namely that 1) retrieval-effort cues predict accuracy over time, 2) retrieval-effort cues decrease over time, 3) confidence increases over time mainly for incorrect memories when memories are repeated, and 4) a retrieval-effort index mediates the relation between confidence and accuracy. However, there were also methodological constraints, leading to smaller-than-expected effects of time and repetition, which also potentially limited generalizability. We will now detail discussions of each of these findings, before moving on to general discussions and limitations.

Retrieval-effort cues predict accuracy over time

Overall, our results add to a large body of research showing that correct memories are more easily retrieved compared to incorrect memories (e.g. [4–6, 8, 13–16]; see Figs 2 and 4). Whereas other studies have examined this relationship after retention intervals up to a few minutes, we demonstrate that this relationship persists at a second recall a couple of weeks later. This was evident also when participants had engaged in repeated retrieval during the retention interval (see Fig 4). The results thus suggest that retrieval-effort cues can be reliable predictors of memory accuracy over extended retention intervals (however, see also section "Retrieval ease decreases over time" below).

We measured six cues to retrieval effort, of which Delays, Hedges and Word fillers were significantly more numerous in incorrect responses compared to correct responses. Delays and Hedges were the strongest predictors (see Table 1), in line with previous findings [4, 8]. Somewhat surprisingly, Non-word fillers showed the opposite result, as there were more numerous incorrect responses (see Fig 2). 

Previous findings on non-word fillers have been slightly inconclusive, as Lindholm et al. [4] and Smith and Clark [7] found non-word fillers to be significantly more common in incorrect responses, whereas there was no statistically significant effect in Gustafsson et al. [8]. Although we have reasoned that non-word fillers are expressed automatically as a consequence of effortful memory retrieval, Clark and Tree [50] compellingly argue that fillers are used intentionally like conventional words in a language and that their usage largely signals turn-taking. 

Thus, non-word fillers might not always signal that one is effortfully attempting to retrieve a memory, but may also signal that one is deciding how to formulate a coming sentence, or that one wants to end a speaking turn. This explains the inconsistencies in the findings regarding accuracy, although it is still somewhat puzzling that we found significantly more non-word fillers and incorrect statements. Nonetheless, an important conclusion from these contrasting results is that non-word fillers are not a reliable predictor of accuracy. 

Another surprising finding was that the two "new" measures of retrieval effort–Prolongations and False starts–did not significantly predict accuracy (see Fig 2). These two cues were inspired by psycholinguistic research on disfluencies, that is, utterances that disrupt the flow of speech (e.g. [51]). We reasoned that a prolonged pronunciation of a word would be a consequence of an inability to retrieve a memory and therefore be more common in incorrect responses. Similarly, we believed that false starts would mainly occur when memory was not fully retrieved, and hence signal inaccuracy. We found no support for these ideas, however. Instead, the overarching evidence points to Hedges and Delays as the most reliable effort cues to indicate memory accuracy.

Given the consistent effort-accuracy relation, a reasonable question to ask is how to use this knowledge in the field as a practitioner. Because there appears to be some variation between individuals in expressing effort (see e.g. "T1" in Fig 4) we suggest that this knowledge is as of now best used carefully, ideally in conjunction with other corroborating evidence, such as physical evidence or other eyewitness reports. A starting point could be to judge statements without hedges or delays as correct, which is supported by the unstandardized odds ratios in Table 2, which show that each hedge should decrease the likelihood of accurate recall by 29%, and each delay should decrease accuracy by about 16%. Such a method has shown some success in improving the judgment accuracy of eyewitness testimonies (see [9]) and is similar to the recommendation given by Wixted and Wells [61] for identification research, where they suggest that highly confident witnesses should generally be believed (given "pristine" conditions).

Retrieval-effort cues decrease over time

Another major finding is that retrieval-effort cues decreased over time regardless of repetition and accuracy. That is, participants who had repeatedly retrieved memories over the two-week interval, as well as those who hadn't, used fewer effort cues in the second interview when recalling correct and incorrect statements alike (d = 0.14). We only expected increased retrieval ease for the repetition group, as repetition is known to facilitate retrieval (e.g. [11, 37]), and expected the no-repetition group to have greater difficulties in retrieving memories, due to memory weakening and forgetting. The increased retrieval ease for the no-repetition group at T2 is not likely due to spontaneous repetition among these participants, as they scored low on the two questions about time spent reflecting on the event, nor due to selective reporting of easily retrieved memories (see [62]), as there was no significant reduction in the number of total unique details reported between T1 and T2 (see Fig 1). Instead, we see three plausible explanations for this effect: a) repeated retrieval opportunities as T1, b) context-dependent learning effects, and c) a switch in the grain size of reported details.

The first explanation-repeated retrieval opportunities at T1-is probably the most important. That is, the participants were allowed to retrieve their memories of the event directly after seeing it, as they were interviewed about the event. They did this several times, first in the free recall session and again in the cued recall session. In addition, there was a third retrieval opportunity, namely during the confidence ratings. During this task, the experimenter read aloud details that the witness had reported and thus allowed the participant to elaborate upon each mentioned detail. This abundance of retrieval attempts likely helped participants consolidate their memory of the event, leading to only minor forgetting at T2 for both groups. Although this could likely have been avoided with a between-group design in which one group was only tested at T2, a within-group was the optimal choice to follow the development of effort cues over time, given the variance in the use of effort cues between individuals (see [9], but see also T1 in Fig 4 above).

The second explanation-context-dependent learning [63]-suggests a more successful retrieval when the retrieval takes place at the same location as the encoding of the event. Our participants were interviewed in the same experimental room during both sessions, so it is possible that this context facilitated their retrieval, minimizing potential forgetting effects during the retention interval.

The third explanation is a shift in the grain size of the details. Koriat and Goldsmith [62] suggest that people can not only decide to withhold or report a memory but also shift the level of detail with which the memory is reported. Thus, a piece of clothing could be described with a high level of detail ("a blue zip-up jacket with green stripes and a hood") or a low level of detail ("a jacket"). In this study, we did not code for grain size, so it is possible that the repetition group remembered things in more detail than did the no-repetition group. We encourage researchers to examine this in future studies, and as our data is openly available, suggest examinations of our data.

We suggest that these explanations also likely led to a smaller overall effect of the repetition manipulation, due to floor effects.

In addition to expecting an easier retrieval overall at T2 for the repetition group, we also expected an increased retrieval ease for incorrect statements compared to correct statements, again due to expected floor effects for correct memories. The findings indeed indicate a greater retrieval ease for incorrect memories (see Fig 4), but the time-accuracy interaction was not statistically significant. If this effect is not due to changes in the grain size of reporting, one might dismiss it as a too-weak manipulation of repetition. However, we believe that this showcases something more important, namely that all memories are not as fragile and easily manipulated as the memory research field might sometimes give the impression (see also [33, 34]). Finally, although it is worth pointing out again that the effort cues still predicted accuracy at T2 despite the general increase in retrieval ease (see Fig 4), a continued decrease over time will likely lead to a point wherein accuracy can no longer be distinguished by retrieval-effort cues. Thus, examining the effects of longer retention intervals with repetition is an avenue for future research.

Higher confidence in incorrect memories after repetition

Regarding confidence, we found no general increase or decrease over time. Instead, we found a three-way interaction between accuracy, time, and repetition, with the notable finding that confidence only significantly increased for incorrect memories in the repetition group (d = 0.27; see Fig 3). The effect was not evident for correct memories (d = 0.03). To the best of our knowledge, this is the first study to demonstrate this effect. Previous studies examining the effect of confidence over time have generally presented an overall confidence score, rather than separate values for correct and incorrect responses (e.g. [23, 24, 30–34]). This is an interesting result because it would suggest that we should effectively trust people's confidence for correct memories, as these judgments were relatively stable over time and across repetition in our study. From a practical standpoint though, there is of course the big caveat that one generally has access to the confidence but not the accuracy, and that the use of confidence is to derive accuracy, not the other way around. Nonetheless, these results could have potential implications for the legal system. For example, Wixted and Wells [61] (see also [64]) have demonstrated that initially confident witnesses should generally be trusted (given "pristine" line-up conditions), as they are often correct, but not initially unconfident witnesses, as they are more likely to be incorrect. Our results add to this research by suggesting that initially confident witnesses could potentially be trusted over time, such as in later interviews with police and jurors, as they should generally remember correctly, and that they are likely to retain similar confidence levels over time. Initially unconfident witness on the other hand would instead be more likely to increase their confidence over time and become overconfident. Some reservations to these suggestions are necessary, however, as we did not observe the typical forgetting effect over time, which could indicate that these results are not representative of general situations (see also "Limitations").

The increase in confidence for incorrect but not correct memories after repetition is presumably due to ceiling effects. That is, confidence in correct memories already approached the max rating of 100% during the first session (Mcorrect = 86.36, of which 60.71% of those ratings were "100"), and thus had less room for increases than confidence for incorrect memories (Mincorrect = 71.05, of which 25.59% of those ratings were "100", see S1 Table).

For the no-repetition group, confidence decreased slightly rather than increased over time, but this decrease was not statistically significant for either correct (d = 0.07) or incorrect memories (d = 0.02; see Fig 3). We had expected confidence to decrease over time as a consequence of more difficult memory retrieval due to memory weakening and forgetting (cf. [65]), but as no major forgetting took place (see Fig 3), it is unsurprising that confidence remained relatively stable.

A final takeaway from the confidence results is that confidence still predicted accuracy at the second interview two weeks after witnessing the original event, despite the increased confidence in incorrect memories for the repetition group. Thus, similar to the results of the retrieval-effort cues, the change induced by a two-week retention interval (and memory repetition) did not largely disrupt the possibility of predicting accuracy from these two variables. This suggests that confidence can remain a reliable predictor over time. It is however important to note that this study was limited to a retention interval of only two weeks, and it is plausible that greater retention intervals and repetitions would eventually lead to an elimination of the confidence-accuracy (as well as the retrieval effort-accuracy) relationship.

Retrieval-effort index mediates between confidence and accuracy

The final major finding was that retrieval effort mediated between accuracy and confidence. Specifically, an index out of the two effort cues Hedges and Delays, mediated 22.7% of the relation between accuracy in reported memories and confidence in those memories. We draw two distinct conclusions from these results. First, the results support the cue-utilization view [25, 26], that is, that people make metacognitive judgments such as confidence based on cues–in this case, cues to retrieval effort, as participants were more confident in memories that were easily retrieved (which in turn were more likely to be correct). Second, the results suggest further bases for confidence in addition to the retrieval cues Hedges and Delays, given the relatively low percentage mediated (cf. [4, 8]). We have previously argued [8] that leftover variance could potentially be explained by confidence being "information-based", that is, based on knowledge and beliefs (for example, relying on the knowledge that it is hard to see colors accurately at night when assessing confidence in the memory of a perpetrator's clothing), in addition to the more automatic experience-based judgments from retrieval effort. However, we deem it unlikely that the majority of confidence judgments would be based on knowledge and beliefs, as these judgments are supposedly deliberate (see [26]), which contrasts with the commonly accepted view within metacognitive research that our ability to assess the bases for our metacognitive judgments is highly limited [27]. Instead, a potential explanation is that participants based their confidence on other automatic cues, and perhaps that our measures of retrieval effort did not fully capture the phenomenological experience of finding a memory hard to retrieve (supplemental analyses revealed that the other effort cues in our study were not major mediators either, see S2 Table). A final explanation is that the coding sheet contained unclarities or inconsistencies regarding the effort cues, which could explain the discrepancies between this study and previous studies regarding the effect size of the mediated relationship between confidence and accuracy (cf. [4, 8]). However, the high intercoder agreement seems to suggest the opposite. Nonetheless, to conclude, we did find that an index for retrieval effort partially mediated the relation between confidence and accuracy, corroborating previous research [4, 8]. It remains an endeavor for future research to examine further bases for confidence.

Memory accuracy and amount of unique details

We also examined the effects of time and repetition on memory accuracy and the amount of unique details provided by the witnesses. The biggest finding was an increase in both correct and incorrect details for the repetition group at T2 during free recall (see Fig 1). This suggests that the repetition manipulation did facilitate memory recall at T2, supporting the idea that the repetition manipulation was successful (as also evident from a greater rating by the repetition group of the two control questions, at d = 1.56 and d = 0.82 respectively). In line with expectations, the repetition group also provided a greater amount of unique details (Mdiff =4.82, d = 0.52) although results were not statistically significant (p = .052). Moreover, supplemental analyses show that the repetition group provided a greater amount of total statements (Mdiff = 14.17, d = 0.68, see S1 File). This corroborates established findings that repetition increases retrieval (e.g. [11, 37]), and indicates that our manipulation of repetition (i.e. asking the participants to write down all they remembered about the witnessed event every second/ third day during the two-week retention interval) was successful, albeit perhaps a bit on the weaker side.

Limitations

A somewhat puzzling result in this study was the greater amount of effort cues for the repetition group at T1 compared to the no-repetition group. This is surprising because the experimental manipulation took place after the T1 session, and participants were randomized to each group. Thus, we would have expected similar levels of retrieval-effort cues for both groups at T1. This observed difference potentially indicates that there is quite a bit of variation between participants in terms of how many effort cues they use (as we have argued earlier–see Discussion in [9]). Therefore, it would be fruitful for future studies to look into individual differences in memory retrieval efforts.

Regarding ecological validity, this experiment was conducted in the comfort (?) of a laboratory setting, with participants informed about the purpose of the study, and they could fully focus their attention on the staged crime video that they saw. This certainly contrasts with many real-life eyewitness experiences, where one might not be prepared for the witnessed event, may not experience great viewing conditions, and may have a looming threat to one's safety. Moreover, our participants were not explicitly exposed to influences from co-witnesses, post-event information, or direct forms of misinformation. Our experiment does therefore not directly generalize to natural eyewitnesses. Nonetheless, we have investigated core processes of memory that do occur outside labs, that is, retention of episodic memories with and without repetition. Moreover, although our manipulation of repetition is not fully representative of spontaneous repeated reflections, we believe that it still effectively approximates the same outcome, namely increased memory strength. We therefore believe and hope that these results will be informative both for cognitive scientists and legal practitioners alike, even with these ecological limitations.

Conclusion

In this experiment, we set out to examine the relationship between retrieval effort and accuracy (and confidence), with a special focus on the effects of time and repetition. Our results indicate that the retrieval-effort cues Hedges and Delays predict accuracy both directly after witnessing an event and two weeks later, as they were continuously more common in incorrect responses. Confidence also predicted accuracy (higher for correct responses), even though repetition led to increased confidence for incorrect responses over time. Moreover, we found support for the idea that confidence is based on retrieval-effort cues, but results suggest additional factors beyond hedges and delays.

References

1. DePaulo BM, Lindsay JJ, Malone BE, Muhlenbruck L, Charlton K, Cooper H. Cues to deception. Psychol Bull. 2003; 129(1);74–118. https://doi.org/10.1037/0033-2909.129.1.74 PMID: 12555795 

2. Sporer SL, Schwandt B. Paraverbal indicators of deception: A meta-analytic synthesis. Appl Cogn Psychol. 2006; 20(4):421–446. https://doi.org/10.1002/acp.1190 

3. Vrij A, Fisher RP, Blank H. A cognitive approach to lie detection: A meta-analysis. Legal Criminol Psychol. 2017; 22(1):1–21. https://doi.org/10.1111/lcrp.12088 4. Lindholm T, Jo¨nsson FU, Liuzza MT. Retrieval effort cues predict eyewitness accuracy. J Exp Psychol Appl. 2018; 24(4):534–542. https://doi.org/10.1037/xap0000175 PMID: 30024208 

5. Paulo RM, Albuquerque PB, Saraiva M, Bull R. The enhanced cognitive interview: Testing appropriateness perception, memory capacity and error estimate relation with report quality. Appl Cogn Psychol. 2015; 29(4):536–543. https://doi.org/10.1002/acp.3132 

6. Paulo RM, Albuquerque PB, Bull R. Witnesses' verbal evaluation of certainty and uncertainty during investigative interviews: Relationship with report accuracy. J Police Crim Psychol. 2019; 34(4):341– 350. https://doi.org/10.1007/s11896-019-09333-6 

7. Smith VL, Clark HH. In the course of answering questions. J Mem Lang. 1993; 32(1):25–38. https:// doi.org/10.1006/jmla.1993.1002 

8. Gustafsson PU, Lindholm T, Jo¨nsson FU. Predicting Accuracy in Eyewitness Testimonies With Memory Retrieval Effort and Confidence. Front Psychol, 2019; 10:703. https://doi.org/10.3389/fpsyg.2019. 00703 PMID: 30984087 

9. Gustafsson PU, Lindholm T, Jo¨nsson FU. Judging the Accuracy of Eyewitness Testimonies Using Retrieval Effort Cues. Appl Cogn Psychol. 2021; 35(5):1224–1235.

For more information:1950477648nn@gmail.com