Memory error speed predicts subsequent accuracy for recognition misses but not false alarms.

The current study aims to test whether faster recognition memory errors tend to result from stronger misleading retrieval, making them harder to correct in subsequent decisions than slower errors, and whether this pattern holds for both miss and false-alarm errors. We used a paradigm in which each single-item Old/New recognition decision was followed by a two-alternative forced-choice (2AFC) test between a target and a lure. Each 2AFC trial had one item that had just been tested for an Old/New judgment and one item that had not been previously tested. Across 183 participants, the RTs for single-item recognition errors were used to predict accuracy in the 2AFC test using a hierarchical logistic regression model. The results showed a relationship between error RT and subsequent 2AFC accuracy that was qualified by an interaction with error type. Slower miss responses were more likely to be corrected than faster misses, but no accuracy differences were observed between slower and faster false alarms. The implications of these findings are discussed as they relate to assumptions about memory processes underlying inaccurate retrieval, using the diffusion model and the two-high-threshold model as examples of accounts that explain errors in terms of misleading retrieval and failed retrieval, respectively.

A complete method for assessing the effectiveness of eyewitness identification procedures: Expected information gain.

We present a method for measuring the efficacy of eyewitness identification procedures by applying fundamental principles of information theory. The resulting measure evaluates the expected information gain (EIG) for an identification attempt, a single value that summarizes an identification procedure's overall potential for reducing uncertainty about guilt or innocence across all possible witness responses. In a series of demonstrations, we show that EIG often disagrees with existing measures (e.g., diagnosticity ratios or area under the receiver operating characteristic) about the relative effectiveness of different identification procedures. Each demonstration is designed to highlight key distinctions between existing measures and EIG. An overarching theme is that EIG provides a complete measure of evidentiary value, in the sense that it factors in all aspects of identification performance. Collectively, these demonstrations show that EIG has substantial potential to inspire new discoveries in eyewitness research and provide a new perspective on policy recommendations for the use of identifications in real investigations. (PsycInfo Database Record (c) 2023 APA, all rights reserved).

Boosting confidence without boosting performance: item strength creates the illusion of source accuracy.

People often express high confidence for misremembered sources. Starns and Ksander ([2016]. Item strength influences source confidence and alters source memory zROC slopes. Journal of Experimental Psychology: Learning, Memory, and Cognition, 42(3), 351-365; hereafter SK16) found that this happens more often when a person is highly confident in memory for the item itself, and that simply increasing item memory can increase high-confidence source errors. Under the decision heuristic account, this pattern emerges because strong item memories contaminate source judgments by promoting high confidence responses even when source evidence is relatively weak. Consequently, strengthening item memory is predicted to increase confidence for both correct and incorrect source responses; however, SK16 could not assess this key prediction because their item-strength manipulation also impaired source memory. We report two experiments with new item-strengthening manipulations designed to minimise source memory impairments. Results replicated the evidence for the decision heuristic account reported by SK16 and provided additional support by showing a boost in source confidence for both correct and error responses when item memory was strengthened without accompanying source impairments .

Bias effects in a two-stage recognition paradigm: A challenge for "pure" threshold and signal detection models.

We explored a two-stage recognition memory paradigm in which people first make single-item "studied"/"not studied" decisions and then have a chance to correct their errors in forced-choice trials. Each forced-choice trial included one studied word ("target") and one nonstudied word ("lure") that received the same previous single-item response. For example, a studied-studied trial would have a target that was correctly called "studied" and a lure that was incorrectly called "studied." The two-high-threshold (2HT) model and the unequal-variance signal detection (UVSD) model predict opposite effects of biasing the initial single-item responses on subsequent forced-choice accuracy. Results from two experiments showed that the bias effect is actually near zero and well out of the range of effects predicted by either model. Follow-up analyses suggested that the model failures were not a function of experiment artifacts like changing memory states between the two types of recognition trials. Follow-up analyses also showed that the dual process signal detection model made better predictions for the forced-choice data than 2HT and UVSD models. (PsycInfo Database Record (c) 2022 APA, all rights reserved).

Playing With BEARS: Balancing Effort, Accuracy, and Response Speed in a Semantic Feature Verification Anomia Treatment Game.

Purpose The purpose of this study was to develop and pilot a novel treatment framework called BEARS (Balancing Effort, Accuracy, and Response Speed). People with aphasia (PWA) have been shown to maladaptively balance speed and accuracy during language tasks. BEARS is designed to train PWA to balance speed-accuracy trade-offs and improve system calibration (i.e., to adaptively match system use with its current capability), which was hypothesized to improve treatment outcomes by maximizing retrieval practice and minimizing error learning. In this study, BEARS was applied in the context of a semantically oriented anomia treatment based on semantic feature verification (SFV). Method Nine PWA received 25 hr of treatment in a multiple-baseline single-case series design. BEARS + SFV combined computer-based SFV with clinician-provided BEARS metacognitive training. Naming probe accuracy, efficiency, and proportion of "pass" responses on inaccurate trials were analyzed using Bayesian generalized linear mixed-effects models. Generalization to discourse and correlations between practice efficiency and treatment outcomes were also assessed. Results Participants improved on naming probe accuracy and efficiency of treated and untreated items, although untreated item gains could not be distinguished from the effects of repeated exposure. There were no improvements on discourse performance, but participants demonstrated improved system calibration based on their performance on inaccurate treatment trials, with an increasing proportion of "pass" responses compared to paraphasia or timeout nonresponses. In addition, levels of practice efficiency during treatment were positively correlated with treatment outcomes, suggesting that improved practice efficiency promoted greater treatment generalization and improved naming efficiency. Conclusions BEARS is a promising, theoretically motivated treatment framework for addressing the interplay between effort, accuracy, and processing speed in aphasia. This study establishes the feasibility of BEARS + SFV and provides preliminary evidence for its efficacy. This study highlights the importance of considering processing efficiency in anomia treatment, in addition to performance accuracy. Supplemental Material https://doi.org/10.23641/asha.14935812.

Does speed of recognition predict two-alternative forced-choice performance? Replicating and extending Starns, Dubé, and Frelinger (2018).

Does the speed of single-item recognition errors predict performance in subsequent two-alternative forced-choice (2AFC) trials that include an item with a previous error response? Starns, Dubé, and Frelinger found effects of this kind in two experiments and accounted for them in terms of continuous memory-strength signal guiding recognition decisions. However, the effects of error speed might just as well only reflect an artefact due to an error-correction strategy that uses response latency as a heuristic cue to guide 2AFC responses, elicited through confounding factors in their experimental design such as error-correction instructions and feedback. Using two conditions, a replication condition, replicating the procedure from Starns et al., and an extension condition (each n = 130), controlling for the named shortcomings, we replicated the error speed effect. In both conditions, speed of errors in a single-item recognition task was predictive of subsequent 2AFC performance, including the respective error item. To be more precise, fast errors were associated with decreased 2AFC performance. As there was no interaction with the factor condition, the results support the idea that speed of single-item recognition responses reflects the amount of memory information underlying the respective response rather than being used for a simple error-correction strategy to improve 2AFC performance.

sdtlu: An R package for the signal detection analysis of eyewitness lineup data.

In a standard eyewitness lineup scenario, a witness observes a culprit commit a crime and is later asked to identify the culprit from a set of faces, the lineup. Signal detection theory (SDT), a powerful modeling framework for analyzing data, has recently become a common way to analyze lineup data. The goal of this paper is to introduce a new R package, sdtlu (Signal Detection Theory - LineUp), that streamlines and automates the SDT analysis of lineup data. sdtlu provides functions to process lineup data, determine the best-fitting SDT parameters, compute model-based performance measures such as area under the curve (AUC) and diagnosticity, use bootstrapping to determine uncertainty intervals around these parameters and measures, and compare parameters across two different data sets. The package incorporates closed-form solutions for both simultaneous and sequential lineups that allow for model-based analyses without Monte Carlo simulation. Show-ups are also supported. The package can estimate the base-rate of lineups that include a guilty suspect when the guilt or innocence of each suspect in the data set is unknown, as in "real-world" lineups. The package can also produce a full set of graphs, including data and model-based ROC curves and the underlying SDT model.

High- and low-threshold models of the relationship between response time and confidence.

The author compared high- and low-threshold discrete-state models of recognition memory in terms of their ability to account for confidence and response time (RT) data. The 2-high threshold (2HT), 1-low threshold (1LT), and 2-low threshold (2LT) models were clearly distinguished by the commonly observed inverted-U pattern whereby RTs are longer for low-confidence than high-confidence responses on both sides of the confidence scale (correct responses and errors). The 2HT model was able to match the RT-confidence relationship for correct responses, but it was unable to match the same relationship for errors. The 1LT model could not match the RT-confidence relationship for either correct responses or errors. Only the 2LT model was able to match the full pattern. The differences between models were driven by their fundamental assumptions about memory retrieval: only the 2-threshold models could produce an RT-confidence relationship by mixing relatively fast responses from a detection state with relatively slow responses from an uncertain ("guess") state, and only the 2LT model could do so for both correct and error responses because it allows misleading detection. Quantitative fits also showed that the 1LT model could not account for changes in confidence-rating distributions across memory-strength conditions, and thus this model performed substantially worse than the other two models even when RT data were not considered. (PsycInfo Database Record (c) 2021 APA, all rights reserved).

Estimating the proportion of guilty suspects and posterior probability of guilt in lineups using signal-detection models.

BACKGROUND: The majority of eyewitness lineup studies are laboratory-based. How well the conclusions of these studies, including the relationship between confidence and accuracy, generalize to real-world police lineups is an open question. Signal detection theory (SDT) has emerged as a powerful framework for analyzing lineups that allows comparison of witnesses' memory accuracy under different types of identification procedures. Because the guilt or innocence of a real-world suspect is generally not known, however, it is further unknown precisely how the identification of a suspect should change our belief in their guilt. The probability of guilt after the suspect has been identified, the posterior probability of guilt (PPG), can only be meaningfully estimated if we know the proportion of lineups that include a guilty suspect, P(guilty). Recent work used SDT to estimate P(guilty) on a single empirical data set that shared an important property with real-world data; that is, no information about the guilt or innocence of the suspects was provided. Here we test the ability of the SDT model to recover P(guilty) on a wide range of pre-existing empirical data from more than 10,000 identification decisions. We then use simulations of the SDT model to determine the conditions under which the model succeeds and, where applicable, why it fails. RESULTS: For both empirical and simulated studies, the model was able to accurately estimate P(guilty) when the lineups were fair (the guilty and innocent suspects did not stand out) and identifications of both suspects and fillers occurred with a range of confidence levels. Simulations showed that the model can accurately recover P(guilty) given data that matches the model assumptions. The model failed to accurately estimate P(guilty) under conditions that violated its assumptions; for example, when the effective size of the lineup was reduced, either because the fillers were selected to be poor matches to the suspect or because the innocent suspect was more familiar than the guilty suspect. The model also underestimated P(guilty) when a weapon was shown. CONCLUSIONS: Depending on lineup quality, estimation of P(guilty) and, relatedly, PPG, from the SDT model can range from poor to excellent. These results highlight the need to carefully consider how the similarity relations between fillers and suspects influence identifications.

How Much Time Do People With Aphasia Need to Respond During Picture Naming? Estimating Optimal Response Time Cutoffs Using a Multinomial Ex-Gaussian Approach.

Purpose Aphasia is a language disorder caused by acquired brain injury, which generally involves difficulty naming objects. Naming ability is assessed by measuring picture naming, and models of naming performance have mostly focused on accuracy and excluded valuable response time (RT) information. Previous approaches have therefore ignored the issue of processing efficiency, defined here in terms of optimal RT cutoff, that is, the shortest deadline at which individual people with aphasia produce their best possible naming accuracy performance. The goals of this study were therefore to (a) develop a novel model of aphasia picture naming that could accurately account for RT distributions across response types; (b) use this model to estimate the optimal RT cutoff for individual people with aphasia; and (c) explore the relationships between optimal RT cutoff, accuracy, naming ability, and aphasia severity. Method A total of 4,021 naming trials across 10 people with aphasia were scored for accuracy and RT onset. Data were fit using a novel ex-Gaussian multinomial RT model, which was then used to characterize individual optimal RT cutoffs. Results Overall, the model fitted the empirical data well and provided reliable individual estimates of optimal RT cutoff in picture naming. Optimal cutoffs ranged between approximately 5 and 10 s, which has important implications for assessment and treatment. There was no direct relationship between aphasia severity, naming RT, and optimal RT cutoff. Conclusion The multinomial ex-Gaussian modeling approach appears to be a promising and straightforward way to estimate optimal RT cutoffs in picture naming in aphasia. Limitations and future directions are discussed.

The Quality of Response Time Data Inference: A Blinded, Collaborative Assessment of the Validity of Cognitive Models.

Most data analyses rely on models. To complement statistical models, psychologists have developed cognitive models, which translate observed variables into psychologically interesting constructs. Response time models, in particular, assume that response time and accuracy are the observed expression of latent variables including 1) ease of processing, 2) response caution, 3) response bias, and 4) non-decision time. Inferences about these psychological factors, hinge upon the validity of the models' parameters. Here, we use a blinded, collaborative approach to assess the validity of such model-based inferences. Seventeen teams of researchers analyzed the same 14 data sets. In each of these two-condition data sets, we manipulated properties of participants' behavior in a two-alternative forced choice task. The contributing teams were blind to the manipulations, and had to infer what aspect of behavior was changed using their method of choice. The contributors chose to employ a variety of models, estimation methods, and inference procedures. Our results show that, although conclusions were similar across different methods, these "modeler's degrees of freedom" did affect their inferences. Interestingly, many of the simpler approaches yielded as robust and accurate inferences as the more complex methods. We recommend that, in general, cognitive models become a typical analysis tool for response time data. In particular, we argue that the simpler models and procedures are sufficient for standard experimental designs. We finish by outlining situations in which more complicated models and methods may be necessary, and discuss potential pitfalls when interpreting the output from response time models.

Speed-Accuracy Trade-Offs and Adaptation Deficits in Aphasia: Finding the "Sweet Spot" Between Overly Cautious and Incautious Responding.

Purpose After stroke, how well do people with aphasia (PWA) adapt to the altered functioning of their language system? When completing a language-dependent task, how well do PWA balance speed and accuracy when the goal is to respond both as quickly and accurately as possible? The current work investigates adaptation theory ( Kolk & Heeschen, 1990 ) in the context of speed-accuracy trade-offs in a lexical decision task. PWA were predicted to set less beneficial speed-accuracy trade-offs than matched controls, and at least some PWA were predicted to present with adaptation deficits, with impaired accuracy or response times attributable to speed-accuracy trade-offs. Method The study used the diffusion model ( Ratcliff, 1978 ), a computational model of response time for simple 2-choice tasks. Parameters of the model can be used to distinguish basic processing efficiency from the overall level of caution in setting response thresholds and were used here to characterize speed-accuracy trade-offs in 20 PWA and matched controls during a lexical decision task. Results Models showed that PWA and matched control groups did not differ overall in how they set response thresholds for speed-accuracy trade-offs. However, case series analyses showed that 40% of the PWA group displayed the predicted adaptation deficits, with impaired accuracy or response time performance directly attributable to overly cautious or overly incautious response thresholds. Conclusions Maladaptive speed-accuracy trade-offs appear to be present in some PWA during lexical decision, leading to adaptation deficits in performance. These adaptation deficits are potentially treatable, and clinical implications and next steps for translational research are discussed.

Adding a speed-accuracy trade-off to discrete-state models: A comment on Heck and Erdfelder (2016).

Heck and Erdfelder (2016) developed a model that extends discrete-state multinomial processing tree models to response time (RT) data. Their model is an important advance, but it does not have a mechanism to produce the speed-accuracy trade-off, the bedrock empirical observation that rushed decisions are less accurate. I present a similar model, the "discrete-race" model, with a simple mechanism for the speed-accuracy trade-off. In the model, information that supports detection of the stimulus type is available for some proportion of items and unavailable for others. Both the amount of time needed for detection to succeed and the amount of time that the decision maker waits before guessing are variable from trial to trial. Responses are based on detection when it is available and has a finishing time before the guess time for that trial. In other words, the decision maker sometimes loses opportunities to respond correctly on the basis of detection by first making a guess. These lost opportunities are more common when the guess-time distribution tends to have low wait times, which decreases accuracy. I report simulations showing that the model can accurately recover parameter values and is strongly constrained by the speed-accuracy trade-offs across conditions with different levels of response caution.

The speed of memory errors shows the influence of misleading information: Testing the diffusion model and discrete-state models.

In this report, we evaluate single-item and forced-choice recognition memory for the same items and use the resulting accuracy and reaction time data to test the predictions of discrete-state and continuous models. For the single-item trials, participants saw a word and indicated whether or not it was studied on a previous list. The forced-choice trials had one studied and one non-studied word that both appeared in the earlier single-item trials and both received the same response. Thus, forced-choice trials always had one word with a previous correct response and one with a previous error. Participants were asked to select the studied word regardless of whether they previously called both words "studied" or "not studied." The diffusion model predicts that forced-choice accuracy should be lower when the word with a previous error had a fast versus a slow single-item RT, because fast errors are associated with more compelling misleading memory retrieval. The two-high-threshold (2HT) model does not share this prediction because all errors are guesses, so error RT is not related to memory strength. A low-threshold version of the discrete state approach predicts an effect similar to the diffusion model, because errors are a mixture of responses based on misleading retrieval and guesses, and the guesses should tend to be slower. Results showed that faster single-trial errors were associated with lower forced-choice accuracy, as predicted by the diffusion and low-threshold models.

Guessing versus misremembering in recognition: A comparison of continuous, two-high-threshold, and low-threshold models.

The two-high-threshold (2HT) model of recognition memory assumes that people make memory errors because they fail to retrieve information from memory and make a guess, whereas the continuous unequal-variance (UV) model and the low-threshold (LT) model assume that people make memory errors because they retrieve misleading information from memory. We explored the nature of memory errors by comparing guessing and memory performance. In 2 experiments, participants studied lists of words followed by a test in which each trial was preceded by a cue indicating the probability that the trial would have a studied word. Participants first guessed whether or not the word would be studied, and then they saw the word and responded again. When the response that was more likely according to the cue was the correct response, participants made more errors after attempting to remember the word than in their initial guesses. This suggests that participants made errors because they retrieved misleading information even when they could guess the correct response on the basis of the probability cue. We also compared the models in terms of their ability to fit ROC functions using parametric bootstrap procedures to correct for model mimicry. These analyses supported both the UV and LT models over the 2HT model. (PsycINFO Database Record

Response biases in simple decision making: Faster decision making, faster response execution, or both?

When people are biased to use one response more often than an alternative response in a decision task, they also make the preferred response more quickly. Sequential sampling models can accommodate this difference in response time (RT) by changing the relative amount of evidence that must accumulate to decide in favor of one versus the other response, but nondecision processes might also play a role, such as the amount of time between selecting and executing a response. We investigated the influence of decision and nondecision processes in two experiments. In Experiments 1a and 1b, arrows appeared on the screen, and participants were asked to move a joystick in the direction of the arrow or make a keypress as quickly as possible. Results showed that motor execution times were faster for expected directions than unexpected directions. In Experiments 2a and 2b, participants decided whether a high or low number of asterisks was displayed on the screen. Decision times were faster for the stimulus class that was more likely to appear, and this effect was larger when participants could anticipate both the likely stimulus class and the motor response needed to identify it than when they knew the likely stimulus class but the associated motor response changed probabilistically from trial to trial. These results show that both decision and nondecision factors contribute to bias effects on RT.

Successful cuing of gender source memory does not improve location source memory.

In three experiments we explored cross-dimensional cuing effects in a multidimensional source encoding and retrieval paradigm. We employed a bias-controlled experimental method of source cuing at retrieval (Starns & Hicks, 2013) in an attempt to improve retrieval of location information indirectly by cuing gender information. Encoded words were situated on the left or right side of a computer monitor and associated with either a male or a female face. When multiple faces were used across the set of encoded words, reinstating the correct face at retrieval alongside an incorrect, opposite-gender face cue improved male/female source decisions for test words. However, this powerful test cue did not improve memory for the encoded location of the words, suggesting that within-dimension cuing does not produce cross-dimensional cuing. This null outcome was found when gender decisions were required (Experiments 1A and 2) or not required (Experiment 1B) prior to location decisions. Nor was cross-dimension cuing found when subjects were told to expect a source test of both gender and location information at retrieval (Experiment 2). Our findings reinforce prior work demonstrating that multiple context dimensions can be bound to item information without any direct binding between the contexts.

Item strength influences source confidence and alters source memory zROC slopes.

Increasing the number of study trials creates a crossover pattern in source memory zROC slopes; that is, the slope is either below or above 1 depending on which source receives stronger learning. This pattern can be produced if additional learning affects memory processes such as the relative contribution of recollection and familiarity to source performance. However, the pattern can also be produced by decision processes if participants are more willing to make high-confidence source judgments when they are more confident that the test item was studied. We explored the role of memory and decision processes by comparing performance across 3 conditions: (a) words seen once with a male or female face (no repetition), (b) words seen once with a face after being presented twice with a picture of either a bird or a fish (different-source repetition), and (c) words seen 3 times with the same face (same-source repetition). zROC functions for the male-female decision showed that different-source repetition produced the same crossover effect as same-source repetition. This pattern was predicted by the decision process account, because it assumes that increasing item memory affects source confidence ratings even if source memory is not improved. Also supporting this account, we found a strong positive relationship between recognition confidence and source confidence even when analyses were limited to items that were attributed to the incorrect source or items that were not studied in either source.

Shifting the criterion is not the difficult part of trial-by-trial criterion shifts in recognition memory.

In recognition memory, participants often fail to shift their response criterion within a test even when they see cues signaling whether they should expect weak or strong memory (e.g., Stretch & Wixted Journal of Experimental Psychology: Learning, Memory, and Cognition, 24, 1397-1410, 1998b). We contrasted two possible accounts for this failure to shift. The first assumes that shifting the criterion requires effortful processing, so participants are unwilling to make shifts even when they expect different levels of strength. The second assumes that participants are unwilling to decide which strength category is indicated by the cue for each trial, so their expectations for memory strength do not change across trials. Targets appeared in different test formats ("cues") depending on whether they were studied once (weak) or five times (strong), and lures were evenly divided between the two formats. Some participants had two response keys for "old" and "new" (2-key), and others had to use different keys to respond "old" for the two strength cues (3-key). The goal of the 3-key condition was to force participants to decide which strength cue was presented on each trial. The 3-key participants had a lower false alarm rate for lures shown with the strong than with the weak cue, but the 2-key participants showed no evidence of a criterion shift. Response times were unaffected by trial-by-trial criterion shifts. We conclude that participants willingly shift their response criterion on the basis of changes in expected strength, but they are unwilling to decide which strength to expect unless they are compelled to do this by other aspects of the task.

A violation of the conditional independence assumption in the two-high-threshold model of recognition memory.

The 2-high-threshold (2HT) model of recognition memory assumes that test items result in distinct internal states: they are either detected or not, and the probability of responding at a particular confidence level that an item is "old" or "new" depends on the state-response mapping parameters. The mapping parameters are independent of the probability that an item yields a particular state (e.g., both strong and weak items that are detected as old have the same probability of producing a highest-confidence "old" response). We tested this conditional independence assumption by presenting nouns 1, 2, or 4 times. To maximize the strength of some items, "superstrong" items were repeated 4 times and encoded in conjunction with pleasantness, imageability, anagram, and survival processing tasks. The 2HT model failed to simultaneously capture the response rate data for all item classes, demonstrating that the data violated the conditional independence assumption. In contrast, a Gaussian signal detection model, which posits that the level of confidence that an item is "old" or "new" is a function of its continuous strength value, provided a good account of the data.