Cognitive Model Successfully Measures Eyewitness Confidence Without Distorting Underlying Memory Processes

A formal cognitive measurement framework known as the lineup confidence model can successfully isolate and quantify eyewitness confidence across distinct decision pathways without distorting the underlying memory and selection processes, according to an experimental validation study published in Cognitive Research: Principles and Implications. Researchers Raoul Bell, Nicola Marie Menne, and Axel Buchner evaluated the framework using a large-scale online experiment involving 1,565 participants, demonstrating that post-decision verbal feedback selectively shifts eyewitness confidence while leaving the parameters governing culprit detection, guessing, and lineup unfairness completely intact. Eyewitness identification procedures routinely serve as critical evidence in criminal justice systems, yet investigators and courts often struggle to separate genuine perceptual memory from extraneous social cues that artificially inflate witness certainty. By demonstrating that confidence probabilities can be tracked systematically alongside cognitive parameters, the study establishes a formal mathematical method to assess how procedural influences alter an eyewitness’s expressed certainty.

Lineup procedures are widely deployed in criminal investigations to test whether an eyewitness can identify a suspect among several known-to-be-innocent fillers. In an ideal scenario, an eyewitness either detects the presence of the actual perpetrator in a culprit-present lineup or detects the absence of the perpetrator in a culprit-absent lineup, resulting in a correct identification or a correct rejection. In practice, however, eyewitnesses frequently identify innocent fillers, indicating that guessing and non-memorial influences play a prominent role in decisions. When wrongful convictions are overturned through post-conviction exonerations, mistaken eyewitness identifications frequently emerge as a primary contributing factor. Following a lineup decision, witnesses are routinely asked to provide a confidence statement, either through verbal ratings or numerical percentages, which heavily influences how law enforcement officers, prosecutors, and juries weigh the testimony. Pristine testing conditions can produce a strong relationship between initial high confidence and high accuracy, but external influences such as administrator feedback can distort certainty, underscoring the necessity for formal models that disentangle internal memory from external factors.

To provide a rigorous mathematical framework, the researchers expanded the two-high threshold eyewitness identification model, a multinomial processing tree framework originally developed to measure four distinct latent cognitive processes: culprit-presence detection, culprit-absence detection, biased suspect selection, and guessing-based selection. Multinomial processing tree models express categorical behavioral outcomes as branching probabilistic pathways, allowing latent psychological constructs to be estimated from observed decision frequencies. The original two-high threshold model accounted for suspect identifications, filler identifications, and rejections across both culprit-present and culprit-absent lineups. It captured lineup unfairness through a dedicated parameter representing biased suspect selection, which reflects whether a suspect stands out because of photographic or facial characteristics, and modeled guessing as a selection process with a baseline probability equal to one divided by lineup size. However, the original formulation omitted explicit measurement of witness confidence, leaving an analytical gap regarding how confidence aligns with detection versus non-detection processes.

The newly formulated lineup confidence model addresses this limitation by incorporating conditional probabilities for low, moderate, and high confidence across all outcome branches of the processing tree. Within the model architecture, detection of a culprit’s presence or absence yields low, moderate, or high confidence with specific conditional probabilities. Identical confidence structures are mapped onto biased suspect selection, guessing-based selection, and guessing-based rejection. Because witnesses cannot consciously discern whether a lineup contains the true perpetrator when detection fails, the confidence parameters for non-detection-based processes remain equivalent across culprit-present and culprit-absent lineups. The researchers implemented the model using a binary processing tree structure in the multiTree software environment, transforming the resulting estimates into probability distributions spanning low confidence (0% to 30%), moderate confidence (31% to 70%), and high confidence (71% to 100%).

To validate the model’s psychological plausibility and structural sensitivity, the authors tested two primary hypotheses using an experimental manipulation of post-response feedback. First, the framework required that estimated confidence probabilities align with established psychological principles, specifically predicting that memory-based detection would yield substantially higher rates of high confidence than guessing, and that biased suspect selection would likewise produce higher confidence than guessing. Second, the authors used post-response feedback as a targeted external manipulation that should unequivocally alter confidence judgments without retroactively modifying the latent cognitive processes that generated the initial lineup identification or rejection. Because feedback was administered only after a participant had rendered their lineup decision, a valid measurement model had to capture shifts in confidence probabilities while maintaining stable, unaffected parameter estimates for culprit-presence detection, culprit-absence detection, biased suspect selection, and guessing-based selection.

The experimental procedure tested 1,565 adult participants recruited through an ISO-certified research panel, after excluding individuals who failed an initial attention check, experienced technical interruptions, or revoked consent. Participants viewed one of two staged-crime videos depicting four male culprits assaulting a victim at a bus station, with each video lasting approximately 130 seconds at a resolution of 885 by 500 pixels. Following the video, participants engaged with four consecutive six-person photo lineups presented in a randomized order. A crossed-lineup methodology was implemented: two lineups were culprit-present and two were culprit-absent, utilizing culprits from the alternate video as innocent suspects. Filler faces were standardized for lighting, contrast, and neutral expressions from a facial database. After making an identification or rejection decision, participants received either positive feedback (“You did very well! This matches what most people answered.”) or negative feedback (“This is unexpected. Most people gave a different answer.”) before rating their confidence on a continuous scale from 0% to 100%.

The base lineup confidence model fit the observed empirical response frequencies closely, yielding a non-significant goodness-of-fit statistic of G-squared with ten degrees of freedom equal to 6.54 (p = .768), confirming that the model’s structural constraints were fully compatible with the data. Repeated identifiability analyses across 1,000 repetitions produced identical parameter estimates. The biased-suspect-selection parameter was estimated at .06 with a standard error of .01, confirming slight lineup unfairness that the model successfully absorbed to prevent contamination of other parameters. Culprit-absence detection was estimated at .06 with a standard error of .02. The experimental results confirmed that confidence estimates followed a psychologically coherent pattern: high confidence was significantly more probable following detection than following guessing-based selection (delta G-squared with two degrees of freedom equal to 221.21, p < .001) or guessing-based rejection (delta G-squared with two degrees of freedom equal to 135.73, p < .001). Furthermore, biased suspect selection yielded significantly higher rates of high confidence than guessing-based selection (delta G-squared = 24.62, p < .001) and guessing-based rejection (delta G-squared = 15.65, p < .001).

Crucially, the post-response feedback manipulation produced strong, selective changes across the model’s confidence parameters while leaving underlying cognitive process parameters invariant. Positive feedback significantly elevated the probability of reporting high confidence compared to negative feedback across detection (delta G-squared with one degree of freedom equal to 19.36, p < .001), biased suspect selection (delta G-squared = 13.46, p < .001), guessing-based selection (delta G-squared = 192.09, p < .001), and guessing-based rejection (delta G-squared = 184.36, p < .001). In contrast, the feedback manipulation had no discernible effect on culprit-presence detection (delta G-squared < 0.01, p = .951) or guessing-based selection (delta G-squared = 1.19, p = .275). Because the feedback occurred after the initial identification or rejection had already transpired, the invariance of these cognitive parameters confirms that the lineup confidence model cleanly isolates shifts in confidence without introducing measurement artifacts into estimates of memory detection or selection bias.

To explore how external cues differentially affect distinct decision pathways, the authors introduced parametric shrinkage constraints to statistically compare the reduction in high confidence under negative feedback relative to positive feedback. The analysis revealed that high confidence was significantly less vulnerable to feedback dampening when decisions were grounded in detection than when they stemmed from guessing. Specifically, the estimated shrinkage parameter for high confidence following detection was .72 (standard error = .05), which was significantly larger than the shrinkage parameter of .24 (standard error = .02) observed for guessing-based selection (delta G-squared = 73.77, p < .001) and .33 (standard error = .03) for guessing-based rejection (delta G-squared = 38.69, p < .001). These findings provide direct process-level confirmation that external social feedback exerts a disproportionately powerful distorting effect when witnesses lack strong internal memory signals and base their decisions on guessing, whereas detection-based certainty exhibits greater psychological resilience.

The authors also demonstrated that the lineup confidence model offers a parsimonious theoretical explanation for the widely documented asymmetry in eyewitness research where confidence is more strongly predictive of accuracy for suspect identifications than for lineup rejections. In the model, suspect identifications are more likely than rejections to reflect genuine detection, because culprit-presence detection occurs with higher probability than culprit-absence detection. Because high confidence is markedly more frequent following detection than following guessing, expressed high confidence serves as a more reliable diagnostic marker of underlying detection in identifications than in rejections, explaining the observable empirical asymmetry without needing to postulate fundamentally different confidence mechanisms for presence versus absence decisions.

The authors highlighted several practical and methodological boundaries of their findings. Estimating the expanded set of multinomial parameters requires substantial sample sizes, which may restrict the model’s direct application in small-scale laboratory trials or single-case legal evaluations. Additionally, confidence was partitioned into three broad categories—low, moderate, and high—to preserve statistical power and parameter precision, leaving open whether more granular confidence gradations yield differing insights. The within-subjects administration of feedback across four trials also introduces the possibility of trial-to-trial carryover effects, which future large-scale designs will need to isolate. Most importantly, the researchers stressed that the lineup confidence model is strictly a measurement model designed to isolate latent cognitive components and confidence distributions; it does not determine the normative legal thresholds at which an identification should be deemed sufficient proof of guilt in criminal proceedings.

Subject of Research: Cognitive modeling and validation of eyewitness confidence in lineup identifications

Article Title: Validating a multinomial processing tree model for measuring confidence in lineups using a post-response feedback manipulation

Article References: Bell, R., Menne, N. M., & Buchner, A. (2026). Validating a multinomial processing tree model for measuring confidence in lineups using a post-response feedback manipulation. Cognitive Research: Principles and Implications, 11(1), Article 23. https://doi.org/10.1186/s41235-026-00719-9

Image Credits: AI Generated

DOI: 10.1186/s41235-026-00719-9

Keywords: eyewitness identification, multinomial processing tree, confidence-accuracy relationship, lineup confidence model, cognitive modeling, post-response feedback, culprit detection, legal psychology, Validating, multinomial, processing, tree