Confidence ratings and non-diagnostic information : a theoretical account of confidence for recognition memory and perceptual discrimination tasks

In a variety of domains, ratings have been used (in the absence of a binary yes/no or old/new judgement) to index recognition. More recently, this approach has been implemented in the face recognition and eyewitness identification literature. Such an approach requires participants to provide a confidence rating (e.g., from 0-100%; Sauer, Brewer & Weber, 2008; Sauer, Weber & Brewer, 2012) as to whether they have viewed a specific face before, without asking for a binary identification judgement. Evidence suggests that such ratings can be used to indicate the likelihood that an image or lineup member has been seen before, and may even provide a more sensitive index of recognition than categorical responses (e.g., Brewer, Weber, Wootton & Lindsay, 2012; Cleary & Greene, 2000; Sauer, Brewer & Weber, 2008; Sauer, Weber & Brewer, 2012). However, there has been little direct investigation of the theoretical mechanisms that underlie these ratings. Research into the mechanisms that shape Feelings of Knowing (judgements that non-recalled information might be cued in a later recognition task; Koriat, 1997) and Judgements of Learning (assessments of the likelihood of future recall for studied information; Koriat, 1993) suggest that metacognitive judgements can be shaped by non-memorial cues, such as heuristics pertaining to perceptions of fluency at test, or naïve theories about how the learning/encoding conditions are likely to affect later recall or recognition. Such heuristic processes reflect perceptions of the experiences associated with encoding and retrieval, rather than reflecting evidence gleaned directly from memory. Thus, they are vulnerable to systematic distortions. Conversely, theories with their origins in the domain of psychophysics suggest that confidence indexes stimulus discriminability and/or memory strength (e.g., signal detection theory, Bernbach, 1971; Egan, Schulman, & Greenberg, 1959; Green & Swets, 1966; or accumulator models using a balance of evidence mechanism to account for confidence, Vickers, 1970; Van Zandt, 2000). The primary focus of this thesis is to test the extent to which rating-only judgements (i.e., confidence in recognition or perceptual discrimination) are reflective of the physical properties of the stimuli (as per psychophysical models of confidence) or shaped by heuristic processes (i.e., as per inferential models of metacognition; appraisals of the reliability of, or our confidence in, our own cognitions). Specifically, we are interested in whether confidence reflects the diagnosticity of the information at test (i.e., the extent to which the accumulated evidence favours one response over another) or simply the amount information available at test. Further, we are interested in whether these effects are consistent across both recognition memory tasks and perceptual discrimination tasks. While psychophysical models of confidence assume that effects observed in psychophysics should generalise to recognition tasks, the reliance on memory adds noise to the process, opening the door for more inferential processes associated with fluency etc. Chapter 3 reports three experiments testing the theoretical mechanisms that shape recognition confidence ratings. Experiment 1 uses a face recognition paradigm, Experiment uses a landscape/house recognition paradigm to test whether effects from Experiment 1 generalise beyond faces to other complex stimuli that might be less reliant on holistic processing, and Experiment 3 uses a perceptual discrimination task asking participants to indicate their confidence that a dynamic grid was primarily blue or primarily orange. Extending upon findings by Busey, Tunicliff, Loftus and Loftus (2000) suggesting that the presence of non-diagnostic information at test can inflate confidence, we in turn implemented a manipulation of non-diagnostic information. In Experiments 1 (N = 60) and 2 (N = 113), this was operationalised through the use of full and partial images. To elaborate, participants may have viewed the top half of a face (or house/landscape) during the study phase, but been presented with the full version of the image during the test phase. In this example, we would view the additional bottom-half of the image as “non-diagnostic” information, as it was not present at encoding. In Experiment 3 (N = 86), participants were asked to indicate their confidence as to whether a dynamic grid was predominantly blue or predominantly orange (i.e., “Indicate your confidence that the flashing grid is predominantly orange”, with the target colour differing between blocks of trials). Non-diagnostic information was included through the addition of white pixels to the grid. To find support for an inferential model of confidence, we would expect to see an increase in confidence associated with the addition of non-diagnostic information because, according to inferential models, metacognitive judgements are based at least in part on the volume of information available at test regardless of the accuracy or diagnostic value of that information. To find support for psychophysical models, we would need to see no changes in confidence associated with the addition of non-diagnostic information at test, as the amount of diagnostic information is unaffected by the addition of non-diagnostic information. Instead, what we discovered across all three experiments was that confidence decreased when the presence of non-diagnostic information increased. These findings align with a little-used psychophysical theory: Baranski and Petrusic’s (1998) doubt-scaling model, which suggests that confidence is inversely proportionate to the amount of non-diagnostic information present at the time of decision. Having seemingly found support for Baranski and Petrusic’s (1998) doubt-scaling model – a model first formally proposed in 1998 and then largely ignored – we further explored the nature of the relationship between non-diagnostic information and confidence. Thus, we conducted a series of perceptual tasks (Chapter 4) that allowed us to examine the effects of different levels of non-diagnostic information on confidence, reaction time (RT) and accuracy in a highly controlled experimental paradigm. In Experiment 1 (N = 56), participants completed 720 trials, in which they were presented with a dynamic grid consisting of flashing blue, orange and white pixels and asked to indicate whether the grid was predominantly blue or orange (using a response scale ranging from low confidence blue, moderate confidence blue, high confidence blue, low confidence orange etc.), with the white pixels constituting non-diagnostic information. Unlike the experiments detailed in Chapter 3, this study asked participants to make a simultaneous confidence/decision judgment rather than a standalone confidence judgement to allow us to observe the effects of non-diagnostic information on accuracy, response time (RT), and confidence. The results showed a reduction in both confidence and accuracy associated with an increase in the proportion of non-diagnostic information, together with an increase in RTs. Further, CAF analyses (which plot accuracy as a function of RT) showed that as the proportion of non-diagnostic information increased, the accuracy of fast responses decreased in comparison to the accuracy of slow responses. To determine whether non-diagnostic information has a direct effect on accuracy (as opposed to the decrease in accuracy being a by-product of the simultaneous confidence judgement and decision judgment), we ran a second experiment (N = 21). This time, participants viewed the same stimuli, but were asked only to indicate whether they believed the dynamic grid to be majority blue or majority orange. We found a pattern of results similar to those for Experiment 1 regarding the relationship between non-diagnostic information and accuracy (RTs). Again, CAFs showed that the accuracy of fast responses decreased in comparison to the accuracy of slow responses as the proportion of non-diagnostic information increased. Much like the experiments outlined in Chapter 3, these results add to the existing literature by demonstrating a negative relationship between non-diagnostic information and confidence. In line with predictions made by Baranski and Petrusic, the current study demonstrated a negative relationship between non-diagnostic information and accuracy, as well as a general slowing of RTs as the proportion of non-diagnostic information increased. Chapter 5 reports two experiments that tested whether a rating-only, confidence approach (similar to that used in Chapter 3) could be used in an applied visual search task to indicate the likelihood of a target being present. In Experiment 1 (N = 24), we confirmed that a rating-only approach could be used in a basic perceptual paradigm to indicate the likelihood of a dynamic grid being majority blue or majority orange (the stimuli was similar to that used in Chapters 3 and 4, but without the additional white pixels). We found that participants were well calibrated in their responding, indicating that they were able to use the rating-only scale to convey the likelihood that the stimuli is majority blue or majority orange. In Experiment 2 (N = 99) we tested whether this could be extended to an applied visual search task: Searching luggage X rays for evidence of a weapon. Specifically, we tested whether confidence ratings could discriminate trials containing a weapon from those that did not, and whether using ratings could mitigate the low prevalence effect (where shifts in decision criterion mean rarely-occurring targets often go undetected). The results showed that despite increases in overconfidence associated with lower prevalence rates, a rating-only approach effectively discriminated target present from target absent trials, even when target prevalence was low. In sum, the experiments outlined here expand the existing literature on confidence judgements in two main ways. First, our research suggests that confidence judgements (whether accompanied by a decision judgment or not) are influenced by the presence of non-diagnostic information, providing support for Baranski and Petrusic’s (1998) doubt-scaling model of confidence. Second, our findings suggest that a rating-only approach can be used in applied perceptual discrimination tasks to indicate the likelihood of a target being present in a complex array