Consequences of Confidence

The current project is a follow up study based on a set of intriguing findings from an initial study that was designed to examine the effects of rendering confidence during recognition decisions. In that study subjects engaged in two sequential recognition tests following the study of a single list of word stimuli. During Test 1 subjects viewed studied words intermixed with new words and rendered old/new recognition judgments. Following this, all the words from Test 1 were represented during Test 2, along with a small number of new lures, and subjects were asked to recognize any words that had been seen before regardless of their prior recognition decisions about them. Critically, one group of subjects rendered confidence judgments during both tests whereas the other only rendered confidence during the second test. The current study focuses on this former condition. An exploratory analysis in this group found two interesting findings. First, confidence during Test 1 for hits (and false alarms) positively predicted the likelihood that these items would again be recognized during Test 2. This is consistent with basic signal detection/strength models of recognition. However, this same model, unintuitively, predicts that confidence during Test 1 for correct rejections (and misses) should be anti-predictive of Test 2 recognition. This is because high confidence that an item is new during Test 1, under this model, means that the item has exceptionally low strength, and hence, compared to low confidence that an item is new, should be harder to subsequently recognize. However, there was no relationship between the confidence of correct rejections at Test 1 and their likelihood of being recognized at Test 2. These findings were confirmed using multilevel modeling for correct rejections and hits from Test 1, where Test 1 confidence was used to predict Test 2 hit or miss outcomes. The model yielded an interaction of Test 1 status (hit vs correct rejection) and confidence during Test 1. Simple effects follow ups demonstrate the afore mentioned positive relation for Test 1 hit confidence, and no reliable relationship for Test 1 correct rejection confidence. Errors were too sparse to be effectively modeled. Thus these data suggest that signal detection strength models make a categorically incorrect prediction about the confidence expressed during correct rejections and future item recognition. The second interesting finding was that for Test1 hit and correct rejection items that were also correctly recognized on Test 2, there was a strong positive correlation between Test 1 and Test 2 confidence. Under a signal detection strength model this is unsurprising for Test 1 hits. However, as noted above, Test 1 correct rejection confidence did not predict Test 2 hit probabilities. Hence, the positive relationship between confidence for Test 1 correct rejections and Test 2 recognition confidence, conditional on recognition, reflects some other phenomenon. We hypothesize that this reflects an item effect unrelated to actual recognition evidence. That is, we hypothesize that there is some characteristic of items, unrelated to memorial evidence, that tends to spur high or low confidence when those items are endorsed. One potential source of this could simply be memory for the prior confidence report action. The current study addresses the potential role of item effects and eliminates the possibility of subjects using remembrance of Test 1 responses to render Test 2 decisions. Both the finding that confidence for new judgments does not predict future recognition in accordance with signal detection strength models, and that confidence reports may demonstrate large item effects during recognition, have implication for metacognition and decision models.