Experiment 2 was based on prior work by Hemmer and Steyvers (2009a) who examined the impact of prior knowledge on episodic memory. In their work, Hemmer and Steyvers compared memory for the size of familiar items (fruit, vegetables) with memory for the size of unfamiliar items (random shapes). The task used a form of continuous recognition where participants were presented with 72 item lists. Study and test trials were randomly interleaved so that studied items were tested at random intervals within the list; on a test trial, participants were first asked if they recognized the item as having been studied before and were then asked to resize recognized items to their original studied size. The lag between study and test could vary between one and 24 trials; it follows that most lags would be outside what is typical in the study of immediate/working memory. Moreover, performance at all lags was averaged in the analyses. The results suggested that episodic memory of the studied items was affected by (a) fine-grained, item-specific representations and (b) two levels of categorical information. For both familiar and unfamiliar shapes, there was a central-tendency bias as the recalled size was systematically influenced by the mean size of the stimuli in the category. The results with familiar stimuli demonstrated the influence of a second categorical factor: item-level prior knowledge (e.g., the average size of apples).
In Experiment 2, we asked if the findings of Hemmer and Steyvers (2009a) could also be found in a VWM task. We used lists containing familiar items (photographs of vegetables) or unfamiliar ones (random shapes). As before, six items were sequentially presented, but in this case, at test, participants were to reconstruct the size of one of the studied objects.
From Hemmer and Steyvers (2009a) normative data were available for the familiar items; these included the normative average size (norm hereafter) for each item as well as the largest and smallest realistic sizes. We assumed these norms were reasonable approximations of the knowledge participants brought to the experiment regarding familiar item sizes. With the help of these data, items could be presented either above or below the norm. This made it possible to predict the direction of any knowledge-based bias at the item level. Specifically, we expected that the remembered size of a familiar object (i.e., the just-seen apple) would drift towards the object’s norm (i.e., the average apple size). Moreover, as before, we expected a central-tendency bias for both familiar and unfamiliar items whereby small items (a mushroom or a small shape) and large items (a cabbage or a large shape) would drift slightly towards the average size within the category. In essence, we tested predictions relating to two levels of knowledge: (1) for the familiar items, an object-level bias, where the size of each item is remembered as being slightly closer to its prototypical size and (2) for both types of items, a central-tendency bias where memory is influenced by the overall mean of item sizes presented within the experiment. Figure 3 summarizes the assumed influence of knowledge at both object and experiment levels.
Forty-two undergraduate students volunteered for the study. Some received course credits for their participation.
Stimuli were taken from Hemmer and Steyvers (2009a) and consisted of 24 high-resolution color photographs of vegetables against a white background as well as 24 images of random blue shapes.
These images were used to create 48 six-item lists, 24 lists of familiar items and 24 lists of unfamiliar ones. The familiar and unfamiliar items were yoked such that the presentation size of shapes was matched to that of the vegetables.
Study sizes of familiar items were determined as follows. In each list, two items were presented at their normative mean size, two items were larger than their normative mean size, and two items were smaller than said mean. All items were studied as often in all three sizes; however, tested items were always studied smaller or larger than their normative mean.
The sizes that were “larger” and “smaller” than the norm were obtained as follows. Recall that normative data contained three estimates: a normative mean size, a normative “smallest reasonable size” (e.g., the smallest realistic size for a radish), and a normative “largest reasonable size” (e.g., the largest realistic size for a radish). For each item, the range from the mean to the smallest reasonable size and the range from the mean to the largest reasonable size were calculated. Items presented smaller than their normative mean were presented at the size that was at 0.6 of the range from the mean to the smallest realistic size. So, if the mean size for a beetroot was 0.25 and the smallest realistic size for a beetroot was 0.15, then the range was 0.10, and a small beetroot would be presented at 0.19, that is [0.25 − (0.6 × 0.10)]. Likewise, the size of an item studied larger than its normative mean was set to be at 0.6 of the range between the mean and the largest realistic size for said item.
As for list composition, the 24 familiar items (and their yoked unfamiliar shapes) were divided into two groups based on their normative size; one group contained the 12 largest items while the other held the 12 smallest items. Forty-eight six-item lists were constructed so that: (a) three items were from the large group and three were from the small group. Also, items were divided into two sets of 12 items, matched for size. To improve experimental control, one set was used as targets for half the participants and the other set was used for the other half; this strategy allowed us to test the same item twice for each participant, once in a size above and once in a size below its normative mean; in essence, each item could be its own control and across participants, all items were used as targets. On each trial, a single item was selected from the list of six for testing. Each sequential position was tested equally often. Lists of unfamiliar items mirrored the construction of the familiar lists. There were also two practice trials created from the same stimuli that had the same structure.
The procedure was as in Experiment 1 except for the following. Images were presented for 1,000 msec each, followed by a 500-msec blank screen. Following the last item of a list, there was a further 1.5 s with a blank screen and then one of the items was presented again in a new size, randomly set to .2 (i.e., at a size corresponding to 20% of the display), .4, .6, or .8 of the presentation window. To reconstruct target sizes, participants moved a cursor placed in the center of a horizontal sliding bar (at bottom of the screen). Moving the mouse-controlled cursor to the left made the target smaller and moving it to the right made it larger.
Results and discussion
As in Hemmer and Steyvers (2009a), the presentation size was subtracted from the remembered size to obtain reconstruction error. A positive error indicates the item was remembered larger than studied; a negative error indicates the reverse. Figure 4 (a and b) presents the mean reconstruction error associated with each studied size separately for items studied larger than their norm, and items studied smaller than their norm. (For the unfamiliar shapes, the norm was taken to be the norm of the vegetable with which they were yoked). The left panel shows the data for familiar items (vegetables) and the right panel shows the data for the unfamiliar colored shape items. As expected, negative slopes were obtained in all conditions; small items were reconstructed larger and large items smaller, consistent with a central-tendency bias. For familiar items there were two distinct regression lines, corresponding to the items studied larger or smaller than their norm. In other words, two items studied at the same objective size can be remembered differently. If the studied size of one item was smaller than its norm, then it tended to be remembered as larger than it actually was. If the size of the corresponding item studied was larger than its normative size, it tended to be remembered as being slightly smaller than it was at study. For the unfamiliar items, the two regression lines were superimposed. As familiar and unfamiliar items were yoked in size, the difference must be due to the knowledge associated with the familiar items.
As before, we ran per participant regressions to analyze these findings. We first compared the two slopes obtained for the familiar items as well as those obtained for the unfamiliar items. We ran separate per participant regressions for the familiar items studied smaller than the norm and for those studied larger than the norm as well as the corresponding analyses for the unfamiliar conditions; the dependent variable was the error score and the predictor was the studied size. We then compared the slopes in a 2 (relative size, lager / smaller) × 2 (familiarity, vegetables / shapes) repeated measures ANOVA. There was no effect of relative size (F(1,41)<1, p=.61), a significant effect of category (F(1,41)= 56.1, p<0.001, and these factors did not interact (F(1,41)<1, p=.70). As Fig. 4 suggests, the slopes for each category (familiar / unfamiliar) are similar for each relative size; however, the mean slope for the unfamiliar items (−.58) is steeper than the mean slope for the familiar items (−.27). The slopes for both vegetables, t(41) = −8.6, p < .001, and shapes, t(41) = −15.6, p < .001) were significantly different from zero.
Following Hemmer and Steyvers (2009a), we then tested for the expected interaction between category (familiar/unfamiliar) and relative size (smaller/larger) for the intercepts. The hypothesis was that familiar items would show a knowledge-based bias through a difference in intercept as illustrated in Fig. 4a. As the unfamiliar items cannot benefit from equivalent knowledge, there should be no difference in intercept in this case, as suggested in Fig. 4b. Further per participant regressions were run for the familiar and unfamiliar items with the error score as the predicted variable. The predictors were the studied size along with a binary variable corresponding to whether an item was smaller or larger than its normative size. When averaged across participants, for the familiar items the two average intercept values were 0.12 and 0.16 respectively for the items studied larger and smaller than their normative means. For the unfamiliar items, the corresponding intercept values were .20 and .21. The means slopes were as above.
A 2 (familiar/shapes) × 2 (relative study size: smaller/larger than norm) ANOVA on the intercepts produced a significant effect of familiarity, F(1,41)= 20.2, p < .001, study size, F(1,41)= 42.1, p < .001 and more importantly the two factors interacted, F(1,41)= 7.7, p= .008. T-tests showed a significant difference between the intercepts observed for the familiar objects, t(41)= −6.0, p < .001, but not for shapes, t(41)= −1.8, p=.08. Thus, for the familiar items, objects studied at the same size, but respectively larger and smaller than the norm were not remembered in the same way. Items presented larger than the norm tended to be underestimated while items that were smaller than the norm were overestimated.