Visual identity and uncertainty in repetition blindness

Visual identity and uncertainty in repetition blindness GARY A. BRILL, ARNOLD L. GLASS, HANIN RASHID, and ERIKA HUSSEY Rutgers University

Repetition blindness (RB) was investigated in 6 experiments. In the first 3 experiments participants detected vowel targets in 11-letter sequences. When all letters were uppercase, detection was poorer for same (e.g., AA) than for different (e.g., AO) targets. However, when one target was uppercase and the other lowercase, RB was found only for targets visually identical except for size (e.g., Oo), not for visually different pairs (e.g., Aa). Experiment 4 found RB for visually identical versus different consonant–vowel–consonant words. Experiments 5 and 6 replicated Kanwisher’s (1987) experiment in which RB was insensitive to word case but revealed these effects to be artifacts of poor recognition of 5-letter words coupled with a biased guessing strategy. Overall, these experiments found RB only at a low level of visual information processing.

The term repetition blindness (RB) was coined by Kanwisher (1987) to describe the phenomenon in which an observer fails to report the repetition of an item in a rapidly presented sequence of items. In the first experiment of her study, when people viewed a sequence of words presented at a rate of eight items per second, the probability of detecting a repeated word was a positive function of the number of intervening items. That is, RB labeled a paradoxical effect of lag on repetition detection: The smaller the number of intervening items, the less often the repetition was detected. Kanwisher’s (1987) next experiment uncovered difficulty in reporting repetitions in a different type of task: People sometimes failed to report a repeated word during verbatim recall of a briefly presented sentence. Because lag was not a factor in this experiment and because comparison of repeated to unrepeated words was not a factor in her Experiment 1, the term repetition blindness was being applied to a different effect in quite a different task. The use of the RB label for these different effects was not an issue for Kanwisher (1987) because of the general nature of her conceptualization of these deficits involving repetition. She hypothesized that the perception of any item requires two stages that operate on different sorts of mental representations (“types” and “tokens”). In the first stage, called AMERICAN JOURNAL OF PSYCHOLOGY Fall 2008, Vol. 121, No. 3, pp. 409–449 © 2008 by the Board of Trustees of the University of Illinois

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43

410 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43

brill et al.

type identification, sensory information must be matched with a type representation in memory. In the second stage, called token individuation, the type representation must be combined with episodic details (“tokens”). Kanwisher considered the repetition deficit effects in both experiments to be aspects of a single phenomenon explained by a failure of token individuation for the second repeated target in a rapid serial visual presentation sequence. However, the common RB label has the unfortunate tendency to obscure the fact that Kanwisher (1987) found different effects in different tasks. The cognitive processing of sentences differs greatly from the processing of semantically unrelated lists of items. The syntax and meaning inherent in sentences create expectations in the mind of the reader that have significant ramifications for how information is sampled and processed. In this article, we are concerned with the explanation for RB effects in the domain of tasks that involve judgments of a visually presented sequence of semantically unrelated items, as in Kanwisher’s (1987) Experiment 1. We will not consider the relationship of the RB effects in this domain to those in sentence processing tasks. Since Kanwisher’s (1987) first experiment, the kinds of semantically unrelated, sequentially presented visual materials found to produce repetition deficits include orthographically similar words (Chialant & Caramazza, 1997; Harris & Morris, 2000; Kanwisher & Potter, 1990), orthographically dissimilar but phonologically similar words (Bavelier & Potter, 1992), letters (Anderson & Neill, 2002; Armstrong & Mewhort, 1995; Bavelier & Potter, 1992; Chun & Cavanagh, 1997; Kanwisher, Kim, & Wickens, 1996; Luo & Caramazza, 1996; Neill, Neely, Hutchinson, Kahan, & VerWys, 2002; Park & Kanwisher, 1994), inverted letters (Kuwana, 2004), digits (Bavelier & Potter, 1992), pictures (Bavelier, 1994), sublexical units (Yeh & Li, 2004), colors (Kanwisher, 1991; Kanwisher, Driver, & Machado, 1995), simple familiar shapes (Kanwisher, 1991), novel shapes (Arnell & Jolicoeur, 1997), novel objects (Coltheart, Mondy, & Coltheart, 2005), locations (Epstein & Kanwisher, 1999), and nonwords (Harris & Morris, 2004; but see Coltheart & Langdon, 2003, and Campbell, Fugelsang, & Hernberg, 2002, for an opposing view on nonwords). The domain of tasks that involve a visually presented sequence of unrelated items can also be divided according to the way repetition must be identified and reported. In some tasks, observers not only must perceive the items in a sequence but also retain them in memory and report them. We shall call these identification tasks, because participants report either the identity of the repeated item or the identities of all the items in the sequence. Kanwisher’s (1987) first experiment used an identification task, as did other early studies of RB. The involvement of memory encoding and retrieval, even for the brief periods entailed by these experiments,

repetition blindness

411

raised the possibility that RB findings were caused by memory effects and guessing biases. For example, Armstrong and Mewhort (1995) replicated the retrieval deficit associated with a repeated item relative to an unrepeated item when participants were asked to provide a full report of all the items presented but showed that the deficit disappeared when participants were asked to produce each critical item after being cued with the preceding item in the sequence. Similarly, Fagot and Pashler (1995), Whittlesea, Dorken, and Podrouzek (1995), and Whittlesea and Podrouzek (1995) altered the reporting requirements in their experiments and eliminated the RB effect. These researchers concluded that RB is the result of processing that takes place after the stimulus items are already encoded. Some of them likened the RB results to a previously demonstrated memory phenomenon known as the Ranschburg effect (Crowder & Melton, 1965; Jahnke, 1969, 1972), in which retrieval deficits are found for repeated items presented at rates slow enough to ensure sufficient time for perception and encoding of each item. Although these studies by researchers favoring an interpretation of RB based on postperceptual processes were criticized on methodological grounds (Downing & Kanwisher, 1995), at a minimum they raised the possibility that all RB to that point was the result of memory and response effects. Consequently, Park and Kanwisher (1994) introduced a new task that reduced memory requirements. Their seventh experiment required an observer to report whether a categorically defined target (in this case, vowels among consonant distractors) occurred once or twice in a sequence. We will refer to this type of task as a counting task. When a vowel occurred twice, its two occurrences could either be identical (e.g., AA) or different (e.g., AO). In this study, RB was defined as poorer detection of same (e.g., AA) than of different (e.g., AO) pairs. Kanwisher et al. (1996)1 refined this procedure by presenting the possible targets for a block of trials in advance of each block. Park and Kanwisher (1994) and Kanwisher et al. (1996) found that two occurrences of a vowel were less likely to be detected when the same vowel was repeated (e.g., AA) than when two different vowels each appeared once (e.g., AO). They also found a lag effect when the observer did not know which vowels might be presented (Park & Kanwisher, 1994, Experiment 4; Kanwisher et al., 1996, Experiment 1). When the two targets were enumerated before trials (Kanwisher et al., 1996, Experiments 1 and 2), RB was also found at a lag of four intervening letters. That is, a lag effect was not found. The study of Kanwisher et al. (1996) is crucial to elucidating the RB effect because the counting task it used was developed in response to criticisms of Kanwisher’s (1987) original explanation of RB. Other researchers (Anderson & Neill, 2002; Coltheart & Langdon, 2003; Hochhaus & Johnston, 1996; Johnston, Hochhaus, & Ruthruff, 2002; Kuwana, 2004;

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43

412 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43

brill et al.

Morris & Harris, 2004; Neill et al., 2002) have also recently demonstrated RB effects using counting tasks. In an experimental refinement that further reduced memory demands, Johnston et al. (2002) and Morris and Harris (2004) showed RB effects even when their participants’ task was to make a response (key press) immediately upon the detection of the targets. Johnston et al. (2002) suggested that their immediate response method leads to even stronger conclusions than the counting task, but we agree with Kanwisher et al. (1996) that it is implausible that an observer would perceive a repeated letter but forget it before realizing that it is one of the targets. For convenience, then, the counting task category will be considered to include the immediate response method as well. This brings us to the purpose of this study. First we replicated and extended the crucial study of Kanwisher et al. (1996) in order to evaluate the extent to which the perceptual and memory explanations could account for the observed RB effects. This effort comprised Experiments 1, 2, and 3 of the study. Next, in the fourth experiment we attempted to extend the RB effect found for individual letters to three-letter words. In Experiments 5 and 6 we replicated and extended the seminal first experiment of Kanwisher (1987) in order to analyze the cause of the RB effect in that study. Finally, we reflect on the best explanations of the RB effects in the tasks studied here, and we consider the range of tasks to which these explanations can be plausibly applied. EXPERIMENT 1 As shown in Table 1, Experiments 1, 2, and 3 were nearly identical in design to Experiments 1 and 2 of Kanwisher et al. (1996). Kanwisher’s experiments were the first among the aforementioned experiments that provided strong evidence for a perceptual locus for RB by minimizing factors that had previously suggested a memory locus. The experiments used a counting task: The participant was to report whether one or two vowels appeared in a letter sequence otherwise consisting of consonants. First, because an observer had to remember only briefly whether one or two vowels had been seen, postperceptual retention of the correct response was kept to a minimum. Second, in Kanwisher et al.’s Experiment 2, the presentation duration of the first occurrence of the target was 240 ms, twice as long as the durations of the other letters in the sequence, in order to ensure that it was encoded. Therefore, an RB effect would have to be the result of a failure to detect or report the second occurrence of the target, not the first occurrence. Third, the experiment was cleverly designed so that the one-vowel letter sequences could be treated as a noise baseline for which reports of a repeated target were false alarms

5, 2 1, 4 A, E, I, O, U 120 ms Upper

5 1, 4 A, E, I, O, U 240 ms Upper

Kanwisher et al. (1996), Expt. 1 Kanwisher et al. (1996), Expt. 2

Target uncertainty Lag Targets Duration of Critical Item 1 Case of Critical Item 2



Table 1. Comparison of experimental designs 2 1, 4 A, E, I, O, U 240 ms Upper

Expt. 1

2 1, 4 A, E, O, U 240 ms Lower

Expt. 2

1 1, 4 A, E, O, U 120 ms Lower

Expt. 3

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43

414 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43

brill et al.

and two-vowel sequences could be treated as signal trials. Thus, a signal detection paradigm could be applied, and d rather than mere response rates could be computed. The large ds found in this task demonstrated an RB effect even under high levels of accuracy, as the ds for identical targets (e.g., AA) were consistently lower (though still high) than for nonidentical targets (e.g., AO). Nevertheless, there remained ways in which nonperceptual response factors could influence the results of the task. In Kanwisher et al.’s (1996) first experiment, two levels of target uncertainty were used. Recall that Park and Kanwisher (1994) used a counting task that involved what we will call a five-alternative uncertainty condition: Participants were simply told that any two vowels could appear in the sequence and that they could be the same or different. In their first experiment, Kanwisher et al. (1996) compared the five-alternative uncertainty condition to a two-alternative condition and found an RB effect in both. In the two-alternative uncertainty condition, participants were given a target set of two vowels for a block of 36 trials, and their task was to count the number of targets, 1 or 2, on each trial. For example, at the beginning of the AO block, the participant was told that each sequence over the next 36 trials would contain A, O, AA, OO, AO, or OA. The task was to discriminate the trials containing two vowels from those containing one vowel. In Experiment 2 of Kanwisher et al. (1996), only the five-alternative uncertainty condition, the higher level of task uncertainty from Experiment 1, was used. Relevant to this concern with target uncertainty is Experiment 2 reported by Whittlesea and Masson (2005). In a counting task for a repeated word, the least underreporting of the repeated word was found in a onealternative low–target uncertainty condition. That is, the underreporting of the repeated word was reduced when it was shown to participants before they viewed the test sequence. Thus, Whittlesea and Masson’s finding is consistent with the possibility that target uncertainty also played a role in Kanwisher et al.’s (1996) result. Also, because perceptual theories, including Kanwisher’s token individuation theory, assume that the first occurrence of the target, typically called Critical Item 1 (C1) in RB studies, affects the processing of the second occurrence, C2, they predict that placing a sufficient number of items between C1 and C2 should eliminate RB. However, in one condition of Kanwisher et al.’s (1996) first experiment and in their second experiment, significant RB was found when four items intervened between C1 and C2. The significant effect at a lag of four suggests that a nonperceptual factor not dependent on lag contributed to the RB effect. However, the effects of target uncertainty in Whittlesea and Masson’s (2005) study and of lag in Kanwisher et al.’s (1996) are only suggestive. The only way to determine whether target uncertainty was the cause of

repetition blindness

415

RB would be to replicate Kanwisher et al.’s study under conditions of reduced uncertainty. We chose Kanwisher et al.’s Experiment 2 because in this experiment C1 was presented for 240 ms, thus ensuring its detection. Experiment 1 reported here was identical to Kanwisher et al.’s Experiment 2 except for two changes introduced to test the role of target uncertainty in RB. First, a two-alternative, instead of a five-alternative, detection task was used. Second, we modified the experimental design to include, before every trial, instructions that explicitly showed the participants the first target and the two possible second targets. For all trials a participant always knew the identity of C1 in advance of the sequence. METHOD Procedure On every trial in the experiment participants viewed a sequence of uppercase white letters displayed on a black background. Each trial began when the participant pressed any key on the computer keyboard. The instructions for the trial were then displayed. After reading the trial instructions, the participant pressed any key to continue. A plus sign (+) appeared in the center of the monitor for 495 ms as a fixation point, followed by the sequence of letters. Each letter in the sequence, except for the first critical item (C1), was presented for 120 ms. C1 was shown for 240 ms to ensure that it was encoded. Thus, any RB found would be the result of the inability to detect the second rather than the first occurrence of the repeated item. At the end of the letter sequence a number (#) sign was shown for 225 ms as a mask. After 250 ms, a prompt appeared on the screen asking “One or Two?” The participants responded by typing a 1 or 2 on the keyboard. A new prompt asking, “How confident are you?” appeared on the screen 495 ms after the response. Participants then clicked on one of three button choices that were labeled “Sure,” “Probably correct,” and “Guessing.” At this point a participant could initiate a new trial, except that at the end of each block of 36 trials, participants were first given feedback on the percentage of correct responses they had obtained on the detection task on that block.

Design The factors defining the design of the experiment are shown in Table 1. The study used a counting task in which a participant had to determine whether one or two vowels appeared in a sequence of 11 letters. The letter sequence always contained 9 or 10 different consonants. The consonant sequence for each trial was created by random sampling without replacement from the set of all consonants except F, Q, and V (because these letters too closely resemble E, O, and U, respectively). The task instructions for participants in this experiment were as follows: Determine if there were one or two target letters in the lists you will see. Your task is simply to decide for each list whether there were one or two target

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43

416 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43

brill et al.

letters somewhere in that list. Before seeing each list, you will be prompted with the possible targets that you will be asked to detect on that trial. When you are certain that you have learned the targets, press any key. The targets will then disappear and the list will be shown. There were 360 trials in the experiment. Participants saw a prompt before each trial that identified the C1 vowel along with the two possible C2 vowels. For example, the prompt when the target pair consisted of the letters O and U was, “On the next trial, the list you will see will be all consonants, except for one or two vowels. The list will contain an O, followed somewhere in the list by either no other vowel, a U or another O.” In a 360-trial experiment, for 120 trials two different vowel targets occurred in the test sequence, for 120 trials the same target vowel was repeated in the test sequence, and for 120 trials a single target occurred in the test sequence. As shown in the second row of Table 1, the 120 trials in which the test sequence contained two different vowels consisted of 60 trials in which one consonant separated the two occurrences of the target (lag one) and 60 trials in which four consonants separated the target (lag four). For the 60 trials with the same lag, each of the 20 possible pairs of target vowels occurred three times, once each with C1 in positions 3, 4, and 5. The sequences with repeated targets were created from the 120 different-target sequences by simply replacing the vowel in C2 with a second occurrence of the vowel in C1. The sequences with a single target were created from the 120 repeatedtarget sequences by simply replacing the vowel in C2 with a consonant. The 360-trial experiment was partitioned into 10 blocks of 36 trials each. Each block had all the trials associated with each of the 10 possible target pairs. For example, the EO block contained the 12 EO and OE different-target trials, the corresponding 12 EE and OO same-target trials, and the 12 E and O single-target trials. The block order and the trial order within each block were randomized. A total of 18 participants, all undergraduate students at Rutgers University with normal or corrected-to-normal vision, participated in the experiment. All but one learned English by the age of 5, and the first language (Polish) of the nonnative English speaker uses the same alphabet as English. Participants either volunteered or received course credit for their participation in the experiment. The experiment was run on a standard personal computer, and E-Prime was used for programming and controlling the experiment.

RESULTS Four separate d calculations were done for each participant for each level of target similarity and lag. A hit was the reporting of a repeated target when in fact a target was repeated. A false alarm was the reporting of a repeated target when there was only one target in the sequence. Standard corrections were applied to false alarm rates of 0 and hit rates of 1. Specifically, a false alarm rate of 0 was converted to 1/2N, where N is the number of observations, and a hit rate of 1 was converted to 1 – 1/2N. Table 2 shows the hits and false alarms for each condition.

417

repetition blindness Table 2. Mean (SE ) hit and false alarm rates, Experiment 1 Lag 1   Same   Different Lag 4   Same   Different

Hits

False alarms

.66 (.046) .90 (.022)

.16 (.035)

.81 (.026) .90 (.020)

.16 (.034)

Notice that participants were asked to rate their confidence in their responses in this and all subsequent experiments. However, across all the experiments, the level of high-confidence responses was more than 70%. Because of the predominance of high-confidence responses, the results of this and the subsequent experiments were the same whether or not confidence was taken into account in the computation of d or the analysis was restricted to high-confidence responses. As a result, only the analyses on all the data are reported. Figure 1 shows the mean d values for each condition. A 2 (C2, same vs. different) × 2 (lag, one vs. four) repeated-measures analysis of variance (anova) was performed on the data. For p ≤ .004, there were significant main effects for target similarity, F(1, 17) = 31.8, MSE = 0.293, and lag, F(1, 17) = 11.1, MSE = 0.158. A significant interaction was found between target similarity and lag, F(1, 17) = 16.3, MSE = 0.060, p