Grouping of list items reflected in the timing of recall: implications for models of serial verbal memory

Transcription

1 Journal of Memory and Language 47 (2002) Journal of Memory and Language Grouping of list items reflected in the timing of recall: implications for models of serial verbal memory Murray T. Maybery, a, * Fabrice B.R. Parmentier, b,1 and Dylan M. Jones c a School of Psychology, University of Western Australia, Crawley, WA 6907, Australia b Cardiff University, Wales c Cardiff University, Wales, and University of Western Australia, Australia Received 21 January 2001; revision received 30 October 2001 Abstract Three experiments examined the effect of temporal grouping on the timing of recall in verbal serial memory. Compared to an ungrouped condition, recall in a grouped condition produced a peak in latency between the groups (Experiment 1). However, the ratio of within- to between-group intervals at presentation was not reflected in recall (Experiment 2), contrary to the predictions of some oscillator models (Brown, Preece, & Hulme, 2000; Burgess & Hitch, 1999). In Experiment 3, grouped and ungrouped lists of different lengths were compared to assess a recent version of the ACT-R model applied to serial recall (Anderson, Bothell, Lebiere, & Matessa, 1998). Recall latencies showed a cost at group onset related to group size and a cost for all items of the first group associated with carriage of a second group. Results are discussed with reference to oscillator models, the ACT-R model, and augmented versions of it. Ó 2002 Elsevier Science (USA). All rights reserved. Keywords: Short-term memory; Verbal memory; Serial recall; Recall timing; Temporal grouping; Oscillator models * Corresponding author. Fax: addresses: murray@psy.uwa.edu.au (M.T. Maybery), f.parmentier@plymouth.ac.uk (F.B.R. Parmentier), jonesdm@cardiff.ac.uk (D.M. Jones). 1 Fabrice Parmentier is now at the Department of Psychology, University of Plymouth. Reproduction, over the short term, of the order of verbal sequences has been a preoccupation of both classic (Ebbinghaus, 1964) and contemporary research (see, e.g., Baddeley, 1986; Brown et al., 2000). Generally, this work has been subsumed under the rubric of Ôworking memory,õ a component of cognitive architecture embodied universally in general theories of human performance. The current paper explores the structure of working memory by measuring the timing of response production in a serial recall task. Response timing has been used only rarely in studies of serial recall the measurement of errors has been the dominant method but recently the measurement of timing has become more relevant to theory testing with the emergence of models that embody elements of temporal representation in the process of retrieval. Specifically, several recent theories suggest that the timing of the presented sequence should be echoed in the timing of the recall sequence. This paper reports experiments that manipulate the temporal grouping of auditorily presented lists and assess how this is reflected in the timing of responses X/02/$ - see front matter Ó 2002 Elsevier Science (USA). All rights reserved. PII: S X(02)

2 M.T. Maybery et al. / Journal of Memory and Language 47 (2002) In studies of temporal grouping, series of items, such as words, are presented for retention and their timing is manipulated to form distinct sub-series (e.g., ABC DEF GHI versus ABCDEFGHI). This form of grouping has been found to improve the accuracy of serial recall (e.g., Frankish, 1985, 1989), to create effects of primacy and recency within the sub-series (Hitch, Burgess, Towse, & Culpin, 1996; Ng & Maybery, 2002; Ryan, 1969a, 1969b), and to reduce the probability of items migrating to adjacent positions if migration involves crossing from one sub-series to another (Henson, 1998; Ng and Maybery). Several models of serial memory can account for the above data (Brown et al., 2000; Burgess & Hitch, 1996, 1999; Henson, 1998; Henson & Burgess, 1997). However, these data are almost exclusively error scores. At a general level, current models of shortterm memory account for the retention of serial order in radically different ways. Order is presumed to be retained: (i) by encoding associations between items, as with TODAM and extensions to it (Lewandowsky & Murdock, 1989; Murdock, 1995), (ii) by encoding successive items with differential strength, as with the primacy model (Page & Norris, 1998a), (iii) by associating items with a context signal, as with oscillator models (Brown et al., 2000; Burgess & Hitch, 1996, 1999; Henson & Burgess, 1997), or (iv) by encoding items within declarative knowledge units, as with the Adaptive Character of Thought Revised (ACT-R) production system model (Anderson et al., 1998; Anderson & Matessa, 1997). Recent reviews have identified several phenomena that are problematic for the first two classes of model (Brown et al., 2000; Burgess & Hitch, 1999; Henson, 1998, 1999), among them temporal grouping. Our focus, therefore, is on the context models specifically the oscillator models of Burgess and Hitch and Brown et al. and declarative knowledge models specifically the ACT-R model of Anderson and his colleagues. Contemporary context models assume that temporal oscillators provide the context that serves as the basis for retrieval. Temporal oscillators are neural cell assemblies that fire in regular cycles, but with different oscillators cycling at different frequencies. Oscillators therefore provide a rich signal that changes continually over time. Two recent models embodying oscillators are capable of making predictions about the effect of grouping on the timing of responses. Following a brief description of each of these models, we consider the alternative perspective involving declarative knowledge provided by the ACT-R model. The phonological loop model Burgess and Hitch (1996, 1999) and Hitch et al. (1996) include temporal oscillators as one of several features in a computational model which elaborates the phonological loop (PL) construct (Baddeley, 1986). In this model the current signal from a set of oscillators is associated with each item as the to-be-remembered list is presented. Then, at recall, the oscillators are reset to the points they had occupied in their cycles at the commencement of list presentation. As the oscillators are replayed, their changing signal then provides cues for the serial recall of the list items. To illustrate, assume that successive items i and i þ 1 had been presented at times t i and t iþ1 after the onset of list presentation. When the oscillators are reset at recall and allowed to evolve, after time t i has elapsed, the signal they provide will closely resemble the signal that had been associated with item i during list presentation. At this point in recall, item i should be cued more strongly than the other list items. As the oscillators continue to evolve to time t iþ1, they assume a state that is optimal for cueing the recall of item i þ 1. Serial recall therefore proceeds under the direction of the signal provided by the oscillators. According to the PL model, the oscillators provide a more sophisticated signal when lists are presented in temporal groups. If there is a constant interval from onset of one group to onset of the next, a group of oscillators becomes entrained to this frequency. The signal of these oscillators repeats itself in phase with the groups. This group of oscillators effectively encodes within-group position. Consequently, items within a group have highly differentiated contextual codes, but items occupying the same position in different groups have very similar codes. However, only some of the available oscillators become entrained to the inter-group frequency. The remaining oscillators continue to operate with varying frequencies, and are used to encode the listwise positions of items, as described in the preceding paragraph. At recall, all the oscillators are reset to the points they had occupied in their cycles at the start of the list, and their composite signal evolves to provide the cues for serial recall. Because the oscillators evolve at recall in the same way as they had done during list

3 362 M.T. Maybery et al. / Journal of Memory and Language 47 (2002) presentation, the expectation is that the timing of recall should mimic the timing of list presentation. OSCAR The OSCillator-based Associative Recall (OS- CAR) model of Brown et al. (2000) also adopts temporal oscillators as providing a context signal that is associated with list items at encoding. As for the PL model, at recall the oscillators are reset to the states they had occupied at the start of the list, and their unfolding activation cues serial recall of the list items. When lists are presented with temporal grouping, OSCAR also assumes the involvement of oscillators with a cycle synchronized to group onset. However one feature of OSCAR differentiating it from the PL model is that the overall rate of evolution of the oscillators is presumed to be under strategic control. Therefore, a different rate can be adopted at recall to what had pertained at encoding. This means that recall need not preserve the absolute time intervals that separated items during list presentation. However, recall should preserve relative time intervals, that is, the intervals between items at recall should be scaled by a constant in relation to the intervals present at encoding. Despite the dependence of the PL and OSCAR models on time, it is surprising that predictions concerning the timing of recall have not been evaluated. Nevertheless, the models have been shown to account for differences in the serial position function on accuracy and in the pattern of order errors when lists with and without temporal grouping are compared (for reviews, see Brown et al., 2000; Burgess & Hitch, 1999; Henson, 1998, 1999). (Note that we consider another oscillator model described by Henson & Burgess, 1997, in General discussion. This model uses oscillators more flexibly than do the PL and OSCAR models, but as a consequence does not generate precise predictions concerning response timing.) ACT-R Ironically, it is a model not based on temporal oscillators that has been used to generate predictions on the timing of recall for grouped lists. According to the ACT-R model (Anderson et al., 1998; Anderson & Matessa, 1997), lists of memory items are organized hierarchically in groups, and within these groups, in items. The hierarchical structure is encoded as declarative knowledge and the output of items at recall is achieved by executing a set of production rules of the form IF- THEN. When a list is presented in groups (ABC DEF GHI), chunks (or knowledge units) for the groups are activated. These knowledge units have links with elements such as ÔList,Õ ÔGroup1,Õ and ÔSize3Õ. The group chunks are themselves linked to chunks representing the items, so ÔGroup1Õ would be linked with the three chunks representing the first three items. Each of the item chunks is associated with elements coding for the position of the item in the group and its identity (e.g., for the first item, Ô1stÕ and ÔAÕ). The representation of a list can therefore be thought of as a tree where the list unit connects to units representing the groups, which themselves connect to units representing the items. This hierarchical structure has some similarity to the Perturbation Model of Estes (1972). For the ACT-R model, the time course of response production is determined by constraints of the architecture on knowledge retrieval. Its predictions derive from the specification of a set of production rules that can be used to navigate through the hierarchical structure and retrieve knowledge. Errors arise from noise affecting the link between a production rule and an element, which makes possible errors commonly found in serial memory such as the migration of an item to a serial position close to its correct position or the migration of an item from one group to another, but preserving correct serial position relative to the group boundaries. Predictions in terms of response latencies are directly related to mechanisms at play at retrieval. The steps described below correspond to the mechanisms for forward serial recall described by Anderson et al. (1998). To recall a list in order, first, production rules retrieve the knowledge units corresponding to the groups comprising the list. Next, another production rule sets the goal to retrieve the contents of the groups in order. For each group, knowledge units for the items belonging to the group must be retrieved before a sub-goal is set to output the group items in order. A final production rule outputs an item before switching to the next. The time taken in each phase of recall is a direct function of the number of knowledge units retrieved. Response time (RT) is typically maximal when recalling the first item of the list because, prior to outputting it, the system must first retrieve the knowledge units for all groups and then the knowledge units for all items of the first group. In comparison, RT is somewhat shorter when

4 M.T. Maybery et al. / Journal of Memory and Language 47 (2002) recalling the first item of a subsequent group, prior to which only information about the constituent items of its group needs to be retrieved. Once the first item of a group has been reported, the time to output the remaining group items is minimal because these have been retrieved at the start of the group and so can be reported relatively rapidly. In summary, one can identify two predictors of RT for the forward serial recall of grouped lists: the number of groups in the list (affecting the time to report the first item of the first group), and the number of items in each group (affecting the time to report the first item of each group). The three experiments reported below systematically examine the effect of temporal grouping on the timing of serial verbal recall. Experiments 1 and 2 address the predictions of the PL and OS- CAR models by comparing RTs for grouped and ungrouped lists. In prospect, we show that these models cannot account for the RT data of Experiment 2 in which we manipulated the temporal compression of the groups. The ACT-R model sits more comfortably with the results of the first two experiments, so in Experiment 3 we submit ACT- R to further test by collecting RTs for grouped and ungrouped sequences of varying length. Reference is also made to the work of Cowan and his collaborators (Cowan, 1992; Cowan et al., 1994, 1998) concerning a memory-search process purported to operate during serial recall. In a section prior to General discussion, we report fits to our data using regression models derived from ACT- R and extensions to it. Experiment 1 In this experiment, we examined the effect of temporal grouping on RTs for the serial recall of auditory letter sequences. Sequences were six items in length, and we compared performance from two conditions. In the ungrouped condition, all sequences were presented with an item-to-item stimulus onset asynchrony (SOA) of 1000 ms. In the grouped condition, SOA was 750 ms for the transitions between items 1 3 and also between items 4 6, whereas the SOA was 2000 ms for the transition from item 3 to item 4. This timing therefore provided two temporal groups, each of three items. Only a few studies have reported serial position effects for RT in the context of a verbal serial recall task, and in each case analysis was limited. First, Cowan et al. (1998, Appendix A), reported data on the silent intervals separating oral responses for sequences of length 3 and 4 in a word span task. For each list length, the interval between the last two items was shorter than the interval(s) earlier in the response sequence, an effect the authors interpreted with reference to the marking of earlier responses. This general idea that early items, once recalled, are suppressed and this then facilitates responses to later items is shared with several contemporary models of short-term memory (see Lewandowsky, 1999). Second, Kahana and Jacobs (2000) reported RTs as a function of serial position for highly practiced participants on over-learned item lists of consonants. Although not analysed systematically, peaks in RT were interpreted as reflecting the onsets of spontaneously organized groups. Third, Anderson and Matessa (1997) reported a pilot study in support of the ACT-R model. Adults were presented with sequences of 5, 7, or 9 digits grouped following 3 2, 3 4, or structures, respectively. Grouping was induced by presenting the sub-series in different screen locations, that is, sub-series were not differentiated temporally. RTs were elevated for the first item of the list, and for the first item of each group. Also, at serial position 4, the first item of the second group for all lists, RTs were longer for larger lists. Finally, Anderson et al. (1998) reported similar data for 3 12 item lists, with forward and backward recall. The studies of Anderson et al. (1998) and Anderson and Matessa (1997) are difficult to compare with other studies on grouping in serial memory, however. Indeed, they used a visuo-spatial rather than temporal manipulation to induce grouping. Further, no systematic comparisons with ungrouped lists were reported, nor was information regarding errors reported. Our experiments investigate the effect of temporal grouping on RTs in serial recall. The focus on temporal grouping is critical because the RT data can then be used to evaluate oscillator models of serial memory, as well as the ACT-R model. The aim of the initial experiment was simply to demonstrate for the first time that temporal grouping affects the timing of serial recall. For the grouped condition, all models (PL, OSCAR, and ACT-R) predict an elevated RT for the fourth item relative to items 2, 3, 5, and 6. For the PL and OSCAR models, this is because the timing of recall should mimic the timing of presentation and at presentation the fourth item is preceded by an extended pause. For the ACT-R model, an

5 364 M.T. Maybery et al. / Journal of Memory and Language 47 (2002) elevated RT is expected for the fourth item because extra processes are assumed in accessing the items comprising the second group. Method Participants Thirty-two University of Western Australia undergraduates aged years participated as part of a first-year psychology course. Apparatus Equipment consisted of a 400 MHz Intel PC fitted with a Soundblaster16 soundcard, a 35 cm Microtouch touch-sensitive screen, and Arista FS300 headphones. Stimuli The stimulus ensemble comprised the 20 consonants other than the trisyllabic W. Each consonant was recorded in English using an adult male speaker. The sound files were recorded in 16- bit mono format using a 44.1 khz sampling rate, and were normalized and made to fit a 450 ms envelope. Two sets of memory sequences (A and B) were constructed. In each set, there were 24 sequences (4 practice, 20 test), each comprising six consonants. The 20 test sequences were constructed by selecting consonants at random, but with two constraints: (1) across the 20 sequences, each consonant appeared exactly once in each serial position; and (2) no pair of consonants occurred in the same order in consecutive trials. Procedure Sixteen participants received ungrouped sequences and 16 received grouped sequences, and in each case stimulus sets A and B were used equally often. A short preliminary task provided familiarity with the auditory stimuli and touch screen. On each of 60 trials, a consonant was presented through the headphones and the required response was to touch that consonant on a letter board displayed on the screen. The letter board consisted of a 5 4 matrix of the consonants arranged alphabetically, with each consonant printed inside a 37 mm square. For the memory task, the participant initiated each trial by touching a start button on the screen. The screen then cleared, and the sequence of six consonants was presented over the headphones. Then the letter board appeared, with an additional box, labeled MISS, appended to the bottom-right corner. Six empty boxes were also presented along the bottom of the screen. As the participant touched consonants (or MISS) on the letter board, the six empty boxes filled from left to right with these responses. An RT (with 6ms accuracy) was recorded for each response (i.e., from presentation of the letter board to the first response, and then the intervals between responses, where each response was recorded at touch onset). Responses could not be corrected, and the final response caused the screen to clear and the start button to appear for the next trial. Instructions emphasized correct serial recall and also stressed that responses were to be made using the index finger of the preferred hand, which, at the start of each trial, was to be positioned on a dot centered on the bottom frame of the monitor. The critical manipulation was the timing of sequence presentation. For ungrouped sequences, each 450 ms consonant was followed by 550 ms of silence (i.e., SOA ¼ 1000 ms). For the grouped sequences, the silent intervals separating consonants were 300, 300, 1550, 300, and 300 ms, respectively (i.e., within-group SOAs ¼ 750 ms; between-group SOA ¼ 2000 ms). In both conditions the total duration of the sequence was 5450 ms. Also, 1000 ms separated the touch of the start button and presentation of the first consonant, and 550 ms separated the end of the last consonant and appearance of the letter board. Results The RT data are of critical interest, but first we present analyses of accuracy and error types to show that effects typically reported for temporal grouping are also present in our data. Accuracy Recall was more accurate in the grouped condition, but this superiority was less evident in the first and last serial position due to the more pronounced bowing of the serial position curve in the ungrouped condition. A 2 (condition: grouped versus ungrouped) 6 (serial position) analysis of variance (ANOVA) was conducted on the number of items recalled in correct serial position, summed across the 20 test trials. All effects were significant: F ð1; 30Þ ¼16:22, MSE ¼ 22.73, p <:001, for the main effect of condition; F ð5; 150Þ ¼26:21, MSE ¼ 3.42, p <:001, for the main effect of serial position; and F ð5; 150Þ ¼ 5:60, MSE ¼ 3.42, p <:001, for the interaction. The interaction

6 M.T. Maybery et al. / Journal of Memory and Language 47 (2002) (see Fig. 1) shows the typical primacy and recency components for the ungrouped condition, whereas, consistent with other studies (e.g., Hitch et al., 1996; Ng & Maybery, 2002), grouping elevated accuracy for intermediate serial positions. This observation was confirmed by contrasting the two conditions with reference to the quadratic trend over serial position, F ð1; 30Þ ¼9:72, MSE ¼ 3.44, p <:01. Order errors The migration of items away from their correct serial positions was constrained by grouping, just as has been found in previous studies. For ungrouped lists, errors frequently involve the migration of items to neighboring serial positions (see Brown et al., 2000). These errors include reporting item 1 at serial position 2 or vice versa (we collapse these two errors together, and symbolize them as 12), reporting item 2 at serial position 3 or vice versa (23), and so on (34, 45, and 56). Furthermore, migrations 23, 34, and 45 typically occur more frequently than migrations 12 and 56. However, for grouped lists, there tend to be relatively few adjacent migrations across the group boundary, that is, for a 3 3 grouping structure, item 3 rarely gets reported at serial position 4, or vice versa (see, e.g., Ng & Maybery, 2002). To see if these patterns were present in our data, we calculated for each participant the mean number of errors per case for the eight cases represented by 12, 23, 45, and 56, and also the mean number of errors per case for the two cases represented by 34. A 2 (condition: grouped versus ungrouped) 2 (error type: 12, 23, 45, 56, versus 34) ANOVA was conducted on these means. The effect of error type was not significant, F ð1; 30Þ < 1, MSE ¼.43, p >:1. However, the effect of condition was significant, F ð1; 30Þ ¼16:25, MSE ¼.84, p <:001, as was the interaction, F ð1; 30Þ ¼4:81, MSE ¼.43, p <:05. Fig. 2 shows that both types of error occurred more frequently for the ungrouped sequences compared to the grouped sequences. However for the ungrouped sequences, error 34 occurred more frequently than errors 12, 23, 45, and 56, whereas the reverse was the case for the grouped sequences. Therefore, consistent with other studies, temporal grouping led to relatively fewer migrations across the group boundary. (The two following experiments also show the pattern of migrations expected from grouping. However, in the interests of economy of exposition, these analyses are reported in the Appendix.) Response times Response timing showed effects of grouping with some slowing of responses for ungrouped Fig. 1. Experiment 1: Number of correct responses (maximum 20; 95% confidence interval for withinsubjects effects ¼:65). Fig. 2. Experiment 1: Mean number of errors calculated separately for error types 34 and 12, 23, 45, 56 (95% confidence interval for within-subjects effects ¼:24).

7 366 M.T. Maybery et al. / Journal of Memory and Language 47 (2002) relative to grouped presentation and a delay at onset of the second group in addition to a general delay at list onset. For each participant, a median RT was calculated for each serial position, across the 20 test trials, but restricted to items recalled correctly. These medians were then submitted to a 2 (condition: grouped versus ungrouped) 6 (serial position) ANOVA. The effect of serial position was significant, F ð5; 150Þ ¼ 62:62, MSE ¼ , p <:001, as also was the interaction, F ð5; 150Þ ¼ 3:54, MSE ¼ , p <:01. However, the main effect of condition was marginal, F ð1; 30Þ ¼ 4:02, MSE ¼ , p ¼ :054. The interaction (see Fig. 3) shows that RTs tended to be shorter for grouped compared to ungrouped lists at all serial positions except the fourth, where the difference was reversed. There is a clear elevation in RT at serial position 4 for the grouped lists but not for the ungrouped lists. A contrast of serial position 4 with serial positions 3 and 5 was significant for the grouped condition, F ð1; 30Þ ¼ 24:02, MSE ¼ , p <:001, but not the ungrouped condition, F ð1; 30Þ ¼2:08, MSE ¼ , p >:1. For the grouped condition, serial position 4 represents the first item of the second group, so there is a substantial cost in latency associated with onset of recall of a new group. Finally, simple effects analyses showed that the difference in means for the grouped and ungrouped lists was marginal for serial position 1, t(30) ¼ 1.99, p ¼ :056, but achieved the conventional level of significance for serial positions 3 and 5, with t(30) ¼ 2.12, p <:05, and t(30) ¼ 2.62, p <:05, respectively. Interestingly, the RTs for the ungrouped condition appear to be somewhat shorter for serial positions 2 and 6 compared to serial positions 3 5, a pattern repeated in the following experiments. Discussion The results of Experiment 1 are consistent with previous studies on serial memory: Temporal grouping improves accuracy and causes a systematic shift in the nature of errors in a serial recall task. 2 The results also include a new finding: Temporal grouping affects RTs, as indicated by the presence of an elevated RT at serial position 4 in the grouped condition (i.e., at the first item of the second group). Grouping items temporally at presentation leads participants to group their responses accordingly. Anderson and Matessa (1997) and Anderson et al. (1998) reported an increase in RT at the first serial position of each group when grouping was induced visuo-spatially. The similarity of their effects to those of temporal grouping found in Experiment 1 points to the generality of the effects across modalities. By including a control condition in the present experiment, we are able to report the precise shift in the serial position function on RT that results from imposition of a temporal group structure. This information also permits a more direct evaluation of the oscillator and ACT-R models. The detailed pattern of RTs is broadly consistent with the ACT-R and oscillator models, however there are some deviations from expectations. In relation to ACT-R, as expected, peaks in RT were observed for the first item of the list in each condition and for the fourth list item in the grouped condition. Also, for the grouped Fig. 3. Experiment 1: Response times for correct serial recall (95% confidence interval for within-subjects effects ¼108). 2 Note that Anderson et al. (1998) did not report differences in accuracy between their ungrouped and grouped conditions, but their ungrouped condition consisted of the simultaneous visual presentation of lists of digits, which including repetitions. This method may have encouraged spontaneous grouping. The advantage in accuracy for temporally grouped lists over regular lists has been reported in many other studies, however (e.g., Frankish, 1985; Hitch et al., 1996; Ng & Maybery, 2002).

8 M.T. Maybery et al. / Journal of Memory and Language 47 (2002) condition, the peak in RT at the first list item exceeded the peak at the fourth item. These results are consistent with the retrieval of knowledge units for groups at the beginning of the list and knowledge units for items at the beginning of each group. However if each list in the ungrouped condition is encoded as a single group, then according to ACT-R, RTs for items 2 6 in this condition should be comparable to the RTs for items 2, 3, 5, and 6 in the grouped condition. In all these cases, the only processing expected is the enactment of a production that simply outputs the next item. However, at serial positions 3 and 5, RTs are significantly longer for the ungrouped condition compared to the grouped condition, and nonsignificant differences in the same direction are present at serial positions 2 and 6. At serial position 4, the RT for the ungrouped condition approaches the peak at that serial position for the grouped condition. Another feature is that RTs for the ungrouped condition are elevated for serial positions 3 5 compared to serial positions 2 and 6 (Fig. 3). One possible explanation for this pattern is that participants spontaneously group the items into either twos or threes. According to ACT-R, grouping in twos should lead to elevated RTs at serial positions 3 and 5, whereas grouping in threes should lead to an elevated RT at serial position 4. A mix of these two grouping strategies could therefore contribute to the particular form of the serial position function on RT observed for the ungrouped condition, that is, elevated RTs for serial positions 3 5 compared to serial positions 2 and 6. In regression modeling reported later, we successfully fit the data from ungrouped lists by adopting the assumption of spontaneous grouping in twos or threes. Grouping was induced temporally to evaluate the oscillator models. The specification of the retrieval mechanisms for these models implies effects of grouping on the timing of recall. Indeed, the rate of retrieval should match the rate of list presentation (for the PL model) or be scaled by a constant with reference to it (for the OSCAR model). The RT functions in Fig. 3 are broadly consistent with the two models: (1) There is a peak in the RT function at serial position 4 for the grouped condition, consistent with the longer SOA at presentation for that item (2000 ms) compared to the other items (750 ms); and (2) the RTs for the other items (items 2, 3, 5, and 6) are lower for the grouped compared to the ungrouped condition, consistent with the difference in SOAs (750 ms versus 1000 ms). Experiment 1 complements other research in confirming that the study of response timing can add significantly to our understanding of serial recall processes. Elsewhere it has been shown for vocal recall that children and adults produce longer silent intervals between consecutive items as list length increases (Cowan, 1992; Cowan et al., 1994, 1998). Cowan et al. (1994) also showed that the preparatory interval the time preceding the output of the first item of the sequence also increases as a function of list length. Several papers have also reported that individual and developmental differences in recall accuracy can be predicted by indices derived from the timing of recall (Cowan, 1992; Cowan et al., 1994, 1998; Chuah & Maybery, 1999; Dosher & Ma, 1998; Hulme, Newton, Cowan, Stuart, & Brown, 1999). According to Cowan (1992, 1999), these results suggest that a search process is operating during the preparatory interval and in each pause between the recall of items in the sequence. This process is thought to identity and reactivate memory items by recirculating them through the focus of attention (Cowan, 1999). The RT data of Experiment 1 are broadly consistent with the account of serial recall advanced by Cowan (1992, 1999; Cowan et al., 1994, 1998) if an additional assumption is made in extending this account to grouping. Cowan argued that the search process operating in the preparatory and inter-response intervals takes longer to complete for longer sequences. If the assumption is made that this process operates within each group in the grouped condition, but across the full list in the ungrouped condition, then our RT data fit Cowan s account: The trend is for RTs for ungrouped lists (searching 6 items) to be longer than RTs for grouped lists (searching 3 items). This holds for the preparatory interval and all subsequent intervals except at serial position 4, where presumably other processes could intervene for the grouped condition (e.g., processes associated with bringing the second group into the focus of attention). However, this observation in relation to CowanÕs (1992, 1999) account is secondary to our immediate aim, which is to evaluate predictions derived from the oscillator and ACT-R models. In order to test more rigorously the oscillator-based accounts of RT functions resulting from temporal grouping, Experiment 2 manipulated the ratio of within- to between-group SOA.

9 368 M.T. Maybery et al. / Journal of Memory and Language 47 (2002) Experiment 2 In this experiment, the PL and OSCAR models were put to a more stringent test by varying the ratio of the interval separating successive items within each group to the interval separating the last item of one group and the first item of the next. That is, Experiment 2 seeks to establish if there is any correspondence, whether absolute or relative, between input and output timing. In a ÔtightÕ condition, the within-group SOAs (i.e., for transitions 1 2, 2 3, 4 5, and 5 6) were 600 ms whereas the between-group SOA (i.e., for transition 3 4) was 3000 ms. In a ÔlooseÕ condition, the within-group SOAs were 900 ms and the betweengroup SOA was 1800 ms. The ratio of within- to between-group SOA was therefore 1:5 in the tight condition and 1:2 in the loose condition. Given the nature of the retrieval mechanisms at play in the PL and OSCAR models, RTs should differ for the tight and loose conditions. In particular, the different ratios of within- to between-group SOA should be reflected in the RT profiles for these two grouped conditions: Compared to the loose condition, the tight condition should show a much more pronounced elevation in RT at serial position 4 relative to the neighboring serial positions. This difference in the within- to between-group ratio of RTs for the two grouped conditions is expected even if retrieval is slowed or speeded overall compared to the actual timing of presentation (a variation possible in OSCAR). With reference to the ACT-R model, no differences in RT are expected for the tight and loose conditions provided the SOA ratio in the loose condition is sufficient to induce grouping (a reasonable assumption given that marked grouping was reported by Ng & Maybery, 2002, for the same 1:2 ratio of within- to between-group SOA). The timing of recall in ACT-R depends only on the grouping structure (3 3 for both loose and tight), so overlaid RT functions are expected for the two grouped conditions. However, these functions should differ from the function for a third, ungrouped condition, in that peaks in RT should be found at list onset and at group onset for the two grouped conditions. Method Participants Forty-eight additional University of Western Australia undergraduates aged years participated as part of a first-year psychology course. Apparatus, stimuli, and procedure The only changes from Experiment 1 were as follows. First, each stimulus set, A and B, was extended to 4 practice and 32 test trials. This meant modifying one of the constraints applying to these sets: Across the test trials, each of the 20 consonants occurred at least once but no more than twice in each serial position. Second, there were three between-subjects conditions, with 16 participants assigned at random to each (8 using set A and 8 set B). For each trial in each condition, 1000 ms of silence separated the participantõs touch of the start button and presentation of the first consonant, the total duration of the sequence was 5850 and 630 ms of silence separated the end of the last consonant and appearance of the letter board. The three conditions differed only in the timing of sequence presentation. In the ungrouped condition, each 450 ms consonant was followed by 630 ms of silence (i.e., SOA ¼ 1080 ms). In the loose-grouped condition, the silent intervals separating consonants within a group were 450 ms (SOA ¼ 900 ms), whereas the interval was 1350 ms (SOA ¼ 1800 ms) for the items spanning the two groups (a 1:2 SOA ratio). In the tight-grouped condition, the within-group intervals were 150 ms (SOA ¼ 600 ms) and the between-group interval was 2550 ms (SOA ¼ 3000 ms, and a 1:5 SOA ratio). Results Accuracy Accuracy was higher for grouped compared to ungrouped presentation, with tight grouping yielding more accurate performance than loose grouping. Both grouped conditions showed shallower serial position functions than the ungrouped condition. A 3 (condition: loose-grouped; tight-grouped; ungrouped) 6 (serial position) ANOVA was conducted on the number of items recalled in correct serial position, summed across the 32 test trials. All effects were significant, with F ð2; 45Þ ¼ 8:49, MSE ¼ 54.36, p <:001, for the effect of condition; F ð5; 225Þ ¼40:36, MSE ¼ 6.01, p <:001, for the effect of serial position; and F ð10; 225Þ ¼6:70, MSE ¼ 6.01, p <:001, for the interaction. Means are displayed in Fig. 4. Two follow-up ANOVAs were conducted. The first contrasted the two grouped-list conditions. There were significant main effects for grouping condition, F ð1; 30Þ ¼4:39, MSE ¼ 50.87, p <:05, and serial position, F ð5; 150Þ ¼16:04, MSE ¼ 4.68, p <:001, but a non-significant interaction,

10 M.T. Maybery et al. / Journal of Memory and Language 47 (2002) F ð5; 150Þ ¼ 1:44, MSE ¼ The tight-grouped condition, in providing a clearer temporal definition of the groups, yielded more accurate performance compared to the loose-grouped condition. However, the two conditions provided similar serial position functions (see Fig. 4). The second follow-up analysis compared the ungrouped condition with the two grouped conditions combined. In this case, all effects were significant: F ð1; 46Þ ¼12:06, MSE ¼ 58.03, p <:001, for the comparison of the ungrouped condition to the two grouped conditions; F ð5; 230Þ ¼ 49:43, MSE ¼ 6.02, p <:001, for the effect of serial position; and F ð5; 230Þ ¼12:24, MSE ¼ 6.02, p <:001, for the interaction. When the interaction was examined by calculating trends across serial position, by far the largest component to the interaction involved the quadratic trend, with F ð1; 46Þ ¼40:47, MSE ¼ 6.48, p <:001. Fig. 4 shows that accuracy was substantially higher for the two grouped conditions compared to the ungrouped condition, especially at serial positions 3 and 4, which represent items on the ends of groups. The Appendix reports analyses of order errors. These show that migrations across the group boundary (34) were curtailed relative to other adjacent migrations (12, 23, 45, and 56) for the loose- and tight-grouped conditions compared to the ungrouped condition. Response times The trend observed in Experiment 1 for temporal grouping to increase the speed of response generally but to slow response at the onset of each group was again observed in Experiment 2. Importantly, the RT functions across serial position were almost identical for the loose and tight conditions (Fig. 5). Correct median RTs were submitted to a 3 (condition: loose-grouped; tight-grouped; ungrouped) 6 (serial position) ANOVA. All effects were significant: F ð2; 45Þ ¼ 3:37, MSE ¼ , p <:05, for the effect of condition; F ð5; 225Þ ¼ 169:04, MSE ¼ , p <:001, for the effect of serial position; and F ð10; 225Þ ¼3:42, MSE ¼ , p <:001, for the interaction. The means for the interaction (Fig. 5) show close concordance for the two grouped conditions. Although both conditions showed a peak in the RTs for recall at serial position 4 (the first item of the second group), there was no evidence that this peak was more pronounced in the tight condition compared to the loose condition. This is despite the substantial differences in SOAs for these conditions; specifically, the ratio of within- to between-group SOA was 1:5 for the tight condition compared to 1:2 for the loose condition. There was no indication that the timing of recall mirrored these two ratios. Consistent with Experiment 1, the two grouped Fig. 4. Experiment 2: Number of correct responses (maximum 32; 95% confidence interval for withinsubjects effects ¼:70). Fig. 5. Experiment 2: Response times for correct serial recall (95% confidence interval for within-subjects effects ¼63).

11 370 M.T. Maybery et al. / Journal of Memory and Language 47 (2002) conditions provided RTs that were shorter at each serial position than the RTs for the ungrouped condition, with the exception of serial position 4. However, but again consistent with Experiment 1, the ungrouped condition showed some elevation in RTs for serial positions 3 5 compared to serial positions 2 and 6. These observations were confirmed in two follow-up analyses. First, in an analysis restricted to the RT data for the two grouped conditions, the main effect comparing the tight and loose conditions was not significant, F ð1; 30Þ ¼:53, MSE ¼ , p >:1, nor was the interaction, F ð5; 150Þ ¼:85, MSE ¼ , p >:1. However, not surprisingly, the main effect of serial position was significant, F ð5; 150Þ ¼131:39, MSE ¼ , p <:001. Because the PL and OSCAR models predict an elevated peak in RT at serial position 4 relative to neighboring serial positions for the tight compared to the loose condition, the interaction of tight versus loose with a comparison of serial position 4 versus serial positions 3 and 5 was computed. This interaction yielded F ð1; 30Þ ¼:06, p ¼.82. Thus there was no evidence of differences between the two grouped conditions in the timing of recall. The second follow-up analysis compared the ungrouped condition with the two grouped conditions combined. This comparison provided F ð1; 46Þ ¼ 6:39, MSE ¼ , p <:05, and also entered into an interaction with serial position, F ð5; 230Þ ¼6:27, MSE ¼ , p <:001. The main effect of serial position was also significant, with F ð5; 230Þ ¼ 164:81, MSE ¼ , p <:001. Simple effects analyses showed that the comparison of the ungrouped condition with the two grouped conditions combined was significant at serial positions 1, 3, and 5, smallest t(46) ¼ 2.53, p <:05. The comparison at serial position 6 yielded t(46) ¼ 1.84, p ¼ :07. Finally, a contrast of serial position 4 with serial positions 3 and 5 was significant for the tight condition, F ð1; 45Þ ¼ 15:64, MSE ¼ , p <:001, and the loose condition, F ð1; 45Þ ¼13:45, MSE ¼ , p < :001, but not the ungrouped condition, F ð1; 45Þ ¼ 1:59, MSE ¼ , p >:1. In summary, consistent with Experiment 1, RTs were generally shorter for the two grouped conditions compared to the ungrouped condition, except for serial position 4. This serial position corresponded to the onset of a new group in the grouped conditions. For these two conditions, pronounced peaks were observed in the RT functions at this point in recall. These peaks did not reflect the duration of the interval between groups at presentation, because the RT functions were in close correspondence for the loose and tight conditions. In the ungrouped condition there was again evidence of elevated RTs for serial positions 3 5 compared to serial positions 2 and 6. In this experiment, the preparatory interval was significantly longer for the ungrouped condition compared to the two grouped conditions, confirming the difference of marginal significance reported in Experiment 1. Power analyses Does our design have adequate power to detect differences in RT of the order predicted by the oscillator models in comparing the loose and tight conditions? One pertinent observation is that the differences predicted in contrasting the loose and tight conditions (SOA ratios of 1:2 and 1:5) should be at least as pronounced as the differences predicted in contrasting the ungrouped and loose conditions (SOA ratios of 1:1 and 1:2). 3 This means that analyses comparing the ungrouped and loose conditions can provide conservative estimates of power for analyses comparing the loose and tight conditions. An ANOVA comparing the ungrouped and loose conditions yielded a significant interaction of condition and serial position, with F ð5; 150Þ ¼3:10, MSE ¼ , p ¼ :01, and power ¼.87 (estimated using the procedures provided by Bakeman & McArthur, 1999). Furthermore, the interaction of ungrouped versus loose with a comparison of serial position 4 versus serial positions 3 and 5 yielded F ð1; 30Þ ¼ 11:55, p <:01, and power ¼.91. Thus based on these comparisons of the ungrouped and loose conditions, our design had adequate power to detect the larger differences predicted by the oscillator models in comparing the loose and tight conditions. Yet recall that when the loose and tight conditions were contrasted, the omnibus interaction with serial position yielded F < 1, as did the interaction with the more precise contrast of serial position 4 with serial positions 3 and 5. The non-significance of these effects would not appear to be explicable in terms of low power. 3 Our analysis also holds if conditions are compared as to absolute differences in SOA rather than SOA ratios. For the critical transition from the third to the fourth item, the SOAs for the ungrouped, loose and tight conditions were 1080, 1800, and 3000 ms, respectively. These SOAs differ more when comparing loose with tight than when comparing ungrouped with loose.

12 M.T. Maybery et al. / Journal of Memory and Language 47 (2002) Discussion The results of Experiment 2 are unambiguous and can be summarized in three points. First, both tight and loose manipulations of temporal grouping improved accuracy compared to the ungrouped condition, the effect being stronger for the tight manipulation than for the loose. Second, both of the grouped presentation conditions showed a reduction in errors consisting of migrations across the boundary between groups (see Appendix). Third, and most importantly, temporal grouping modified the timing of responses but the latency functions for the tight and loose conditions were indistinguishable. The latter finding is inconsistent with the position adopted by the oscillator models of Brown et al. (2000) and Burgess and Hitch (1992, 1999), which is that variation in RT should mimic differences in timing at list presentation. Even when allowing for variations in reinstatement speed of the oscillators (Brown et al., 2000), the substantial difference in the ratio of within- to between-group SOA should have led to a longer pause at serial position 4, relative to the other serial positions, in the tight compared to the loose condition. The results of Experiment 2 suggest that temporal grouping affects the timing of recall in a way that excludes the hypothesis of a simple reproduction of the presentation timing or a scaled version of it. The discrepancy between presentation and recall timing is clear, as illustrated in Fig. 6. The oscillator models of Brown et al. (2000) and Burgess and Hitch (1992, 1999), which place time at the core of encoding and retrieval mechanisms, appear unable to account for the lack of correspondence evident in Fig. 6. Moreover, variations of RT similar to those reported in Fig. 6. Experiment 2: Stimulus onset asynchronies (presented) and response onset asynchronies (observed). this study have been observed in serial recall when grouping was imposed not temporally but visuospatially (Anderson et al., 1998; Anderson & Matessa, 1997). According to Anderson and colleagues, this effect is the result of the hierarchical organization of information presented sequentially. A list is organized in groups that are themselves organized in terms of item representations. This allows the prediction from ACT-R of longer RTs for the first item of a group independently of the timing of presentation. Consistent with Experiment 1, RTs for the two grouped conditions tended to fall below those for the ungrouped condition, except at serial position 4. We discussed earlier how this difference could reflect the memory search process postulated by Cowan (1992) if, for grouped conditions, an additional assumption is made that search is restricted to the group currently being recalled. One important aspect of the search process is that the time it takes to identify the next item for recall is purported to depend on list length. In Experiment 3, we varied list length to test the idea of a search process operating at a group level and dependent on group size. Experiment 3 In this experiment, temporal grouping was examined for lists of different lengths. In particular, sequences of length 3, 4, 5, and 6 were presented either with or without an extended temporal gap between the third item and any subsequent items. This manipulation was aimed at assessing the ACT-R model (Anderson et al., 1998), and the search process proposed by Cowan (1992, 1999), with an extension made to it in respect of grouped lists. With regard to the search process identified by Cowan (1992), several predictions can be made for ungrouped lists. First, the preparatory interval should increase with list length, as also should the intervals between successive responses (Cowan, 1992; Cowan et al., 1994, 1998). Further, the intervals between responses should decrease with serial position (Cowan et al., 1998). The first two predictions arise because it is assumed that the search process enacted prior to each response is more prolonged for longer lists. The third prediction follows from the assumption that there is marking/suppression of items as recall unfolds. Predictions can be generated for grouped lists if the search process is presumed to operate within