Task switching: The effect of task recency with dual- and single-affordance stimuli

Transcription

1 THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY 2006, 59 (7), Task switching: The effect of task recency with dual- and single-affordance stimuli Petroc Sumner and Lubna Ahmed Imperial College London, London, UK When we switch to a new task, performance is transiently relatively poor, but improves dramatically after one trial. Such a switch cost may result from the preceding task being highly primed while the new task is not yet primed. This predicts that it should become more difficult to switch back to Task A when more trials of Task B have intervened. Such a lag effect has been found in some but not in most previous experiments, and to resolve this discrepancy we examined the effects of task lag with different stimuli. We found that when stimuli uniquely and clearly cued the task minimizing the need for control switch reaction time increased with task lag. However, when the need for control was increased by using similar or identical stimuli in the two tasks, this lag effect was abolished or reversed. Thus only when control processes are minimized can priming explain the difficulty of switching back from Task B to Task A. Second, we asked how the impact of control is mediated in conditions where it is not minimized. If it is mediated through altering the relative activation states of competing tasks, then as it becomes easier to do one task the relative task-set activation state is tipped in that task s favour it should always become harder to do the other task. On the other hand, if control bias affects switch performance directly, this relationship need not hold. We found that as it becomes easier to perform one task it can become easier, not harder, to switch to the competing task. Thus control bias must act directly on switch performance, rather than only through its influence on relative task-set activation. Cognitive control is often studied by requiring participants to switch repeatedly between simple tasks. Average performance on task-switch trials usually suffers compared to that on task-repeat trials, and the cause or causes of this switch cost remain debated (e.g., Allport, Styles, & Hsieh, 1994; Allport & Wylie, 1999; De Jong, 2000; Gilbert & Shallice, 2002; Goschke, 2000; Meiran, 1996, 2000a; Meiran, Chorev, & Sapir, 2000; Rogers & Monsell, 1995; Ruthruff, Remington, & Johnston, 2001; Sohn & Anderson, 2001; Yeung & Monsell, 2003b). Here we test the extent to which switch costs can be explained by relative task-set priming and, second, whether the influence of control can be explained by an alteration of the relative task-set activation state. Priming may contribute to switch costs in several ways. For example, performance on repeat trials should be facilitated by task-set priming Correspondence should be addressed to Petroc Sumner, School of Psychology, Cardiff University, Tower Building, Park Place, Cardiff, CF10 3AT, UK. sumnerp@cardiff.ac.uk We thank Eric Ruthruff, Nick Yeung, Andy Parton, Clive Rosenthal, and Charlotte Golding for comments on drafts of this article. We are also very grateful to Eric Ruthruff for advice on the design of Experiment 2A. # 2006 The Experimental Psychology Society 1255 DOI: /

2 SUMNER AND AHMED produced by having just performed the same task a simple repetition benefit (e.g., Ruthruff et al., 2001). In addition, it is possible that priming for a previous task-set carries over and causes interference on trials of a new task task-set inertia (e.g., Allport et al., 1994; Allport & Wylie, 1999; Meiran et al., 2000; Sohn & Carlson, 2000). In other words, when Task B is performed after Task A, it is intuitively plausible that as Task B is repeated, priming of Task-Set B builds up while the level of priming reached previously for Task-Set A diminishes. Then on switching back to A, performance is poor for two reasons: Task-Set A is not yet primed no repetition benefit while Task-Set B is highly primed and causes interference (for examples of this idea and mathematical models based on it, see Meiran et al., 2000; Sohn & Anderson, 2001; Sohn & Carlson, 2000; Yeung & Monsell, 2003a). If such an account explains the source of switch costs, a straightforward prediction can be made: If more trials of Task B are performed before switching back to A, the switch reaction time (RT) should be increased. Such a lag effect has been reported by Ruthruff et al. (2001), but it has not been found in other studies. For example, Monsell, Sumner, and Waters (2003) found a small decrease in RT for switch trials with increasing task lag. Task-set priming that occurs simply from repeating the same task, as discussed above, has been labelled autogenous priming (Monsell et al., 2003). This can be logically distinguished from exogenous priming, which is elicited by stimulus task associations. When stimuli are associated with more than one task, a switch trial may suffer not only because a different task has just been performed, but also because the stimulus itself may recently have been associated with this different task, exacerbating the difficulty in switching (e.g., Allport & Wylie, 2000; Waszak, Hommel, & Allport, 2003). This would also predict a lag effect because if more trials of Task B have just been performed, it is more likely that the stimulus on the first trial of Task A will have been recently encountered within Task B and would therefore exogenously prime Task B rather than Task A. On the other hand, an increase of switch RT with increasing task lag is not predicted if the switch cost is to be explained by the application of executive control processes on a switch trial. When switching task, control processes applied in order to perform the new task task-set reconfiguration might cause a switch cost simply by taking time to execute (e.g., De Jong, 2000; Meiran, 1996; Rogers & Monsell, 1995). This reconfiguration process would not be expected to take longer following longer task lags (unless it was influenced by the relative priming state of the competing tasks, in which case the relative priming state, not the reconfiguration process, would be the underlying cause of the extra delay following longer lags). Potential contributions to switch costs may also arise from carry-over effects of inhibition or activation applied to the previous task. For example, if it is necessary to inhibit Task-Set A in order to perform Task-Set B, then the effect of this inhibition might carry over and hinder performance when switching back to Task A (e.g., Allport et al., 1994; Allport & Wylie, 1999). Evidence for such inhibition has been provided by Mayr and Keele (Mayr, 2002; Mayr & Keele, 2000), who found longer RT on the final trial in an ABA task sequence than in a CBA task sequence; they argued that this increase in RT was due to inhibition applied to Task A in order to switch to Task B in the former sequence. If such carryover makes the dominant contribution to switch costs, this predicts either no lag effect or a reverse lag effect increased switch RT with shorter task lags. There is no reason to expect that more inhibition would be carried over if there had been more trials of the intervening task. In fact two lines of thought predict the opposite: If Task-Set A must be inhibited before performing Task B (e.g., Mayr, 2002), then on switching back to A the effect of such inhibition would be greater following fewer, not more, intervening trials of B. Similarly, if control biases against Task A are applied during Task B and in a just-enough manner (Goschke, 2000; Monsell et al., 2003; Schuch & Koch, 2003; Yeung & Monsell, 2003b), the control biases 1256 THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY, 2006, 59 (7)

3 TASK SWITCHING against A may be eased as Task B is repeated and becomes more established, counterintuitively making it easier, not harder, to switch to Task A when Task B is more primed. Note that it is not important here whether such control biases take the form of inhibition of Task A or extra activation of Task B. Neither predict a lag effect. Furthermore, the inhibition or activation could be an example of active top-down control, or an automatic lateral effect (see, e.g., Mayr & Keele, 2000; Schuch & Koch, 2003). The important point is that such inhibition/activation is caused as a result of trying to perform one task in the face of interference from another. We call this inhibition/activation a control bias or control state, because its outcome is to control the interference and lead to a response relevant to the current task. Note also that while the initial application of such control may be top-down or lateral, voluntary or automatic, its carry-over into a new task is considered passive and has been prominent in discussions of taskset inertia (e.g., Allport et al., 1994; Allport & Wylie, 1999; Meuter & Allport, 1999; Yeung & Monsell, 2003a). We use task-set inertia (TSI) as an umbrella term to refer to passive carry-over of any factors associated with performing previous tasks. Thus TSI has at least three components: carry-over of the state of autogenous priming reached during the preceding task; carry-over of exogenous priming (stimulus associations) from previous tasks; and carry-over of control biases invoked to perform previous tasks. Lastly, note that while priming has by some authors been considered under the umbrella of bottom-up processes (e.g., Ruthruff et al., 2001), we avoid doing so here, because priming necessarily entails the involvement of memory traces, and therefore the distinction between bottom-up and topdown processes becomes blurred. Thus the important distinction for our study is between relative task-set priming, caused simply by repeating one task, and carry-over of control biases that were invoked by performing one task in the face of interference from another. The former predicts a positive lag effect an increase of switch RT when more trials of the other task precede the switch while the latter predicts no lag effect, or a reversed lag effect. In addition, control in the form of a discrete reconfiguration process also predicts no lag effect. In this study we took advantage of these different predictions for the effect of lag in order to examine the influence of priming and control on switch performance. We also aimed to explain the discrepancy in lag effects reported in past studies. Ruthruff et al. (2001) found that switch RT increased with longer task lags, while data from other studies show, if anything, the opposite (e.g., Mayr & Keele, 2000; Monsell et al., 2003). The main difference between Ruthruff et al. s (2001) experiments and those of Monsell et al. (2003) and Mayr and Keele (2000) lies in the type of stimuli employed. Ruthruff et al. used single-affordance or univalent stimuli that could be used only for one task (coloured rectangles for a colour task and white letters for a letter task). The other studies used dual-affordance or bivalent stimuli, meaning that either task could be performed with the same stimulus (for example, Monsell et al. used single-digit stimuli, and participants switched between classifying them as high/low and odd/even). It is possible that priming dominates switch costs for singleaffordance stimuli, increasing switch RT with lag as predicted above. However, the effects of control biases may dominate for dual-affordance stimuli, which cause interference between tasksets or responses (e.g., Allport et al., 1994; Mayr, 2002; Mayr & Keele, 2000). As explained above, this would be predicted to eliminate the lag effect. It is known that single-affordance stimuli generally produce much smaller switch costs than dual-affordance stimuli (e.g., Allport et al., 1994; Dreisbach, Haider, & Kluwe, 2002; Jersild, 1927; Rogers & Monsell, 1995; Spector & Biederman, 1976), but there has been little direct comparison of the two stimulus types and no comparison of the effect of lag for each. It is difficult to make direct comparisons because, by definition, the stimuli need to be different in each case, and for dual-affordance stimuli the participant needs a separate cue to indicate which task to perform. THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY, 2006, 59 (7) 1257

4 SUMNER AND AHMED We have taken the following approach: In Experiment 1 we compared single- and dualaffordance stimuli in a single experiment, making the conditions as similar as possible. We found no lag effect for either dual- or single-affordance stimuli. In Experiment 2A we used more easily distinguishable single-affordance stimuli without cueing, in an attempt to replicate the lag effect found by Ruthruff et al. (2001). A lag effect was found. In Experiment 2B we therefore tested the effect of reintroducing cueing and the other procedural differences between Experiments 1 and 2A, while keeping the more easily distinguished stimuli. The lag effect of 2A was replicated. In Experiment 2C we repeated Experiment 2A but with the less easily distinguished single-affordance stimuli, as used in Experiment 1. The lag effect of Experiment 2A was not replicated. We conclude therefore that the crucial determinant of the lag effect was stimulus similarity (rather than a categorical distinction between single and dual affordance). In Experiment 3 we returned to using both dual- and single-affordance stimuli, as in Experiment 1, to investigate whether there is a direct link between the lag effect and the pattern of performance through a run of one task that is, whether as performance improves on one task it becomes harder to switch to the other task, as would be predicted if both the improvement within each task and the switch performance were dependent on the relative priming of the two task-sets. EXPERIMENT 1 Our hypothesis was that an increase of switch RT with task lag is found when single-affordance stimuli are used, as in the study of Ruthruff et al. (2001), while no lag effect is found when dualaffordance stimuli are employed. We therefore directly compared single- and dual-affordance stimuli in a paradigm based on Ruthruff et al. s control experiment, in which they found the clearest increase of switch RT with task lag. Participants switched between responding to letters {I, S, O, X} and colours {red, green, yellow, blue}. In the dual-affordance condition coloured letters were presented for both tasks (so that either task could be performed on the same stimulus). In the single-affordance condition the letters were presented in black and the colours in the shape of a þ. Since dual-affordance stimuli by definition do not cue the task, we needed to introduce a separate task cue. We displayed this cue in both single- and dual-affordance conditions to keep them matched (see Figure 1). The task switches of interest were relatively rare and were not predictable, occurring with a probability of.1 on each trial. Thus there was a high probability of task repetition (.9), and participants were explicitly informed of this. Method Participants A total of 12 volunteers, aged between 21 and 35 years, participated in the study. All had normal or corrected-to-normal acuity and normal colour vision. Apparatus Stimulus presentation and response collection were performed by a Cambridge Research Systems (CRS) VSG 5 board (housed in a PC host) connected to a Sony GDM-F520 Trinitron monitor and a CRS CB3 response box. Stimulus presentation was synchronized with the screen refresh rate of 100 Hz, and all timings were controlled and measured by the CRS clock so that they were not subject to the errors produced by normal PC operating systems. Stimuli In the centre of a dark-grey screen we presented a light-grey circle or square of approximately 28 diameter, which indicated the task to perform: For half the participants the circle was associated with the colour task and the square with the letter task, and for the other participants this was reversed. The task stimuli were 0.58 in height and were presented in the centre of this cue shape. The letters were drawn from the set {I, S, O, X} and the colours from the set {red, green, 1258 THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY, 2006, 59 (7)

5 TASK SWITCHING Figure 1. To view this figure in colour, please go to the online version of the issue. Examples of trials in Experiment 1. Single-affordance stimuli are shown on the left and dual-affordance stimuli on the right. For each stimulus affordance type, two trials of the letter task are shown, followed by two trials of the colour task, indicated by the cue shape (square and circle, respectively). In the figure, coloured stimuli are represented as white, with example colours indicated by labels. In this example, eight trials of the letter task occurred before the task switched, indicated by the black clock hand, which was vertical for the first trial of a new task and proceeded clockwise 458 around the cue shape on each trial. blue, yellow}. In the single-affordance condition the letters were presented in black and the colours in the shape of a þ. In the dual-affordance condition coloured letters were presented for both tasks. The task to be performed was indicated both by the cue shape and by a clock hand protruding from behind the cue shape (similar to the experiments of Monsell et al., 2003). This clock hand moved discretely clockwise (458) each trial and jumped to 12 o clock at each switch (see Figure 1). Procedure In each trial the stimulus was presented until a response was made, using one of two buttons. The same buttons were used for both tasks, so that, for example, the left button could be assigned to I and S in the letter task and red and green in the colour task, and the right button could be assigned to O and X, and blue and yellow (these assignments were counterbalanced across subjects, and both tasks were practised for 48 trials individually so THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY, 2006, 59 (7) 1259

6 SUMNER AND AHMED that participants were familiar with the button assignments before being required to switch between tasks; additionally a card indicating the response mappings was present just below the screen throughout the experiment). If an error was made, feedback was given in the form of an auditory tone immediately following the button press. After a response there was an interval of 500 ms, and then the next trial began, indicated by the clock hand moving to the next position, or if this was a switch trial, the clock hand jumping to the 12 o clock position and the cue shape changing also. The stimulus appeared immediately. Thus participants had some time to prepare for the expected task, but did not have time to prepare for an unexpected switch. The basic procedure was eight trials of one task and then eight trials of the other task, indicated by the clock hand progressing around the cue shape in steps of 458, and participants were instructed to expect this. However, on each of the eight trials in the run, the probability of an unexpected task switch was 10%, meaning that an unexpected switch occurred in about half of the runs. Participants were explicitly informed of this: that sometimes the task would switch early, but that although this would seem to happen often, on any individual trial the probability was only 10%, and thus it was always best to expect task repetitions (until the end of the run of eight). Embedding the unexpected switches in a basic pattern of fixed run length and explicitly informing the participants about the probabilities were considered to be the best ways to avoid very long runs of one task and the participants changing their own expectancies according to the gambler s fallacy (that as the number of repetitions increases the likelihood of a switch increases, Kahneman & Tversky, 1972). Such changing expectancy could potentially obscure a lag effect. Design Participants performed the dual- and singleaffordance conditions on separate days (counterbalanced across subjects). On each day participants performed a practice block of the relevant condition (10 runs of each task) and then one long experimental block of 120 runs of each task (making a total of 1,360 trials). In the practice there were no unexpected switches (all runs were of eight trials) in order to help the participants to expect this pattern. Also, for the same reason, at the start of the experimental block there were no unexpected switches for the first two runs. During the block there were three breaks, coming at equal intervals (and always followed by runs of eight). In the dual-affordance case, for each switch with a given task lag (for each task), the number of congruent trials (in which the stimulus would map to the same response for both tasks) was constrained to be equal to the number of incongruent trials (except where the switch probability dictated that a category of trial occurred an odd number of times so that the number of incongruent and congruent trials was different by one this was counterbalanced across subjects). The numbers of occurrences of each letter and colour combination were also kept as equal as possible, given the number of trials in that category that occurred overall (if the category of trial occurred 16 times or less, each of the 16 letter/colour combinations could occur a maximum of once; if the category occurred between 17 and 32 times, each letter/colour combination would occur once or twice, and so on this was randomized across subjects). Similarly in the single-affordance condition, the occurrences of each colour or letter were kept approximately equal for each category of trial. In all experiments participants were instructed to perform each individual trial as fast as possible whilst minimizing errors. RT and error rate were monitored during the practice blocks, and participants were given feedback and encouraged to try to be more careful or to respond faster as appropriate. Analysis For all experiments, the basic procedure for data analysis was as follows: The first two switch trials of each block were deleted from the analysis. In single-affordance blocks we deleted all trials on which the stimulus was the same as the previous trial, and in the dual-affordance blocks we deleted any trial in which the relevant stimulus attribute (colour or letter) was identical to that in the previous trial, since trials in which the whole 1260 THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY, 2006, 59 (7)

7 TASK SWITCHING stimulus was identical to that in the previous trial would be rarer than in the single-affordance case (as it happens, the pattern of results was the same when these stimulus repeat trials were not deleted). RTs of less than 200 ms or more than 2,000 ms were also excluded (,1%). Separate analyses of variance (ANOVAs; with factors explained in the Results sections below) were performed for percentage error and RT, and in addition, mean RT and median RT for each cell were analysed separately to check that the pattern of results was similar and that all reported effects were present in both mean and median. As this was the case, only the results for mean RT are reported below. Error trials were not included in the RT analysis. Results The lag effect Figure 2A shows performance in Experiment 1 plotted against task lag. The difference between repeat trials (lag 1, unconnected symbols) and switch trials (lag. 1) is clear, and is discussed further below (under the heading The switch cost). However, more importantly for our study, among the switch trials there was no lag effect for either dual-affordance or single-affordance stimuli. ANOVAs were performed on the RT and error rates of the switch trials, with three within-subject factors of task lag (2 8; lag 1 trials are repeat trials), stimulus affordance type (dual vs. single) and task (colour vs. letter), and a between-subjects factor of block order (whether the dual-affordance condition was performed first or second). There was no effect of lag for either RT, F(6, 60) ¼ 0.4, or errors, F(6, 60) ¼ 0.6; corresponding linear trends: F(1, 10) ¼ 0.02 and 1.6. Neither was there an interaction between lag and stimulus type for either RT, F(6, 60) ¼ 1.6, or errors, F(6, 60) ¼ 0.9, even though there was a clear main effect of stimulus affordance type in both RT, F(1, 10) ¼ 45, Figure 2. The results of Experiment 1. Figure 2A shows mean correct RT and error rates plotted as a function of task lag (log scale), following Ruthruff, Remington, and Johnston (2001). The task lag is the number of trials since the same task was performed (i.e., the number of immediately preceding trials of the other task, plus one) such that lag 1 denotes repeat trials (unconnected symbols) and lag. 1 denotes switch trials (e.g., lag 2 is produced by an ABA sequence, lag 3 by ABBA, and so on; longer lags were rarer so lags 5 6 and 7 8 are grouped; lag 9 trials are not plotted as the probability of a switch was different from other trials, see Methods). Figure 2B separates the different repeat trials by plotting the results as a function of position of the trial in a run. Switch trials all have run position 1 (unconnected symbols), and all other run positions are repeat trials (run positions 5 6 and 7 8 are grouped). THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY, 2006, 59 (7) 1261

8 SUMNER AND AHMED MSE ¼ 258,565, p,.001, and error, F(1, 10) ¼ 25, MSE ¼ 273, p,.01. We also reanalysed the data for the single-affordance stimuli having deleted all trials in which the response was repeated, because this is known to affect switch costs (Rogers & Monsell, 1995), but this made no difference to the lack of a lag effect, F(6, 60) ¼ 1.4. The only hint of a lag effect came in the single-affordance stimuli for those participants that performed the single-affordance condition first (RT for lag 2 8: 705, 716, 712, 713, 738, 733, 746 ms), but it did not approach significance: no effect of lag for just these data, F(6, 30) ¼ 0.4, and in the main analysis no interaction of affordance with lag and block order, F(6, 60) ¼ 1.5. There was no sign of any other effect except 7% shorter RT on the letter task than on the colour task, F(1, 10) ¼ 5.2, MSE ¼ 76,925, p,.05. Repeat trials Figure 2B plots performance against position of the trial in a run. All switch trials have run position of 1, and the other run positions are repeat trials. For dual-affordance stimuli, the switch cost extended beyond switch trial, and this effect was shown to be reliable by an ANOVA on the RTs for repeat trials only, with factors of run position (2 8), stimulus affordance type, and block order (between subjects). There was a main effect of run position, F(6, 60) ¼ 7.8, MSE ¼ 627, p,.01, and an interaction of this factor with stimulus type, F(6, 60) ¼ 7.3, MSE ¼ 509, p,.01, caused by the lack of any reduction in RT in the single-affordance stimuli an ANOVA on this stimulus type alone showed no effect of run position, F(6, 60) ¼ 0.8. A post hoc paired t test showed that the reduction of RT from the first to the second repeat trials of the dual-affordance stimuli was reliable, t(11) ¼ 3.6, p,.01. There was also the expected main effect of stimulus affordance type, F(1, 10) ¼ 50, MSE ¼ 5,304, p,.001, and an unsurprising interaction of this with block order as RT was slightly improved for the block performed second, F(1, 10) ¼ 13, p,.01. The errors did not show the same extension of switch cost into the first repeat trial. In fact there was a slight but reliable increase in error rate throughout the whole run, F(6, 60) ¼ 3.4, MSE ¼ 8.7, p,.05. The only other effect in error rate was the overall difference between dual- and single-affordance stimuli, F(1, 10) ¼ 5.0, MSE ¼ 46, p,.05. The relationship between the pattern of repeat trial performance and the lag effect is discussed in the introduction to Experiment 3. The switch cost The simple switch cost was not the key interest of this experiment, but in order to compare it for dual- and single-affordance stimuli, ANOVAs were performed on RT and error rates of all switch trials versus all repeat trials, divided also by the within-subject factor of stimulus type and the between-subjects factor of block order. The switch cost was robust in both RT, F(1, 10) ¼ 168, MSE ¼ 11,397, p,.001, and error rate, F(1, 10) ¼ 13, MSE ¼ 34, p,.01, and it was about twice as large for the dual-affordance stimuli than for the single-affordance stimuli in both RT (524 vs. 273 ms), F(1, 10) ¼ 37, MSE ¼ 5,070, p,.001, and error rate (8.6%, vs. 3.5%), F(1, 10) ¼ 13, MSE ¼ 5.8, p,.01. There were no effects of block order (all F, 1). Congruency Because we used the same buttons for responses for both tasks, we can analyse the effect of congruency for the dual-affordance stimuli it is generally found that RT is longer for incongruent stimuli (which map to different responses in the two tasks) than for congruent stimuli (which map to the same response). In the switch trials there was a substantial congruency effect in RT (1,181 vs. 1,013 ms), F(1, 10) ¼ 19, MSE ¼ 63,591, p,.01, and error rate (29% vs. 8%), F(1, 10) ¼ 95, MSE ¼ 370, p,.001, but neither of these interacted with lag, F(6, 60) ¼ 0.8, F(6, 60) ¼ 1.5. (The congruency effects in RT and error rate for switch trials with lags 2 8 were: 221, 267, 117, 123, 173, 118, 177 ms and 23%, 22%, 15%, 20%, 22%, 18%, 18%.) In repeat trials, the congruency effect was smaller, and it was significant in the errors (13% vs. 9%), F(1, 10) ¼ 23, MSE ¼ 28, p,.01, but not for RT (531 vs. 519 ms), F(1, 10) ¼ 3.6, MSE ¼ 1,478, 1262 THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY, 2006, 59 (7)

9 TASK SWITCHING.05, p,.1, and there were no differences among repeat trials in either RT or errors, F(6, 60) ¼ 1.7, F(6, 60) ¼ 0.6. This reduction in congruency effect from switch to repeat trials was highly robust in RT, F(1, 10) ¼ 17, MSE ¼ 5,151, p,.01, and in error rate, F(1, 10) ¼ 70, MSE ¼ 29, p,.001. Insum,therewasacongruencyeffect,whichwas larger for switch trials, but which did not interact with the lag effect or the repeat trial performance. See Experiment 3 for discussion of the congruency effect. Discussion In the introduction we argued that if switch performance is determined by carry-over of autogenous priming, then a positive lag effect is predicted (the more trials of A the harder it becomes to switch to B). Furthermore, for dualaffordance stimuli, exogenous priming would be expected to make the lag effect larger when stimuli are associated with more than one task, a switch trial may suffer not only because a different task has just been performed, but also because the stimulus itself may recently have been associated with this different task, exacerbating the difficulty in switching (e.g., Allport & Wylie, 2000; Waszak et al., 2003). The absence of any clear lag effect stands in contrast to these predictions based on autogenous and exogenous priming. On the other hand, the results are consistent with the switch cost being determined by carry-over of control bias, or by a reconfiguration process, neither of which predicts a lag effect. It is worth noting that this is not to say that priming has no influence on performance. The fact that the dual-affordance stimuli were associated with more than one task (exogenous priming) raised RT globally. However, changes in exogenous priming during a run of a task would be predicted to make it harder to switch to a different task that uses the same stimuli a positive lag effect and this was not found. Thus the effects of exogenous priming seem to be relatively long-lived, rather than prone to change over a run of up to eight trials. The results of Experiment 1 do not explain why Ruthruff et al. (2001) found a lag effect while other studies have not. We hypothesized that no lag effect would be found for dual-affordance stimuli, for which effects of control bias may dominate the switch cost, but a lag effect would be present for single-affordance stimuli, where little control bias occurs, and effects of priming dominate the switch cost. Why was there no lag effect for single-affordance stimuli? Two main differences between Experiment 1 and Ruthruff et al. s (2001) experiments were introduced so that our dual- and single-affordance conditions would be as similar as possible. One potentially important difference was that our single-affordance stimuli (black letters and coloured þ symbols) were chosen to be similar to the dual-affordance stimuli (coloured letters), whereas Ruthruff et al. used much more easily discriminable stimuli for each task (letters and coloured rectangles). The latter are more likely to minimize the need for control bias that prevents activation of the inappropriate task. Even though the single-affordance stimuli in Experiment 1 were, by definition, not the same in each task, they may have been similar enough to each other and to the dualaffordance stimuli to evoke some control biases in a similar way to dual-affordance stimuli. This is consistent with the much larger switch cost obtained for the single-affordance stimuli of Experiment 1 than has been obtained in previous studies using single-affordance stimuli (another reason for large switch costs is that the participants had no warning of the unexpected and relatively rare switches before the stimuli appeared). It is also consistent with the hint of a lag effect in the single-affordance case for those participants that had not yet encountered the dual-affordance stimuli (but note that block order can only be a small contributory factor because the lag effect did not approach significance, and the switch cost was just as large for these participants: 270 vs. 277 ms). A second potentially important difference was that task cues were presented in all blocks in Experiment 1, and even though cues are by definition not necessary for the single-affordance THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY, 2006, 59 (7) 1263

10 SUMNER AND AHMED case, participants may still have paid attention to them. There were also other methodological differences between Experiment 1 and Ruthruff et al. s (2001) control experiment (in which the clearest lag effect was found), which might possibly have influenced the outcome: a shorter intertrial interval; and a slightly lower chance of an unexpected switch (.1 vs..16; E. Ruthruff, personal communication, April 25, 2003). Finally, it is possible that using only two response buttons might encourage participants to use a strategy of if same category stay, if other category shift to govern their responses rather than consistently responding according to the intended response mappings (N. Yeung, personal communication, February 9, 2004), and this could potentially obscure a lag effect. Therefore in Experiment 2 we first attempted to maximize our chances of measuring a lag effect at all, and then we tested whether the key factor was stimulus similarity, cueing, or any other factor mentioned above. EXPERIMENT 2A Experiment 2A was based on Ruthruff et al. s (2001) control experiment, in which they found the clearest increase of switch RT with task lag. Participants switched between responding to letters {I, S, O, X} and coloured rectangles {red, green, yellow, blue}. Thus the stimuli clearly and unambiguously cued which task should be performed and were never associated with the other task. Four response buttons were used so that participants could not adopt a strategy of if same category stay, if other category shift. Method Participants A total of 8 volunteers, aged between 21 and 27 years, consented to participate in the study. All had normal or corrected-to-normal acuity and normal colour vision. None had participated in Experiment 1. Procedure Black letters or coloured rectangles (all 0.58 in height) were presented in the centre of a grey screen. Each of the letters {I, S, O, X} was associated with one of four response buttons {top left, bottom left, top right, bottom right}, which participants were instructed to press using the middle and index fingersofeachhand.similarly,eachcolour{red, green, blue, yellow} was associated with one button, and all the response mappings were counterbalanced over subjects. No other stimuli were presented in addition to the letters or the rectangles. In each trial, the stimulus was presented until a response was made. Errors were indicated by an auditory tone immediately following the button press. After each response there was an interval of 1,500 ms, and then the next stimulus was presented. On each trial there was an 84% chance that the task would be the same as the previous one (letter or colour), and a 16% chance of a task switch. Participants were informed of this and were encouraged to expect the same task (the starting task was counterbalanced across subjects). Each subject performed 2,000 trials in 100-trial blocks. Participants had 100 trials of practice on each task and then a further 200 trials of practice at the switching regime before data collection. A card indicating the response mappings was present just below the screen throughout the experiment. Results The lag effect The results for Experiment 2A are plotted as solid black squares in Figure 3. Figure 3A plots performance against task lag, showing that RT was shorter in the repeat trials (lag 1, unconnected symbol) than in the switch trials (lag. 1, connected symbols). This is the expected switch cost and in itself is not of key interest in this study. Our main interest lay in whether RT on switch trials increased as the task lag increased, and indeed such a lag effect is clearly evident. This was shown to be significant, F(5, 35) ¼ 5.6, MSE ¼ 1,904, p,.01; linear trend, F(1, 7) ¼ 24, MSE ¼ 2,194, p,.01, by a repeated measures ANOVA on RT, with factors of task (colour or 1264 THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY, 2006, 59 (7)

11 TASK SWITCHING Figure 3. The results of Experiments 2A, 2B, and 2C. Figure 3A shows mean correct RT and error rates plotted against task lag, as in Figure 2A (data for rarer trials with lag. 5 are grouped into three bins: lag 5 6, lag 7 8, lag 9 11; note that the 9 11 bin does not exist for Experiment 2B because it followed the method of Experiment 1). Figure 3B separates the repeat trials as a function of run position, and all switch trials are plotted together as run position 1 (run positions 5 6, 7 8, and 9 11 are grouped). letter) and lag (lag 1 trials were not included). There was no interaction between task and lag, F(5, 35) ¼ 2.0, even though there was an effect of task, such that responses in the colour task were 50 ms slower overall than those in the letter task, F(1, 7) ¼ 13, MSE ¼ 4,709, p,.01. For the error rates an equivalent ANOVA showed no significant effects. Repeat trials Figure 3B plots performance against position of the trial in a run. Again the difference between switch trials (run position 1) and repeat trials (run position. 1) is evident. More interestingly, there was also a continued decrease in RT amongst the repeat trials as the switch became more distant in time. This was shown to be significant, F(6, 42) ¼ 9.3, MSE ¼ 389, p,.001; linear trend, F(1, 7) ¼ 45, MSE ¼ 400, p,.001, by an ANOVA on RT, with factors of task (colour or letter) and run position (2 8). There was no interaction between this effect and task, F(6, 42) ¼ 1.1, and the main effect of task was not significant either, F(1, 7) ¼ 2.0, although overall RT for the letter task was 25 ms faster than that for the colour task. For the error rates an equivalent ANOVA showed no significant effects. EXPERIMENT 2B In contrast to Experiment 1, Experiment 2A succeeded in producing a lag effect broadly replicating the results obtained by Ruthruff et al. (2001). Since both the stimuli and the procedure were changed between Experiment 1 and Experiment 2A, the exact cause of the differing lag effect remains to be determined. Therefore in Experiment 2B we kept the stimuli employed in Experiment 2A, but used the procedure of Experiment 1, in order to discover whether the lag effect was affected most by the similarity of stimuli or by the presence or absence of a cueing regime (or any other factor such as the number of response buttons, intertrial interval or the slight difference in switch probability). THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY, 2006, 59 (7) 1265

12 SUMNER AND AHMED Method Participants A total of 8 volunteers, aged between 21 and 29 years, participated in the study. All had normal or corrected-to-normal acuity and normal colour vision. None had participated in Experiments 1 or 2A. Procedure Participants switched between responding to letters {I, S, O, X} and coloured rectangles {red, green, yellow, blue}, as in Experiment 2A. All other aspects of the experiment, including the switching regime, the shape cues, and clock-hand cues, were the same as those in the singleaffordance condition of Experiment 1, except that more trials were performed (2,000 per subject). Results The lag effect The grey squares in Figure 3 show the results for Experiment 2B, which are remarkably similar to those obtained for Experiment 2A. As well as the usual difference between repeat trials (lag 1) and switch trials (lag. 1), there was a clear lag effect within the switch trials (Figure 3A). This was shown to be significant, F(4, 28) ¼ 3.6, MSE ¼ 1,645, p,.05; linear trend, F(1, 7) ¼ 9.5, MSE ¼ 2,226, p,.05, by an ANOVA on RT, with factors of task (colour or letter) and lag (lag 1 excluded). Switch RT for the letter task was 60 ms faster than that for the colour task, F(1, 7) ¼ 9.5, MSE ¼ 7,635, p,.05, but there was no interaction between task and lag, F(4, 28) ¼ 1.3. For the error rates a parallel ANOVA showed no significant effects. Repeat trials Figure 3B shows that there was again a progressive decrease in RT through repeat trials. Interestingly, the point at which RT on repeat trials stopped decreasing (run position 6) seems to correspond with the point at which switch trial RT stopped increasing with lag (lag. 6). An ANOVA was performed on RT, with factors of task (colour or letter) and run position (2 8). The main effect of run position was significant, F(6, 42) ¼ 7.2, MSE ¼ 327, p,.001; linear trend, F(1, 7) ¼ 4.7, MSE ¼ 472, p ¼.67; quadratic, F(1, 7) ¼ 26, MSE ¼ 206, p ¼.001. The 24- ms difference (letter faster) between the tasks was not significant, F(1, 7) ¼ 2.8, but there was an interaction between task and run position, such that the reduction in RT through the repeat trials was more pronounced in the colour task than in the letter task, F(6, 42) ¼ 4.1, MSE ¼ 251, p,.01; linear contrast, F(1, 7) ¼ 12, MSE ¼ 299, p ¼.01. For the error rates a parallel ANOVA showed no significant effects. EXPERIMENT 2C Experiment 2B replicated the results of Experiment 2A, suggesting that the crucial factor in determining the lag effect was the degree of similarity between the single-affordance stimuli, rather than the cueing regime or any other methodological difference between Experiments 1 and 2A. To test this in Experiment 2C, we repeated Experiment 2A but with the more similar stimuli used in Experiment 1. This would be predicted to reduce the lag effect. Method Participants A total of 8 volunteers, aged between 22 and 27 years, participated in the study. All had normal or corrected-to-normal acuity and normal colour vision. None had participated in Experiments 1, 2A, or 2B. Procedure Participants switched between responding to letters {I, S, O, X} and colours in the shape of a þ {red, green, yellow, blue}, as in the singleaffordance blocks of Experiment 1. All other aspects of the experiment were the same as those in Experiment 2A, including the analysis and ANOVAs reported below THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY, 2006, 59 (7)

13 TASK SWITCHING Results The lag effect The open squares in Figure 3 show the results for Experiment 2C. There was little if any lag effect, F(5, 35) ¼ 0.9; linear trend, F(1, 7) ¼ 2.6, a result clearly different from that of Experiment 2A. Indeed, including the results of Experiment 2A and 2C in the same ANOVA revealed a robust interaction between the within-subject variable lag and the between-subjects variable experiment, F(5, 35) ¼ 3.6, MSE ¼ 630, p ¼.01; linear trend interaction, F(1, 7) ¼ 17.0, MSE ¼ 637, p ¼.004. There were no significant effects in the error rates. Repeat trials There was no reduction in RT through repeat trials, F(5, 35) ¼ 0.5; linear trend, F(1, 7) ¼ 0.4. This also contrasts with the results of Experiment 2A; interaction between lag and experiment: F(5, 35) ¼ 6.5, MSE ¼ 113, p,.001; linear trend interaction, F(1, 7) ¼ 24.2, MSE ¼ 127, p ¼.002. Discussion Experiments 2A and 2C were the same except for the stimuli employed, but the lag effect differed. Likewise Experiments 1 and 2B used the same procedure, but different stimuli, and the lag effect differed. Experiments 2A and 2B used the same stimuli, but differed in several other respects, such as the presence or absence of cues, but the lag effects were remarkably similar. Therefore the most important factor in determining the lag effect seems to be the stimuli employed. Thus we can conclude that carry-over of priming influences the switch cost as long as the stimuli for the two tasks are clearly discriminable, and control bias is minimized. When the stimuli for the two tasks become similar enough to provoke control biases, carry-over of the effects of such control seems to dominate, reducing or eliminating the lag effect (and it seems likely that this is exacerbated if the participants also have experience of dualaffordance blocks using similar stimuli). A discrete reconfiguration event, occurring only on switch trials, might also contribute to the switch cost as this would similarly predict no lag effect. This stimulus-dependency would explain the discrepancies in previous studies. Rather than a simple categorical difference between dual- and single-affordance stimuli, the critical factor seems to be the extent to which stimuli used in the two tasks are similar to each other, with the dualaffordance case being the extreme where the stimuli for each task are identical. It is worth asking whether single-affordance stimuli elicit task switching at all. Since there are no incongruent stimuli (which map to different responses in each task), participants could encode all the stimulus response (S R) mappings as a single task and thus be switching only between stimulus categories, not between task-sets. The switch cost might then be explained in one of two ways: (a) The cost comes from reactivating the perceptual processing of colour or the perceptual processing of letters (even though all the S R mappings are kept active); (b) the cost is associated with revising S R mappings, but on an individual basis, not as a task grouping. These arguments have in common the idea that what is not happening at a switch is reactivation of a group of S R mappings (i.e., a task-set). The first argument would find it difficult to explain the large switch costs of our experiments, which are considerably larger than those obtained previously with single-affordance stimuli (e.g., Meiran, 2000b; Ruthruff et al., 2001). We are not aware of any precedent for it costing anywhere near as much to switch processing of perceptual categories. The second argument, that S R mappings are revived on an individual basis, could explain the raised RT of a switch trial, but not the fact that RT reduces again on the first repeat trial. On changing from letters to colours, for example, none of the colour S R mappings will have been used during the run of letter stimuli, so the S R mapping needed on the switch trial will not have been recently activated. However, neither will the S R mapping needed on the second colour trial (remember that stimulus repeat trials were deleted from the analysis), so the RT should not be reduced for the second trial in a run, unless THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY, 2006, 59 (7) 1267

14 SUMNER AND AHMED the second trial s mapping was partially retrieved during the switch trial. Thus the association between colour S R mappings must be greater than that between letter and colour mappings, and this is one way of expressing what is meant by a task-set. EXPERIMENT 3 Experiments 1 and 2 demonstrated that carry-over of priming state does influence switch performance, but that as soon as the stimuli for the competing tasks become similar, the effect of control dominates. Now we ask whether the state of control might itself be carried over to affect switch performance or whether it affects switch performance indirectly via an influence on relative task-set activation levels that are carried over. In Experiments 1 and 2 there seems to have been a simple relationship between the lag effect and the pattern of performance throughout a run of one task. In Experiments 2A and 2B, the increase in switch RT with lag was accompanied by a decrease in RT through repeat trials. In Experiments 1 and 2C there was little or no lag effect and a little or no reduction in RT through repeat trials. Such a relationship between preswitch RT and switch RT fits well with the hypothesis that performance is determined simply by the relative activation states (or priming) of each task-set the easier it becomes to do one task the harder it becomes to do the other (e.g., Meiran et al., 2000; Sohn & Anderson, 2001; Sohn & Carlson, 2000; Yeung & Monsell, 2003a). Thus the lack of any lag effect in Experiment 1 could be explained not by the carry-over of control states, but rather by the fact that performance on the preswitch task was not improving with more repeat trials. Some researchers have previously suggested that autogenous priming often reaches asymptote very quickly (e.g., Allport & Wylie, 1999; Gilbert & Shallice, 2002; Yeung & Monsell, 2003b), and if the relative activation or priming of each task-set does not change with longer lags, then this can explain why there was no lag effect. This is not to say that control bias is not involved it influences how soon the effect of autogenous priming reaches asymptote (see Gilbert & Shallice, 2002; Monsell et al., 2003; Yeung & Monsell, 2003b) but its effect on switch costs may be indirect. Thus it may be that control biases, such as inhibition, push autogenous priming to asymptote early and thus destroy the lag effect indirectly, rather than abolishing it directly via carry-over of the control bias itself. To distinguish between these two possibilities, we need to create a situation in which we have not minimized the need for control and in which performance through a run of one task continues to improve during repeat trials. If this improving performance then makes it increasingly hard to switch to the next task, this would indicate that the major direct influence on switch costs is simply the relative activation state of task-sets when the activation state of Task A goes up, and the activation state of Task B goes down, it becomes easier to do A and harder to do B. But if there is still no such lag effect that is, no direct relationship between the ease of doing Task A and difficulty of doing Task B this would show that relative task-set activation states cannot fully account for switch costs, and thus the control bias must carry over to influence the switch directly, not only via its influence on relative task-set activation. Thus the aim of Experiment 3 was to produce a gradual reduction in RT through a run of dualaffordance stimuli and to test what effect this might have on switch trials with different task lags. To do this, we made the task switches and repetitions entirely unpredictable because Monsell et al. (2003) found that whereas predictable switching produced equivalent performance for all repeat trials, with unpredictable switching there was a more gradual approach to baseline performance. We found a hint of this effect in Experiment 1: The first repeat trial for dualaffordance stimuli had a longer RT than later repeat trials. This result lay between the two patterns found by Monsell et al., consistent with the fact that our task repetition probability of.9 lay between their extremes of complete predictability 1268 THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY, 2006, 59 (7)