Divided attention and interruption: the multi-resource model and its practical costs

An interface that asks for two things at once is making a claim about attention. A notification can be read while a form is completed, or a call taken while driving. Work on divided attention and interruption says the cost of such pairings is real, and its size turns on which demands the two tasks share rather than on how many tasks there are. Unit 3 asked what selection leaves out. This one asks what happens when two tasks meet. The strongest counter is structural. Pashler (1994) finds a single central bottleneck, so interference is largely all-or-nothing rather than graded by shared channels, and the cost may fall on affect rather than speed, since interrupted tasks were completed faster with no loss of quality (Mark et al., 2008). The cheaper remedy is a suspended task that the user can take up again.

One pool of attention, or several

The oldest answer treats attention as a single general capacity. Kahneman (1973) argued that a mental activity needs an information input specific to its structure, and a second, non-specific input that he called effort, capacity or attention, drawn from a limited supply and allocated with considerable freedom across concurrent activities. On his own summary, interference is non-specific and depends only on the demands of the two tasks.

Norman and Bobrow (1975) supplied the test. A process whose performance improves as more resources are supplied is resource-limited. One that does not improve however much is supplied is data-limited. A data-limited task therefore shows no interference however allocation changes.

Work a pairing to see the difference. A user listens to a spoken instruction while typing a reply. On the single-capacity account the two draw on one supply of effort, so their combined demand predicts the cost. Now let the same user listen while clicking a button. The demand is much the same, so the account predicts much the same interference. That prediction is what the multi-resource model overturns.

The multi-resource model is the field's working answer. Wickens (2002) sets out four dimensions along which two tasks compete. The dimensions are stage (perceptual and cognitive against response), sensory modality (auditory against visual), code (visual against spatial) and visual channel (focal against ambient). Interference is therefore graded rather than all-or-nothing. Those channels belong to a body, and a user who works through a screen reader has fewer of them. A spoken message and a spoken reply compete more than a spoken message and a manual click.

Work the pairing on a driving task. A simulated driver holds a spoken conversation in one condition, and taps a touchscreen to answer the same questions in another. The words are identical and the demand is comparable. The first pairing shares its vocal channel with nothing, while the second takes the hands and eyes that steering already needs. Interference is then graded by the dimensions the two tasks hold in common, and not by their number. Wickens (2008) reports that the four dimensions accounted for driving-simulation interference. The case shows the prediction, and not that those are the only dimensions that matter.

Interference is a pairing, not a count

Wickens (2008) restated the model and drew its boundary: it accounted for task interference and predicted driving-simulation data with high correlations, but it is only partly relevant to mental workload, and it matters most under multi-task overload.

A rival account dispenses with resources. Pashler (1994) reviewed the dual-task literature and concluded that the psychological refractory period indicates a stubborn bottleneck. That bottleneck covers action selection and probably memory retrieval, while perceptual and motor stages overlap. Interference is then structural rather than graded, appearing whenever both tasks need central processing, whatever their modalities.

Salvucci and Taatgen (2008) offer a synthesis. In threaded cognition each task is a thread coordinated by a serial procedural resource and run across perceptual and motor resources. Interference appears when two threads call on the same resource at once, and no separate executive is required.

An interruption suspends a goal, and resumption pays for the decay

Interruption is the everyday form of the problem, and the best account of its cost is about memory rather than capacity. Altmann and Trafton (2002) proposed that a goal is held with an activation level that decays while it is suspended. Returning to an interrupted task therefore requires reactivating the goal, and the residual activation of older goals interferes with a new one. They tested the account against latency and error data from the Tower of Hanoi.

Work the resumption on a long form. A user is midway through a booking, leaves to answer a message, and returns. The suspended goal has decayed while the interruption ran. The older, near-identical goal of the previous booking is still interfering, so the user re-reads the page to work out where they were. That is decay and interference in one event. The account was tested on the Tower of Hanoi rather than on a booking form, so the size of the effect here is a prediction.

Trafton and colleagues (2003) tested the lever that follows. Participants who received an eight-second warning before a pending interruption resumed the interrupted task more quickly than those interrupted without warning. The unwarned group caught up with practice, so a cue that re-establishes the goal is a designable object.

Monk, Trafton and Boehm-Davis (2008) varied what the interruption demands, and not only how long it lasts. Across three experiments, longer and more demanding interruptions produced longer resumption times. The resumption profile followed a logarithmic curve that began to level off between about thirteen and twenty-three seconds. The interaction of duration and demand pointed to rehearsal slowing the decay.

The cost appears in different places

Bailey and Konstan (2006) ran a controlled experiment with fifty participants in which only the timing of a peripheral task was manipulated. Interrupting execution made primary tasks between three and twenty-seven per cent slower and doubled the errors committed across the session, and it raised annoyance by between thirty-one and one hundred and six per cent while roughly doubling the rise in anxiety. Only timing changed, so the affect cannot be attributed to a heavier workload.

Mark, Gudith and Klocke (2008) found that interrupted tasks were completed in less time with no difference in quality, apparently because people worked faster to compensate, though after about twenty minutes they reported more stress, frustration, time pressure and effort. A system measured by time on task alone would read that work as efficient. Leroy (2009) adds a cost that outlasts the interruption: in two experiments people found it hard to disengage from an unfinished task, and their performance on the next task suffered.

When an interruption arrives, not only that it arrives

Timing is a separate variable. Adamczyk and Bailey (2004) interrupted participants at moments chosen from task models of predicted cognitive load. They found that the moment changed emotional state and positive social attribution. On this evidence the claim about timing is a claim about affect, and not about the speed of resumption. Iqbal and Horvitz (2007) logged the activity of twenty-seven users for two weeks. Return to the application of a suspended task averaged nine minutes thirty-three seconds after an email alert, and the resumption phase after an immediate response averaged sixteen minutes thirty-three seconds, with twenty-seven per cent of suspensions running beyond two hours.

Contested question: shared resources, one bottleneck, or neither?

The live dispute is not whether interruption costs something, but what makes two tasks interfere. The multiple-resource position (Wickens, 2002, 2008) holds that interference scales with the number of shared dimensions. The single-bottleneck position (Pashler, 1994) holds that central processing is serial. On that view interference is largely all-or-nothing, and it appears whenever both tasks need response selection. A single-capacity account, the ancestor of both, holds that interference is non-specific and depends only on combined demand (Kahneman, 1973). Salvucci and Taatgen (2008) show that one threaded mechanism can reproduce both patterns. A synthesis that fits both is not a test that separates them. The three positions are not equally supported. The resource account fits studies that vary a modality, and the bottleneck account fits the psychological refractory period, where response selection is the shared demand.

The evidence that would settle it is a dissociation. Take two dual-task pairs equated for combined demand. The first shares a sensory modality but not response selection. The second shares response selection but not a modality. Multiple-resource theory predicts more interference in the first, and the bottleneck account more in the second. A study of that shape would discriminate between them, with both models fitted to the same data and their free parameters reported.

While the dispute is open, a designer can act on what all three accounts allow. Reduce the overlap in response selection, because that is the demand every account treats as costly. Warn before an interruption, and leave a visible record of where the user stopped. Judge the remedy by the state of the suspended task, not by the notification's length.

Where the argument is heading

Two moves are visible. The first treats an interruption as a process with separable parts. Trafton and colleagues (2003) separate the warning lag from the interruption itself. Monk and colleagues (2008) separate duration from demand. Leroy (2009) separates the switch from the residue that survives it. The second move takes the remedy from the person to the system. If the binding constraint is the state of the suspended goal, the fix is a cue that restores it, and the evidence here would not support a briefer notification instead.

The sources this unit rests on

• Adamczyk, P. D. and Bailey, B. P. (2004). If not now, when? The effects of interruption at different moments within task execution. Proceedings of CHI '04, 271-278. doi:10.1145/985692.985727 Interrupted participants at moments selected from task models of predicted load and found that the moment changed emotional state and positive social attribution. On this evidence the timing claim concerns affect rather than the speed of resumption.
• Altmann, E. M. and Trafton, J. G. (2002). Memory for goals: an activation-based model. Cognitive Science, 26(1), 39-83. doi:10.1016/s0364-0213(01)00058-1 Modelled a suspended goal as decaying activation that must be reactivated, and tested the account against latency and error data from the Tower of Hanoi. It supplies the unit's mechanism for resumption cost.
• Bailey, B. P. and Konstan, J. A. (2006). On the need for attention-aware systems: measuring effects of interruption on task performance, error rate, and affective state. Comput. Hum. Behav., 22(4), 685-708. doi:10.1016/j.chb.2005.12.009 Varied only the timing of a peripheral task in a controlled experiment. Interrupting execution slowed primary tasks by three to twenty-seven per cent, doubled errors across the session, raised annoyance by thirty-one to one hundred and six per cent and doubled the rise in anxiety.
• Iqbal, S. T. and Horvitz, E. (2007). Disruption and recovery of computing tasks: field study, analysis, and directions. Proceedings of CHI '07, 677-686. doi:10.1145/1240624.1240730 Logged twenty-seven users for two weeks. Return to the application of a suspended task averaged nine minutes thirty-three seconds after an email alert, the resumption phase after an immediate response averaged sixteen minutes thirty-three seconds, and twenty-seven per cent of suspensions ran beyond two hours.
• Kahneman, D. (1973). Attention and Effort. Prentice-Hall. Available at: https://archive.org/details/attentioneffort0000kahn Argued that attention is a single limited supply of effort allocated flexibly across concurrent activities, indexed by pupil dilation. His summary that interference is non-specific and depends only on the combined demand is the position later qualified.
• Leroy, S. (2009). Why is it so hard to do my work? The challenge of attention residue when switching between work tasks. Organ. Behav. Hum. Decis. Process., 109(2), 168-181. doi:10.1016/j.obhdp.2009.04.002 Ran two experiments on disengagement from an unfinished task. People found it hard to move attention away from unfinished work and their performance on the next task suffered.
• Mark, G., Gudith, D. and Klocke, U. (2008). The cost of interrupted work: more speed and stress. Proceedings of the CHI Conference on Human Factors in Computing Systems (CHI '08), 107-110. doi:10.1145/1357054.1357072 Found that interrupted tasks were completed in less time with no loss of quality, apparently because people worked faster, while they reported more stress, frustration, time pressure and effort after about twenty minutes.
• Monk, C. A., Trafton, J. G. and Boehm-Davis, D. A. (2008). The effect of interruption duration and demand on resuming suspended goals. Journal of Experimental Psychology: Applied, 14(4), 299-313. doi:10.1037/a0014402 Ran three experiments varying interruption duration and demand. Longer and more demanding interruptions produced longer resumption times, and the resumption profile followed a logarithmic curve that levelled off between about thirteen and twenty-three seconds.
• Norman, D. A. and Bobrow, D. G. (1975). On data-limited and resource-limited processes. Cognitive Psychology, 7(1), 44-64. doi:10.1016/0010-0285(75)90004-3 Distinguished processes whose performance improves when more resources are supplied from those whose performance does not, so that competition matters only for the first kind.
• Pashler, H. (1994). Dual-task interference in simple tasks: data and theory. Psychological Bulletin, 116(2), 220-244. doi:10.1037/0033-2909.116.2.220 Concluded from the dual-task literature that the psychological refractory period reflects a stubborn central bottleneck covering action selection and probably memory retrieval. It is the unit's structural rival to the resource model.
• Salvucci, D. D. and Taatgen, N. A. (2008). Threaded cognition: an integrated theory of concurrent multitasking. Psychological Review, 115(1), 101-130. doi:10.1037/0033-295x.115.1.101 Models concurrent tasks as threads coordinated by a serial procedural resource across perceptual and motor resources, with no separate executive, and predicts interference and its absence across dual-choice tasks and driving.
• Trafton, J. G., Altmann, E. M., Brock, D. P. and Mintz, F. E. (2003). Preparing to resume an interrupted task: effects of prospective goal encoding and retrospective rehearsal. International Journal of Human-Computer Studies, 58(5), 583-603. doi:10.1016/s1071-5819(03)00023-5 Gave some participants an eight-second warning before a pending interruption; they resumed the task faster than participants interrupted without warning. It shows a cue restoring the suspended goal is a designable object.
• Wickens, C. D. (2002). Multiple resources and performance prediction. Theoretical Issues in Ergonomics Science, 3(2), 159-177. doi:10.1080/14639220210123806 Sets out a four-dimensional multi-resource model, in stage, modality, code and visual channel, in which interference grows with the number of shared dimensions.
• Wickens, C. D. (2008). Multiple resources and mental workload. Human Factors, 50(3), 449-455. doi:10.1518/001872008x288394 Restates the model for the workload literature, reports that the four dimensions accounted for task interference, and notes that the model is only partly relevant to workload and matters most under multi-task overload.