On the optimality of serial and parallel processing in the psychological refractory period paradigm: Effects of the distribution of stimulus onset asynchronies

Abstract

Within the context of the psychological refractory period (PRP) paradigm, we developed a general theoretical framework for deciding when it is more efficient to process two tasks in serial and when it is more efficient to process them in parallel. This analysis suggests that a serial mode is more efficient than a parallel mode under a wide variety of conditions and thereby suggests that ubiquitous evidence of serial processing in PRP tasks could result from performance optimization rather than from a structural bottleneck. The analysis further suggests that the experimenter-selected distribution of stimulus onset asynchronies (SOAs) influences the relative efficiency of the serial and parallel modes, with a preponderance of short SOAs favoring a parallel mode. Experiments varying the distribution of SOAs were conducted, and the results suggest that there is a shift from a more serial mode to a more parallel mode as the likelihood of short SOAs increases.

Introduction

The “psychological refractory period” (PRP) paradigm has often been used to study the factors limiting cognitive performance in dual-task situations (e.g., Pashler, 1984, Telford, 1931, Welford, 1952, Welford, 1959). In the most typical versions of this paradigm, participants are asked to perform two separate choice reaction time (RT) tasks in each trial. The stimuli for the two tasks—S1 and S2—are presented in rapid succession, and participants are asked to respond to each as quickly as possible.

The PRP paradigm is popular partly because it provides experimenters with fine-grained control over the time interval separating the onsets of S1 and S2, an interval known as the “stimulus onset asynchrony” (SOA). When the SOA is relatively long, participants can simply perform the tasks one after the other, because processing of S1 can finish before S2 is presented. In this case, not surprisingly, the latency of the second response, RT2, is approximately the same as (or only slightly longer than) it would be if that task were performed in isolation. When the SOA is short, however, S1 is still being processed when S2 arrives, and participants must somehow cope with the demands of two simultaneous cognitive tasks. In this case performance generally slows dramatically (for a review see, e.g., Pashler, 1994a). In particular, RT2 increases substantially at short SOAs (Kahneman, 1973), and this increase is generally known as the “PRP effect”. Effects of SOA on RT1 are generally much smaller (e.g., Smith, 1969) and are sometimes essentially absent (e.g., Pashler & Johnston, 1989).

One attractive hypothesis about the cause of the PRP effect is the “response-selection bottleneck model” (Pashler, 1984, Pashler, 1994b, Welford, 1952, Welford, 1959). According to this model, one stage—called the bottleneck—is only capable of processing one task at a time. That is, this stage must process the tasks serially for some structural reason. When the second task needs access to the bottleneck stage while this stage is still busy processing the first task, the second task simply has to wait. Because such waiting time contributes directly to RT2, this model predicts that RT2 should decrease approximately linearly with slope −1 as SOA is increased. Although observed slopes relating RT2 to SOA are often shallower than this (Kahneman, 1973), the observed values are close enough to the predictions for many theorists to conclude that they support the bottleneck model (Pashler, 1994b).

There is still debate about the bottleneck model, however, because other models can also predict that RT2 should increase as SOA decreases, possibly even with a slope of approximately −1. For example, limited-capacity models are often discussed as alternatives to the bottleneck model (e.g., Kahneman, 1973, Navon and Gopher, 1979). The common feature of these models is that processing capacity can be shared between tasks in a graded fashion, with perhaps 70% of processing capacity allocated to one task and 30% to the other. Thus, capacity models are fundamentally different from the bottleneck model in that every stage is capable of processing two tasks in parallel—that is, there is no structural bottleneck.¹ Recent investigations indicate that some versions of these models can predict slopes of approximately −1 and can also accommodate other evidence previously cited as selectively supporting the bottleneck model (e.g., Navon and Miller, 2002, Tombu and Jolicœur, 2003). In addition, several other models allow the possibility of parallel processing, at least under some circumstances (e.g., Logan and Gordon, 2001, Meyer and Kieras, 1997a, Meyer and Kieras, 1997b, Navon, 1984). In general, such models seem more capable than bottleneck models of explaining observations that Task 1 responses may be affected by the nature of the response selection required for Task 2 (e.g., Hommel, 1998, Logan and Delheimer, 2001, Logan and Schulkind, 2000).

One reason that it has proved difficult to test experimentally between the bottleneck model and its competitors that allow parallel processing is that the latter models can closely mimic the bottleneck model (e.g., Meyer and Kieras, 1997b, Navon and Miller, 2002, Tombu and Jolicœur, 2003). To our knowledge, virtually all models that allow parallel processing also allow serial processing, so the fact that two tasks could be processed in parallel does not imply that they always would be.² For example, serial processing might be preferred because it is a natural way to bind together the separate sources of information relevant to each task (e.g., Logan & Gordon, 2001) or because it prevents crosstalk between tasks (e.g., Navon & Miller, 1987). The present article emphasizes another possibility: even if parallel processing were possible, people would be unlikely to use this mode if the serial mode were more efficient. Therefore, theorists should consider the possibility that serial processing leads to better performance than parallel processing before attributing such processing to structural limitations (i.e., a bottleneck).

In this article we focus primarily on the distinction between the bottleneck model, which requires serial processing in a certain stage, and other models that allow parallel processing in all stages. Although a number of studies have been conducted to see whether parallel processing takes place in paradigms designed to encourage it (e.g., Ruthruff et al., 2001, Tombu and Jolicœur, 2002), none of these studies have presented a theoretical framework that could be used to determine when the serial versus parallel processing modes would be optimal. Instead, in devising paradigms intended to encourage parallel processing, researchers have relied on intuitions and indirect evidence suggesting that parallel processing is more likely under some conditions than others—for example, with extensive practice (e.g., Hazeltine et al., 2002, Hirst et al., 1980, Schumacher et al., 2001; but for bottleneck-based accounts of practice effects, see Ruthruff et al., 2001, Ruthruff et al., 2003 ). Others, especially Meyer and Kieras (1999), have determined the conditions under which parallel processing would occur from specific models of processing (see also Logan and Gordon, 2001, Tombu and Jolicœur, 2002).

This article is based on a theoretical analysis of the conditions that determine whether parallel or serial processing is more efficient. In the first section, we present a metatheoretical model of dual-task performance that allows us to assess formally the optimality of serial and parallel processing modes under various circumstances. One surprising implication of this model is that serial processing is almost always more efficient than parallel processing. In light of this implication, repeated demonstrations of seriality do not seem theoretically decisive, because they could result from performance optimization rather than from a structural limitation.

Using our metatheoretical model, we develop an experimental manipulation that can be used to increase the benefit of parallel processing relative to serial processing. In the second section, we present a series of experiments examining the effects of this experimental manipulation on dual-task performance. In general, performance is sensitive to this manipulation in manner consistent with the idea that participants shift to a more parallel mode of processing when such a mode is more likely to be optimal. In short, the results weaken the claim of an immutable structural bottleneck, as do previous findings that at least some participants tend to shift processing modes in response to instructions emphasizing the use of parallel versus serial strategies (e.g., Schumacher et al., 2001).