Intelligence 41 (2013) 467–478

Contents lists available at SciVerse ScienceDirect

Intelligence

Task switching training effects are mediated byworking-memory management

Maayan Pereg⁎, Nitzan Shahar, Nachshon MeiranDepartment of Psychology and Zlotowski Center for Neuroscience, Ben-Gurion University of the Negev, Beer-Sheva, Israel

a r t i c l e i n f o

⁎ Corresponding author at: Department of Psychology,of the Negev, Beer-Sheva 84105, Israel.

E-mail address: maayanru@post.bgu.ac.il (M. Pere

0160-2896/$ – see front matter © 2013 Elsevier Inc. Ahttp://dx.doi.org/10.1016/j.intell.2013.06.009

a b s t r a c t

Article history:Received 17 January 2013Received in revised form 26 May 2013Accepted 11 June 2013Available online xxxx

Task switching is an important executive function, and finding ways to improve it has becomea major goal of contemporary scientists. Karbach and Kray (2009) found that training in theAlternating-Runs Task-Switching (AR-TS) paradigm (in which the task changed every secondtrial) reduced the costs of switching in untrained tasks, as well as led to far transfer tointerference control ability and fluid intelligence. However, AR-TS is known to involve workingmemory updating (WMU). Therefore, we hypothesized that AR-TS training involves WMU andnot task-switching proper. Participants were trained using Karbach and Kray's protocol.Results indicate a highly specific transfer pattern in which participants showed near transfer toswitching cost in the AR-TS paradigm, but did not significantly improve in another version ofthe task switching paradigm in which the tasks were randomly ordered or a version in whichthe task changed every 3rd trial. The results suggest that what has been trained is not a broadtask-switching ability but rather a specific skill related to the uniqueWMU requirements of thetraining paradigm.

© 2013 Elsevier Inc. All rights reserved.

Keywords:Cognitive trainingExecutive functionsWorking memoryTask switching

1. Introduction

Executive functions are cognitive abilities enabling goaldirected behavior. As such, they have broad relevance to issuessuch as general intelligence (Friedman et al., 2006), psychopa-thology (e.g., Kashdan&Rottenberg, 2010;Morgan& Lilienfeld,2000; Pennington & Ozonoff, 1996), psychological develop-ment (e.g., Garon, Bryson, & Smith, 2008; Zelazo, Carlson, &Kesek, 2008), and school performance (e.g., Diamond, Barnett,Thomas, & Munro, 2007). Knowing how to improve executivefunctions is therefore likely to have an enormous impact on awide array of psychological domains.

There is no clear consensus on the taxonomy of executivefunctions, andwhether they represent a single ability or a rangeof abilities (e.g., Baddeley, 1986, vs. Lehto, 1996). Nonetheless,many studies adopt Miyake et al.'s (2000) taxonomy, whichwas based on individual differences within the normal range.

Ben-GurionUniversity

g).

ll rights reserved.

According to Miyake et al., there are three executive functionsincluding updating and monitoring of working memory repre-sentations (WMU), inhibition of prepotent responses (inhibi-tion) and shifting between tasks ormental sets (task switching).

Several studies in the past few years demonstrated thattraining in a cognitive task tapping an executive function couldresult in far transfer to general intelligence (e.g., Jaeggi,Buschkuehl, Jonides, & Perrig, 2008; Klingberg et al., 2005;Schmiedek, Lövdén, & Lindenberger, 2010). By “far transfer”werefer to improvements seen in a structurally different task thanthe training task (that involves different content and taskrequirements, yet tapping similar critical psychological pro-cesses), as opposed to near transfer effects which relate tospecific attributes of the training task. The transfer is allegedlybased on the fact that the training program and the transfertasks have a common element through which the trainingoccurs. Showing far transfer of executive function training isespecially interesting in light of findings suggesting thatindividual differences in executive functions are mainly geneticin origin (Friedman et al., 2008). In line with the geneticfindings, there have been some recent reports showing failures

468 M. Pereg et al. / Intelligence 41 (2013) 467–478

to find beneficial outcomes of various training programs whichhave previously reported as successful (e.g., Redick et al., 2012;Chooi & Thompson, 2012; Shipstead, Redick, & Engle, 2012).However, replication was not our primary goal. It served asmeans to test a specific hypothesis related to principlesunderlying successful executive function training. Specifically,our hypothesis was that computerized training, at least ascurrently implemented, improves only one out of threedomains of executive functioning. This hypothesis is based onthe observation that while most of the published studiesshowing promising results trainedWMU, only very few studiesindicate promising results in the other two domains ofexecutive functions including task switching and inhibition.Actually, it seems that these two executive functions do not gainfrom computerized training, at least as currently implemented;although they may gain from other approaches includingextensive educational interventions, (e.g., Diamond et al.,2007) and meditation (Greenberg, Reiner, & Meiran, 2012,2013).

In the present work, we focused on task switching training.Before getting into the details of the training studies, it isimportant to introduce some key terms related to taskswitching. First, the task switching paradigm (e.g., Kiesel etal., 2010; Meiran, 2010; Monsell, 2003; Vandierendonck,Liefooghe, & Verbruggen, 2010, for review) yields two mainperformance cost estimates: switching cost, the differencebetween switch trials (in which the task has switched) andrepeat trials (in which the task repeats from the previous trial)within mixed-tasks blocks; and mixing cost, the differencebetween repeat trials (from the mixed-tasks block) and trialsfrom single-task blocks (in which only one task is executed).These costs represent the difficulty in switching and thereforeserve as a target for training. Second, the term “task switching”refers to several paradigms that yield somewhat differenteffects. Rogers and Monsell (1995) introduced the AlternatingRuns Task Switching paradigm (AR-TS), in which the tasksalternate between runs of fixed lengths (for example, Run-Length = 2 as in Karbach&Kray, 2009,means anAA–BB–AA…sequence, in which A and B represent the two tasks. That is, thetask changes every 2nd trial). Another paradigm is cued-TS, inwhich the tasks are randomly ordered, an external cue appearsjust before the target stimulus, and the participants areinstructed to perform the cued task (Shaffer, 1965). We willfocus on these two methods, though other paradigms are alsobeing used (for review, see Meiran, 2010). The differencebetween these twoparadigms is that there is a constant need tokeep track of the position in the run in WM in AR-TS, whereasthe cued TS paradigm does not involve such a requirement.

Themost influential task switching training study is Karbachand Kray's (2009), showing that TS training led to widespreadtransfer to switching cost, mixing cost, interference control,verbal and visualWM, and fluid intelligence in three age groups:children, young adults, and older adults. This study stands outpartly because there are only very few studies showing transfereffects of TS training. Actually, twounpublishedworks thatwereconducted well before Karbach and Kray published their paperindicate very limited transfer after task switching training.These include an unpublished Ph.D. dissertation from Gopher'slab (Armony-Shimoni, 2001) as well as an unpublished workfrom our lab (Sosna, 2001). In Armony-Shimoni's Ph.D. study,participants were trained on the randomized-runs paradigm

(Altmann & Gray, 2002, 2008; Gopher, Armony, & Greenshpan,2000) in which task-cues appeared at the beginning of runs oftrials varying in length between 4 and 12 trials. The resultsindicated some transfer of training effects across different kindsof stimuli (e.g., from letters to digits) or across differentcomputational operations as long as they belonged to thesamemodality, such as spatial processing (e.g., from comparingwhich one of two groups has more items to evaluating whethera group has more or less than five items). However, when theprocessing mode changed (e.g., from spatial to semantic) orwhen the judgment goals changed (e.g., from judging high-vs.-low to judging odd-vs.-even) no transfer of training wasfound. Sosna's Master's Thesis included 2 experiments in whichparticipants were trained in a cued-TS paradigm involving twospatial location tasks (up-down and right-left). In Experiment 1,there were three training sessions and switch probability variedbetween training groups. In Experiment 2 (6 training sessions),different versions of the training paradigm were used. In bothexperiments the switch costs were subjected to training effectsbut not to transfer effects. Importantly, in both of these studies,the training paradigm was not AR-TS, suggesting that perhapssome unique features of the AR-TS paradigm are responsible forKarbach and Kray's success. Only one study (Minear & Shah,2008) compared training and transfer effects in cued-TS andAR-TS and the results are inconsistent with the hypothesisabove since cued-TS training but not AR-TS training led to sometransferrable gains, which the authors attributed to theunexpected task switches in cued-TS. Thus, we conclude thatthe empirical picture is far from being clear at present.

In their study, Karbach and Kray (2009) trained participantsduring a sixweek period: During the firstweek, the participantsperformed pretest measurements (switching, inhibition, WMand fluid intelligence tasks); Afterwards, they went throughfour AR-TS training sessions, one perweek; and then came backfor posttestmeasurements in the sixthweek. Four experimentalgroups in each age group were tested: single-task training(control), switching training, switchingwith verbal self instruc-tion and switching with verbal self instruction and variability.The verbal self instruction strategy was incorporated in orderto facilitate the maintenance and selection of the tasks (asrequired in AR-TS). The variable training, in which the taskschanged between sessions, was incorporated in order tofacilitate generalization and thus transfer to new tasks. Ofgreatest interest in the present paper is the fact that, amongyoung adults, largest gains in switching and mixing costs wereseen when both self instruction and variable training wereincorporated. Zinke, Einert, Pfennig, and Kliegel (2012) partiallyreplicated Karbach and Kray's (2009) findings in adolescentsshowing mixing cost reduction, RT decrease in a 2-back task,and choice reaction RT decrease. On the other hand, theyneither showed switching cost reduction, nor gains in inhibitionmeasures.

We found the widespread transfer in Karbach and Kray's(2009) study to be surprising for some reasons. First, theirtraining protocol did not involve an adaptation of task difficulty.This feature stands out as other successful protocols such asJaeggi et al.'s (2008), involved continued adaptation of taskdifficulty, aimed at keeping a high level of difficulty throughoutthe training phase. Moreover, it has been previously claimedthat adaptive task difficultymay be a crucial factor in the successof training (Buitenweg, Murre, & Ridderinkhof, 2012; Shipstead

469M. Pereg et al. / Intelligence 41 (2013) 467–478

et al., 2012). In addition, we found the transfer to fluidintelligence to be unexpected, since Friedman et al. (2006)showed that the entire correlation between switching/inhibitionand fluid intelligence was mediated through WMU. Finally, thetraining dosage was rather minimal. It seems to be too low toenable very far transfer, especially given Jaeggi et al.'s findingthat far transfer depended on a high training dosage.

Of greatest relevance here is that Karbach and Kray (2009)argued that their AR-TS training task relates to at least twoaspects of training: task-set maintenance and selection andtask-set switching. The involvement of maintenance brings tomind the possibility that WM mechanisms were highlyinfluential in generating the transfer effects, and if so, Karbachand Kray's study does not violate the generalization that WMUbut not switching and inhibition can be improved by currentcomputerized training protocols. In fact, Zinke, Einert, Pfennig,and Kliegel's (2012) recent finding concerning improvement inthe 2-Back task seems to support this hypothesis.

All these considerations led us to replicate Karbach andKray's (2009) study, using their exact training program andstimuli with outcome measures that were chosen in order todifferentiate between WMU and switching involvement inthe transfer-of-training effects. Before getting into the detailsof our study it is important to first review the literaturesuggesting WM involvement in the AR-TS paradigm whichwas used as the training task in Karbach and Kray's study.

2. The involvement of WM in AR-TS

There are many differences between AR-TS and cued-TS(e.g. Altmann, 2007; Monsell, Sumner, & Waters, 2003; Tornay& Milán, 2001). Altmann argued that the switching cost that isfound in AR-TS involves two inseparable processes — aswitching cost similar to the switching cost observed in thecued-TS paradigm and a cost specific to the first trial of a run.This latter component is reminiscent of WMU since itpresumably involves changing the active goal in WM. In thecued-TS paradigm, switching cost does not reflect this updatingbecause the cost is computed as the difference in performancebetween trials involving a task switch and trials involving a taskrepetition. When randomly cued, both switch and repeat trialsinvolve goal updating, and thus the difference between themdoes not involve goal updating. Therefore, it might be difficultto interpret results from AR-TS training (with switching costbeing the training effect).

Evidence for WM involvement in AR-TS comes from severaladditional sources. For example, Baddeley, Chincotta, andAdlam(2001), and Bryck and Mayr (2005, see also Goschke, 2000; butsee Miyake, Emerson, Padilla, & Ahn, 2004; Saeki, Baddeley,Hitch, & Saito, 2013, for qualification) showed that AR-TSperformance was influenced by articulatory suppression butcued-TS was not. Based on these findings, Bryck and Mayrsuggest that a critical function in AR-TS is the endogenoussequencing of taskswithinWM.According to Rubinstein,Meyer,and Evans (2001), two distinct processing stages enable taskswitching: rule activation and (more relevantly) goal shifting.Goal shifting keeps track of current and future tasks, through anupdating process. In AR-TS, goal shifting is based on the retrievalof the next task's identity from memory. In cued-TS, the tasksequence is random and therefore the identity of the next taskcannot be retrieved frommemory and is thus based on the task

cue. Moreover, an interesting neuropsychological finding comesfrom Brown and Marsden's (1988) study in which participantsswitched between color naming and word reading of Stroopstimuli (colored color-words, such as the word RED printed ingreen). The results showed that Parkinson's disease patients(known to suffer fromWMdeficits) were impaired in AR-TS butnot in cued TS (in which there was no need to keep track of thetask sequence in WM). This neurological dissociation furthersupports the hypothesis that AR-TS has a considerable WMinvolvement. (For further findings on the involvement ofworking memory in the AR-TS paradigm see Liefooghe,Vandierendonck, Muyllaert, Verbruggen, & Vanneste, 2005;Saeki & Saito, 2004).

The final piece of evidence comes from Miyake et al. (2000)who observed a moderate, yet nontrivial correlation between aWMU factor and a task-switching factor. Importantly, two of thethree switching tasks used by Miyake et al. (2000) to indextask-switching ability incorporated an AR-TS protocol (Plus-Minus and Letter-Number) and one task was externally cued.The raw, zero-order correlations indicated that performance inthe AR-TS taskswas significantly correlatedwith performance inall three WMU tasks; but in contrast, cued-TS performance didnot significantly correlate with any of the WMU tasks, thussupporting the differential involvement of WMU in AR-TS.

3. The present study

To summarize, we hypothesize that Karbach and Kray's(2009) study is not an exception to the rule that computerizedtraining had so far failed to improve wide-ranged switchingabilities because the improvement is related to the WMUaspects inherent in the AR-TS paradigm. For this reason, wedecided to replicate the exact same training protocol used byKarbach and Kray but incorporate transfer tasks that enabled usto test our hypothesis. Thus, we included TS paradigms thatresembled or did not resemble the AR-TS paradigm (withRun-Length = 2) that was used during training. Moreover, weincluded aWMU taskwhich required holding inmind similar (2pieces) amount of information as the AR-TS training paradigm,or required holding in mind a different amount of information(1, 3 pieces). Choice reaction time (CRT) was also assessed intwo versions: one requiring considerableWMdemand (becauseof the need to keep in mind arbitrary Stimulus-Response [S-R]mapping), and the other not WM-demanding since the S-Rmapping was not arbitrary. Finally, we also examined whethertraining would transfer to inhibition measures (Stroop), espe-cially given the high individual differences correlation betweenTS and inhibition reported by Friedman and Miyake (2004).Again, we incorporated two versions of the Stroop paradigm, aversion with arbitrary S-R mapping and a version with anon-arbitrary S-R mapping in order to assess WM involvement.

Following this logic, we reasoned that if the trainingprogram used in Karbach and Kray (2009) does improve“pure switching ability”, then the improvement should be seenin transfer to all the other TS paradigms. If however, what isbeing trained is WMU, then transfer should be seen in all thetasks requiringWM including all WMUmeasures and CRT andinhibition measures with arbitrary mapping. Alternatively, ifwhat is being trained is the (narrower) ability to hold in mindthe identity of the most recent 2 tasks, then training shouldtransfer just to WMU that requires keeping in mind the recent

470 M. Pereg et al. / Intelligence 41 (2013) 467–478

2 pieces of information. Finally, if the transfer is very near, onlyimprovement in AR-TS with Run-Length = 2 is expected, butnot any other transfer effect.

4. Method

4.1. Participants

An adwas spread to all Ben-GurionUniversity of theNegev'sstudents via the internal organization system (after approval)calling undergraduate students to register to a cognitivetraining study. Our prerequisites were that the participantsshould have a Psychometric Entrance Test (PET) score betweenZ-scores 0 and 2 (relying on the normative sample score), nolearning disabilities and no uncorrected vision or hearingimpairments.

Sixty-six Ben-Gurion University of the Negev undergraduatestudents took part in the experiment (42 females, mean age =24.29, SD = 1.93). The self reported PET Z-scores werebetween 0.13 and 1.89 (mean Z-score = 1.52, SD = 0.38,based on the PET normative sample). One control participantdropped out prior to post-test and results of a total of 65participants were analyzed. The participants were paid for theirparticipation (25 NIS per hour, ~6.5 USD). In addition, theyreceived feedback regarding their improvement at the end ofthe posttest session, as partial compensation for participating inthe experiment.

4.2. Materials and procedure

The experiment consisted of two sessions for the controlgroup, and six sessions for the training group. The pretestmeasurement session took approximately 150 min, including a15 min break. Afterwards, the training group participated infour training sessions, 20–25 min each, one per week;whereasthe control group did not perform any training. During thesefour weeks, no contact wasmade with the control group, otherthan remainders regarding posttest to ensure participation inthat session. The posttest session took place in the 6th week ofthe experiment and was slightly shorter than the pretestsession, approximately 120 min, including a 15 min break.

While Karbach and Kray (2009) presented the transfer taskssessions as measurements (meaning that the participants werenot blind to the experimental assignment), we presented themas “long training”, and the training sessions as “short training”.The participants were told that they are enrolling in a study inwhich there are two groups: one going through two long-training sessions, separated in time and one in which thelong-trainings sessions were combined with a short-trainingsession (experimental and control groups, respectively). Wetold the participants thatwewanted to see how the two trainingmethods interact (the “short-training” and the “long training”).We did so becausewewanted to lead the participants to believethat they were going to improve andwanted to also conceal thenature of the pretest and posttest as tests of improvement. Theparticipants were told that they will be informed regardingwhich training they will receive only after the first “longtraining” session. Once the study ended, the participants werefully debriefed regarding the misleading recruitment and noneof them reported suspecting that the measurements were nottraining sessions.

Participants in the two groups were matched according totheir university entrance score, single-task and mixed-taskperformance at pretest andOperation Span scores, in this order(Unsworth, Heitz, Schrock, & Engle, 2005). The matchingprocedure was adapted from Karbach and Kray (2009), withtwo differences. First, we added the O-Span score, in order toequate the participants on WM, since our primary hypothesisconcerns WM. Second, the university entrance score was usedfor matching as well as for pre-screening in order to base thematching on a highly reliable measure and in addition, toreduce the within-group variance in cognitive abilities in orderto increase statistical power.

4.2.1. Training tasksThe training sessions were constructed according to Karbach

and Kray (2009), using their stimuli and procedure. Theexperimental group received the “switching + verbal self-instruction + training variability” protocol, in which youngadults showed the largest improvement in Karbach and Kray'sstudy. The participants were instructed to verbally pronouncethe upcoming relevant task, a strategy that is meant to help theparticipants to keep-track of the currently relevant task. Foursessions of 20–25 min were administered during four weeks oftraining (one per week), in which participants performed twonew, different, tasks in every session. Only mixed blocks wereinvolved in the training, and each session started with twopractice blocks followed by 24 experimental blocks (17 trialseach). Errors were followed by a 400 ms beep tone.

4.2.2. MeasurementsOperation Span (was only administered at pretest for the

purpose of matching. It was the exact task used by Unsworthet al., 2005): In this task, participants were requested toalternate between two tasks. The first task was to verify thecorrectness of an arithmetic operation and the second task wasto remember a letter. At the end of a series of such pairs of tasks,the participants were required to recall the letters in the sameorder inwhich theywere presented. Thiswas done bymeans ofmouse clicking on letters that appeared on the response screen.Actual testing was preceded by practicing the two tasks,separately and combined as well as by estimating the timeneeded to solve the arithmetic operation. This estimate wasthen used to limit the presentation of the arithmetic operationso as to prevent rehearsal. The test phase consisted of three setsof each set size (ranging between 3 and 7, randomly ordered).This made a total of 75 letters to recall. For further elaborationon the task's parameters see Unsworth et al., 2005.

4.2.2.1. Task switching. There were three different TS tasks,each involving different tasks, none of which were the sametasks used during training. All of them began with threepractice blocks (two single-task blocks and one mixed-tasksblock), followed by 18 experimental blocks (8 single-task and10 mixed-tasks blocks). Each block had 17 trials, and theparticipants were instructed at the beginning of each block asto whether it was a single-task or a mixed-tasks block.

4.2.2.1.1. AR-TS2 (Run-Length = 2). Following Karbachand Kray (2009), we used an AR-TS task to measure trainingoutcome, using their design, tasks and stimuli. Task A was a“size” task (indicating whether the target stimulus was smallor large), and task B was a “food” task (indicating whether

Fig. 1. The Choice Reaction task structure. Task A and Task B demonstrate thetwo tasks there were administered. On each task there were arbitrary andnon-arbitrary stimulus-response rules, and each of them was administeredin a two 2-choice versions (right-left or up-down) and in the full 4-choiceversion.

471M. Pereg et al. / Intelligence 41 (2013) 467–478

the target stimulus was a fruit or a vegetable). None of thesetasks was used during training, meaning that any influence oftraining would not be task-specific. Two types of blocks wereused: in the single-task blocks, only one of the two tasks wasperformed, and in the mixed-tasks blocks the runs of pairs oftrials of the two tasks alternated (i.e., AA–BB–AA), meaningthat the task switched on every 2nd trial. The stimuliconsisted of 18 vegetable and 18 fruit pictures; each had alarge and a small version. Each trial began with a 1400 msfixation, following the target that appeared until the responsewas made (pressing one of the two keys — “a” for leftresponses and “l” for right responses on a standard QWERTYkeyboard), and a 25 ms post target interval. As in the trainingsessions, errors were followed by a 400 ms beep tone.

4.2.2.1.2. AR-TS3 (with Run-Length = 3). The participantswere instructed to switch the tasks every three trials(e.g. AAA–BBB–AAA). This test was administered at the endof the AR-TS2 task at posttest only. This part consisted of onepractice block and six experimental blocks of the “size” and“food” tasks, same as the tasks used in AR-TS2. The rest of thetask's characteristics were the same as AT-TS2.

4.2.2.1.3. Cued-TS. In this task, the name of the task appearedon the screen before the target stimulus was presented andserved as the task-cue. The tasks switched randomly with a50:50 chance for switch and repetition. Task A was a “house”task (indication whether the target was a furniture or anelectrical appliance), and task B was a “location” task(indicating whether the target was located in the upper orlower half of the screen). The Stimuli consisted of 18illustrations of electrical appliances and 18 illustrations offurniture items (taken using Google™ searches), each couldappear either in the lower (y = 35%) or the upper (y = 65%)part of the screen. The 1400 ms fixation that was used in theAR-TS was split to a 900 ms fixation and a 500 ms cue at themiddle of the screen (which was the name of the task — i.e.“location” or “house”) followed by the target stimulus until aresponse was given.

Notice that in all of the versions of the TS paradigm, one taskwas semantic (i.e. “house” and “food”), and the other requiringlower-level perceptual analysis in nature (i.e. “location” and“size”). The response rules for the “semantic” tasks were fixed,and the response rules for the “perceptual” tasks werecounterbalanced between participants. For example, half ofthe participants received the right key as a response for “large”,and half received it for “small”, but all of them received theright key for fruit and the left for vegetable.

4.2.2.2. Far transfer tasks4.2.2.2.1. Inhibition. Two versions of the Stroop (1935) task

were applied. One of the four color-words (i.e. red, green,blue and yellow) or a neutral stimulus (i.e. a Hebrew parallelto XXXX) appeared randomly in different colors. Participantswere instructed to only regard to the color, while ignoringthe color-word. The tasks consisted of one practice block andthree experimental blocks, with 20 practice trials and 40trials in every experimental block. 20% of all trials wereneutral, 20% were congruent, and 60% were incongruent. A700 ms fixation separated between the targets and thetargets remained on the screen until a response was made.

Stroop interference was defined as the difference betweenincongruent (e.g. ‘red’ in blue ink) and neutral trials. We

employed a version with a low WM load, a vocal Stroop, inwhich participants said the ink color, a microphone recordedthe response times, and the responses were taped so thataccuracy could be scored offline. The high WM load versionwas a manual task, in which four keyboard keys werematchedto the four possible colors by stickers. The participants usedtwo fingers of each hand in order to respond in this task. Thisversion involves high WM load because of the need to hold inmind the color-to key assignment.

4.2.2.2.2. Working memory updating. An N-back task wasapplied, using N = 1, 2 and 3 (Owen, McMillan, Laird, &Bullmore, 2005). The task was adapted from Zinke et al.(2012) and consisted of 52 animal illustrations (Snodgrass &Vanderwart, 1980). The participants had to indicate by a keypress whether the animal is the same one that was presentedN trials ago. There were three blocks (one block per N-backlevel), such that Block 1 had N = 1, Block 2 had N = 2 andBlock 3 had N = 3. Each block had 124 trials (the first 4 trialswere taken out of the analysis) in which the illustrationswere presented for 1500 ms, followed by a 1000 ms interval.The participants had a 7-trial practice with feedback prior toeach N-back level. Target probability was set to 25%, and thedependent variable was accuracy level.

4.2.2.2.3. Choice reaction time (CRT). There were six CRTtasks that included three tasks with arbitrary S-R mapping(involving a highWM load) and three tasks with non-arbitrarymapping (see Fig. 1). Within each triplet, there were two taskswith two choices and one task with four choices (the samechoices that were used in the two-choice versions). The CRTtasks were never performed one after the other to prevent asmuch as possible negative transfer effects. Each of the six tasksconsisted of 4 blocks of 32 trials. Due to a programming error,the first 8 participants performed 40 trials in each block. Thetasks started with the 2-choice tasks, followed by the 4-choicetask, but the order in which they were performed (arbitrary vs.

300350400450500550600650700750800850900

Qua

rtile

1

Qua

rtile

2

Qua

rtile

3

Qua

rtile

4

Qua

rtile

1

Qua

rtile

2

Qua

rtile

3

Qua

rtile

4

Qua

rtile

1

Qua

rtile

2

Qua

rtile

3

Qua

rtile

4

Qua

rtile

1

Qua

rtile

2

Qua

rtile

3

Qua

rtile

4

Session1 Session2 Session3 Session4

RT

(m

s)

Repeat Switch

0

0.02

0.04

0 .06

0.08

0.1

0.12

0.14

0.16

Qua

rtile

1

Qua

rtile

2

Qua

rtile

3

Qua

rtile

4

Qua

rtile

1

Qua

rtile

2

Qua

rtile

3

Qua

rtile

4

Qua

rtile

1

Qua

rtile

2

Qua

rtile

3

Qua

rtile

4

Qua

rtile

1

Qua

rtile

2

Qua

rtile

3

Qua

rtile

4

Session1 Session2 Session3 Session4

PE

Repeat Switch

A

B

Fig. 2. Training sessions: Reaction Time (RT) in ms and Proportion of Errors(PE) as a function of Transition (switch vs. repeat), Training Session andSession Quartile. The results show a general reduction of RTs, and switchingcost within session but not between sessions.

472 M. Pereg et al. / Intelligence 41 (2013) 467–478

non-arbitrary) was counterbalanced between participants.Trials began with a fixation presented either for 250 ms or500 ms (randomly selected), followed by the target, whichwas presented until a response was made or until 6 s hadelapsed. Errors were followed by a beep for 400 ms.

4.2.2.3. Data analysis. In all of the tasks except for the N-backtask (which is based on accuracy) we analyzed mean reactiontime (RT) and proportion of errors (PE). RT values quickerthan 100 ms or slower than 3000 mswere excluded. Error andpost error trials were not analyzed, as well as the first trial inevery block and the practice blocks. Error rates were alsoanalyzed and all effects are showed both in RT and in errorproportions. Analyses of Variance (ANOVAs) were run on eachof the measurements. In the transfer tasks' results, only effectsinvolving Group are reported. With this sample size, thea-priori power to detect a medium effect size (ρ = .30) with a1-sided t-test (as required for replication) was just over .80given α = .05.

5. Results

5.1. Training results (n = 33)

RTs were analyzed in a three-way ANOVA with the within-subjects independent variables- Session (1–4), Transition(repeat — switch) and Quartile (1–4), referring to thesequential portion of a session, such that each session had 4quartiles, each lasting approximately 7 min (see Fig. 2). TheTransition variable enabled us to detect switching costs.Namely, switch trials were expected to produce poorerperformance relative to repeat trials. Thus, an interactionbetween Session and Transition would show that switchingcosts differed across sessions and the most important trendwould be for these costs to become smaller progressing fromSession 1 to Session 2, 3 and 4. Similarly, an interactionbetween Quartile and Transition would show that switchingcosts changed in the course of the session, again with the mostimportant trend showing that these costs reducedwith sessionprogression. The distinction between Session and Quartile isimportant since the tasks themselves differed between sessions.Thus, if switching costs would reduce from Session 1 to Session2, for example, this would already show some transfer oftraining (to new tasks). If, however, switching costswould onlyreduce within a session, this would imply that the training wastask-specific and not transferrable.

All three main effects were significant including Session[F(3,96) = 16.41, p b .001, MSE = 32,942.21, η2p = .34],Transition [F(1,32) = 95.41, p b .001, MSE = 16,916.02,η2p = .75], and Quartile [F(3,96) = 19.15,p b .001, MSE =4852.40, η2p = .37]. The two-way interactionswere significantaswell including Session by Quartile [F(9288) = 2.33, p = .01,MSE = 3304.43, η2p = .07], Session by Transition [F(3,96) =4.55, p b .01, MSE = 3492.85, η2p = .12], and Quartile byTransition [F(3,96) = 8.89, p b .01, MSE = 1779.92, η2p =.22]. We adopted Karbach and Kray's (2009) protocol inwhich a particular set of two tasks was used in each session.Admittedly, this design makes it difficult to interpret theSession effect since it is perfectly confounded with theparticular tasks. Their results indicate a speeding whencomparing Session 1 to Session 4. In order to compare our

resultswith theirs, we conducted a similar comparison and alsofound this speeding effect [F(1,32) = 5.57, p = .02, MSE =41851.45, η2p = 0.14].

The Session main effect as well as the Session by Transitioninteraction indicate a non-monotonous change in both RT andswitching costs thatwerehigher in Session2, possibly because ofthe specific attributes of the tasks (keep in mind that the tasksdiffered across sessions). The Quartile by Transition interactionindicates a trend for switching cost reduction within the courseof the session progression. A similar three-way ANOVA wasconducted on the PEs. Themain effect of Sessionwas found to besignificant [F(3,96) = 7.98, p b .001, MSE = 0.002, η2p = .20]as were the main effects of Transition [F(1,32) = 42.18,p b .001, MSE = 0.004, η2p = .57] (mean PE on repeat trialswas 0.039 and 0.065 on switch trials) and Quartile [F(3,96) =9.43, p b .001, MSE = 11.75, η2p = .23]. The interactionsbetween Session and Transition [F(3,96) =8.03, p b .001,MSE = 0.001 η2p = .20], and between Quartile and Transition[F(3,96) = 9.25, p = .02, MSE = 0.0009, η2p = .09] weresignificant as well, though they did not indicate a systematicreduction or increase in switching cost with session progressionas did the equivalent RT interactions. The fact that the systematic

550600650700750800850900

T (

ms)

A

473M. Pereg et al. / Intelligence 41 (2013) 467–478

decrease in RT-switching-cost during the session was notaccompanied by a systematic increase in PE-switching-costrules out speed-accuracy tradeoff as an account of the RTinteraction. Thus, we can conclude that our results are quitesimilar to those reported by Karbach and Kray (2009) (seeTable 1).

300350400450500

Tra

inin

g

Con

trol

Tra

inin

g

Con

trol

Tra

inin

g

Con

trol

Tra

inin

g

Con

trol

Pre Post Pre Post

Semantic Task Perceptual Task

R

Repeat Switch

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

Tra

inin

g

Con

trol

Tra

inin

g

Con

trol

Tra

inin

g

Con

trol

Tra

inin

g

Con

trol

Pre Post Pre Post

Semantic Task Perceptual Task

PE

B

5.2. Near transfer (n = 65)

5.2.1. AR-TS2 (run length = 2)We performed two sets of 4-way ANOVAs, one on

switching costs and one on mixing costs. The ANOVA designincluded the between subjects independent variable Group(training- control), and the within subjects independentvariables Measurement (pre-post), Task-Type (semantic,perceptual) and Transition (single-repeat for the mixingcost analyses, and repeat-switch for the switching costanalyses). As in the previous analysis, the Transition variableenabled us to detect performance costs. Switch trials wereexpected to produce poorer performance than repeat trials,and repeat trials were expected to produce poorer perfor-mance than single-task trials. The Task-Type variable showsthe difference between the two kinds of tasks, though nodifference between them was necessarily expected. Thepredicted result in which the training group shows lowercosts at posttest would reflect in an interaction betweenMeasurement, Group and Transition. This interaction couldbe modulated by the Task-Type variable, indicating differen-tial improvements as a function of task type.

No significant main effect for Group was found [F(1,63) =0.42, p = .52, MSE = 173,606.48, η2p = .007]. Other than

Repeat Switch

Fig. 3. AR-TS2: Reaction Time (RT in ms) and Proportion of Errors (PE) as afunction of Transition, Measurement, Group and Task-Type. A significantthree-way interaction between Group, Measurement and Transition wasfound in the perceptual task, such that the training group reduced itsswitching cost more strongly than the control group.

Table 1Effect sizes expressed in Cohen's D, for RT (and proportion of errors).

Traininggroup

Controlgroup

Switching cost reductionbetween Training Sessions

0.15 (−0.33) –

RT reduction betweenTraining Sessions

1.57 –

AR-TS2 SC reduction (pre vs. posttest) 0.52 (0.41) 0.70 (−0.1)AR-TS2 SC reduction(perceptual task pre vs. posttest)

0.93 (0.23) 0.33 (−0.25)

AR-TS2 MC reduction (pre vs. posttest) 0.77 (−0.16) 0.45 (−0.03)Cued-TS SC reduction (pre vs. posttest) 0.39 (0.02) 0.33 (−0.14)Cued-TS SC reduction(semantic task, pre vs. posttest)

0.25 (0.03) 0.23 (−0.16)

Manual Stroop interference reduction 0.31 (−0.18) 0.06 (0.1)Vocal Stroop interference reduction 0.51 (−0.43) 0.23 (−0.18)3-Back accuracy improvement(accuracy difference, pre vs. posttest)

0.40 0.22

CR 4-choice arbitrary RTpre-to-posttest improvement

1.01 (0.06) 0.88 (0.23)

CR4-choice non-arbitrary RTpre-to-posttest improvement

0.51 (−0.18) 0.84 (0.003)

CR 2-choice arbitrary RTpre-to-posttest improvement

0.74 (−0.06) 0.45 (−0.07)

CR 2-choice non-arbitrary RTpre-to-posttest improvement

0.59 (−0.25) 0.79 (−0.06)

Note. Effect sizes were calculated as (post-test − pre test)/pooled (pre-post)SD. Positive values are in the expected direction.

that, the four-way interaction was found to be significant[F(1,63) = 4.76, p = .03, MSE = 2619.64, η2p = .07] (Fig. 3).

The four-way interaction indicates that there was aswitching cost reduction from pretest to posttest in bothgroups but the reduction rate was influenced by Group andTask. This reduction was statistically comparable in 2 groupsin the semantic task, as indicated by a non-significantsimple triple interaction of Measurement Time, Transitionand Group [F(1,63) = 0.31. p = .58. MSE = 3869.21,η2p = .005]. However, the same simple triple interactionapproached significance in the perceptual task [F(1,63) =3.59, p = .06, MSE = 265,777.02, η2p = .05], showinglarger switching cost reduction in the training group ascompared with the control group (Cohen's D = .93 and .33in the two groups, respectively). In comparison, Karbachand Kray (2009) reported D = ~1.60 in the parallel youngadults' training group and D = ~.30 in their single-taskcontrol group. The marginally significant simple interactioncould have been due to poor statistical power. Indeed, using

300350400450500550

600650700750800

850900

Training Control Training Control

Semantic Task Perceptual Task

RT

(m

s)

Repeat Switch

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

Training Control Training Control

Semantic Task Perceptual Task

PE

Repeat Switch

A

B

Fig. 4. AR-TS3: Reaction Time (RT in ms) and Proportion of Errors (PE) as afunction of Transition, Measurement, Group and Task-Type. There was nosignificant advantage for the training group.

474 M. Pereg et al. / Intelligence 41 (2013) 467–478

an RT inverse transformation, as recommended by Ratcliff(1993) to increase the statistical power, resulted in asignificant simple triple interaction in the perceptual task[F(1, 63) = 4.76, p = .032, MSE = 0.02, η2p = .07]. Nosignificant differential group improvement was found formixing cost. This was true regardless of whether mixing costwas defined as we did (repeat vs. single-task) or as Karbachand Kray (2009) did (switch and repeat vs. single-task).

A similar four-way ANOVA was conducted on PEs. Asignificant triple interaction betweenMeasurement, Transitionand Group [F(1,63) = 4.91, p = .03, MSE = 0.002, η2p =.07] was found. This interaction shows that while bothgroups showed switching cost prior to training, a muchsmaller switching cost was found in the training group atposttest (Fig. 3). Thus, the PE results show training-relatedreduction in switching costs in both tasks. No otherinteractions involving Group were found to be significant.A similar ANOVA was calculated in order to estimate mixingcost reduction, though no significant effects involvingGroup and Measurement were found. Thus, we were ableto partially replicate Karbach and Kray's (2009) neartransfer effect in switching costs. Specifically, PE-switchingcosts were more strongly reduced in the training groupthan in the control group, while the parallel differentialimprovement in RT-switching-cost was restricted to theperceptual task for some reason. The fact that we could(albeit partially) replicate Karbach and Kray's results iscritical to the interpretation of the remaining results, whichpresumably suggest the underlying reasons for this transfereffect.

5.2.2. AR-TS3 (Run-Length = 3)A three-way ANOVA was conducted with the same

ANOVA design as above except for not including Measure-ment as an independent variable (since the task wasadministered at posttest only). In this analysis, an interac-tion between Group and Transition would indicate that thetraining group improved their overall performance in theAR-TS paradigm, above and beyond the particular runlength that was used during training. None of the effectsinvolving Group reached significance, meaning that thetraining group lost their advantage (in RT switching-costs)once the run length was different than the length they weretrained on (Fig. 4).

The same ANOVA was conducted on PEs, showing asignificant main effect of Group [F(1,63) = 8.17, p = .006,MSE = 0.007, η2p = .11], indicating that the control grouphad significantly more errors in general. None of the othereffects involving Group reached significance. The PE resultsshow that training transferred to PE in general, yet it did nottransfer to switching costs.

5.2.3. Cued TSWe conducted a four-way ANOVA with the same design

as that used for AR-TS2. As opposed to the previous analysis,Training-related improvement in this task would mean thatthe gain transferred to a structurally dissimilar paradigm.The four-way interaction was only marginally significant[F(1, 63) = 3.19, p = .08, MSE = 3420.8, η2p = .05](Fig. 5). Again, it seems there was a difference betweentask types, only this time the semantic task showed a

greater switching cost reduction in the training group.However, the simple interactions conducted on each tasktype, separately, were non-significant (p = .13 in thesemantic task [p = .14 in an ANOVA with inverse RT toimprove statistical power], and p = .47 in the perceptualtask [p = .67 for inverse RT]).

A parallel ANOVA on PEs revealed a significant interactionbetween Group, Transition and Task [F(1,63) = 4.41, p =.039, MSE = 0.001, η2p = .06]. No interaction involvingMeasurement was found to be significant, however, indicat-ing that the group differences were already found at pretestand could not have resulted from training.

Similar ANOVAs were conducted on mixing costs (Transi-tion representing repeat vs. single-task). No effects involvingGroup were found to be significant either for RT or PE.

In summary, we conclude that the only stable reduction inswitching or mixing costs found post training was in theswitching cost of the version of the AR-TS task (Run-Length =2), which was most similar to the AR-TS task used duringtraining.

300

350

400

450

500

550

600

650

700

750

800

850

900

Traininng Control Training Control

Pre Post

RT

(m

s)

Neutral Incongruent

10

20

30

40

50

60

70

80

90

Pre Post

Str

oo

p In

terf

eren

ce (

ms)

Training Control

A

B

Fig. 6. A. Key Stroop task: The trend showing greater reduction in Stroopinterference in the training group did not reach conventional significancelevels. B. Key Stroop task: As opposed to the training group, the controlgroup did not show a trend for improvement between pre and posttest.

300350400450500550600650700750800850900

Tra

inin

g

Con

trol

Tra

inin

g

Con

trol

Tra

inin

g

Con

trol

Tra

inin

g

Con

trol

Pre Post Pre Post

Semantic Task Perceptual Task

RT

(m

s)

Repeat Switch

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

Tra

inin

g

Con

trol

Tra

inin

g

Con

trol

Tra

inin

g

Con

trol

Tra

inin

g

Con

trol

Pre Post Pre Post

Semantic Task Perceptual Task

PE

Repeat Switch

A

B

Fig. 5. Cued-TS: Reaction Time (RT in ms) and Proportion of Errors (PE) as afunction of Transition, Measurement, Group and Task-Type. The groups didnot differ significantly in the reduction of switching costs or RT betweenpretest and posttest.

475M. Pereg et al. / Intelligence 41 (2013) 467–478

5.3. Far transfer (n = 65)

5.3.1. Manual StroopA three-way ANOVA was performed using the between-

subjects independent variable Group (training–control) andthe within-subjects independent variables Measurement(pre-post), and Congruence (incongruent–neutral). TheCongruence variable enabled us to detect the Stoop interfer-ence. Incongruent trials were expected to produce poorerperformance than neutral trials. Therefore, an interactionbetween Group, Measurement and Congruence could indi-cate that the training group (but not the control group)reduced their Stroop interference at posttest.

None of the effects involving Group were significant,though the triple interaction [F(1,63) = 2.49, p = .12,MSE = 1299.38, η2p = .038], indicated a trend in the hypoth-esized direction (Fig. 6). Therefore, we decided to look furtherinto this effect. Specifically, the trend for a reduction in Stroopinterference was marginally significant in the training group[F(1,63) = 3.85, p = .054, MSE = 1299.38, η2p = .06], but

not in the control group [F(1,63) = 0.8, p = .77, MSE =1299.38, η2p b .004] who actually showed a reversed trend.PE analysis revealed no significant effects, though no in-dications for speed accuracy tradeoff were found.

5.3.2. Vocal StroopA parallel ANOVA was conducted on the vocal Stroop RT

data. As opposed to the previous task, our hypothesis suggeststhat if the training is indeed WM-based then no interaction isexpected between Group, Measurement and Congruence sincethe vocal Stroop task (in which participants said the ink color)does not challenge WM given the fact that S-R rules need notbe retained. No significant effects involving Group reachedsignificance. Actually, both groups showed a trend for Stroopinterference reduction from pretest to posttest (Fig. 7). Thistrend reached significance in the training group [F(1,63) =8.30, p = .005, MSE = 674.83, η2p = .13]; but not in thecontrol group [F(1,63) = 1.79, p = .18, MSE = 674.83,

300

350

400

450

500

550

600

650

700

750

800

850

900

Traininng Control Training Control

Pre Post

RT

(m

s)

Neutral Incongruent

10

20

30

40

50

60

70

80

90

Pre Post

Str

oo

p In

terf

eren

ce (

ms)

Training Control

A

B

Fig. 7. A. Vocal Stroop task: Contrary to the manual Stroop task, both groupsshowed a trend for reduced Stroop interference at posttest in this task. B. VocalStroop task: Both groups showed a trend for reduced Stroop interference atposttest, significant only in the training group.

476 M. Pereg et al. / Intelligence 41 (2013) 467–478

η2p = .03]. An analogous analysis of PE revealed that none ofthe effects involving Group reached significance.

In summary, only the manual Stroop performanceshowed a marginally significant training-related perfor-mance improvement.

5.3.3. CRTThe CRT tasks were divided according to whether the S-R

mapping was arbitrary and whether the task involved 2 or 4choices. A four-way ANOVA was calculated with the between-subjects independent variable Group, and three within-subjects independent variables: Measurement (pretest vs.posttest), Arbitrariness (arbitrary-non arbitrary) and Numberof Choices (2–4). The Arbitrariness variable enabled us todetect differences between a high WM-demanding task(i.e., arbitrary condition) and low WM-demanding task(i.e., non-arbitrary condition). Poorer performance wasexpected in the arbitrary condition. The Number of Choicesvariable refers to WM load, such that 4-choices are harder to

keep in mind than 2-choices and were therefore expected toproduce poorer performance. In all but one condition, therewas a relatively moderate pre-to-posttest improvement. Theonly exception was the 4-choice arbitrary task (presumablyinvolving the highest WM demand) which showed markedimprovement. However, none of the effects involving Groupapproached significance. PE analyses showed similar patterns,with no significant interactions involving Group.

5.3.4. N-backA three-wayANOVAon accuracy levels was conductedwith

the between-subjects independent variable Group, and the twowithin-subjects variablesMeasurement (pre-post) andN-Level(1,2,3). The N-Level variable enabled us to detect difficultyeffects. Since the training was based on a Run-Length = 2, aninteractionwas expected between the three variables, showingimproved performance in the training group's posttest perfor-mance for N = 2, as opposed to the control group.

In this task, there are two ways to calculate the accuracylevel — with or without considering “empty trials”, in whichthe participant did not respond at all (presumably because theydid not know whether the target was shown or not N trialsbeforehand). In addition, two outlier participants were foundat pre test (one from the training group and one from thecontrol group), showing no responses at N = 3 level, probablymeaning that they simply gave up when N was 3. Weconducted four analyses (with and without the empty trials,and with and without the outliers), and we report the analysisin which the empty trials were considered as errors andwithout the outlier participants. Since the results were similarin all of the analyses, we will only report the discrepanciesbetween them, when found.

The triple interaction was not significant [F(2, 122) =1.53, p = .22, MSE = 13.94, η2p = .024]. Despite thenon-significant triple interaction, inspection of Fig. 8 suggeststhat the most pronounced difference between the groupswas found at N = 3, in which the training, but not thecontrol group, showed a higher accuracy gain from pretest toposttest. We conducted two planned comparisons, for N = 2and for N = 3. For N = 2, the simple interaction betweenMeasurement and Group was not significant [F(1,61) b 1], asboth groups showed a significant trend for improvement. Thesimple-simple effects (testing the pretest to posttest im-provement in each group for a given N-level) were significantat this level for both groups ([F(1,61) = 22.27, p b .001,MSE = 11.68, η2p = .34] in the training group, and[F(1,61) = 31.26, p b .001, MSE = 11.68, η2p = .30] in thecontrol group). The simple two-way interaction betweenMeasurement and Group was not significant for N = 3 aswell [F(1,61) b 1], but unlike at N = 2 level, the simple-simple effect was significant in the training group [F(1,61) =6.46, p = .01, MSE = 21.04, η2p = .095], though clearly non-significant in the control group [F(1,61) = 1.31, p = .26,MSE = 21.04, η2p = .02].

The critical triple interaction reached significance only inone of the four analyses, the one in which we included theoutlier participants and excluded the empty trials [F(2126) =3.08, p = .049, MSE = 39.61, η2p = .05]. However, the trendfor pretest-to-posttest improvement seen in the training groupfor N = 3 was only marginal in this analysis (p = .09).

70

75

80

85

90

95

100

N1 N2 N3 N1 N2 N3

Pre Post

Ave

rag

e A

ccu

racy

(%

)

Training Control

Fig. 8. N-back performance (% correct) as a function of Group, N-Level, andMeasurement without empty trials and without outliers: A significantsimple-simple effect was found in the 3-back condition for the training (butnot the control) group.

477M. Pereg et al. / Intelligence 41 (2013) 467–478

6. Discussion

Executive functions are cognitive abilities that are rele-vant to many domains of psychological functioning includinggeneral fluid intelligence (e.g., Friedman et al., 2006).Recently, Karbach and Kray (2009) showed that trainingusing the AR-TS paradigm resulted in a widespread improve-ment in executive functioning and fluid intelligence. Theirresults are inconsistent with a number of studies, includingMinear and Shah (2008). Most importantly, Karbach andKray's results seem to violate our hypothesis that WMU andnot the other two executive functions in Miyake et al.'s(2000) taxonomy benefits from computerized trainingprotocols as currently designed. The goal of this study wastherefore to test the hypothesis that training effects in theAR-TS paradigm reflect WMU improvement rather than animprovement in task-switching proper. We hypothesizedthat AR-TS training, due to the high WMU involvement inthis particular TS paradigm, would only lead to posttestimprovement in tasks involving similar WM managementprocesses. The study included two groups, one was trained inthe AR-TS paradigm following Karbach and Kray (2009), andthe other was a silent control group. We reasonablyreplicated Karbach and Kray's (2009) training and neartransfer effects. Specifically, we found differential trainingrelated switching-cost reduction. In itself, this result isreassuring since Minear and Shah (2008) did not observetransfer effects after AR-TS training. The other transfer resultsin our study show that training did not lead to a significantimprovement in a general ability. Rather, the transfer effectsindicate that the improvement was related to the specificattributes of the AR-TS training task. Training effects did notsignificantly transfer to cued-TS, and significantly transferredto AR-TS only when the testing task had the same run lengthas in training. It should nonetheless be emphasized that eventhe transfer effects seen with Run-Length = 2 should not bedismissed since they were observed with tasks that were nottrained, only tested.

In other domains, we noticed an inconsistent pattern ofimprovement that prevents us from reaching any clearconclusion. Specifically, we hypothesized that if WM is thecore function in AR-TS, transfer would show in manual, butnot in a vocal Stroop task, since in the manual task, one has tohold in mind the stimulus-response rules. Unfortunately,none of the trends reached significance. Accordingly, the CRTresults did not indicate any differential training relatedimprovement even when they were most demanding interms of WM.

Thus, we conclude that although the AR-TS paradigminvolves WMU, it is not very demanding in this respect.Authors who argue that broad WMU skills can be trained(especially Jaeggi et al., 2008) also argue that for this tohappen, WMU demands must be kept high throughouttraining. However, in the current AR-TS paradigm, WMUrequirements were not very high and did not increase withtraining progression. This may explain the fact that somevery specific WMU skills seemed to have been trained in thecurrent study: those related to changing tasks every 2nd trial,but broader WMU skills were not trained.

During a recent trend for replicating training studies, somestudies did not obtain the original outcomes. For instance, bothRedick et al. (2012) and Chooi and Thompson (2012) failed inreplicating Jaeggi et al.'s (2008) fluid intelligence improvementoutcomes. In addition, a recently published paper also failed toachieve far transfer benefits using a training protocol similar toKarbach and Kray's (von Bastian & Oberauer, 2013). Ourcurrent results join this trend in a sense. We were able tosuccessfully replicate some of the near transfer effects found byKarbach and Kray (2009) but do not indicate any far transfereffects. Thus, our findings imply that Karbach andKray's (2009)training protocol is perhaps not generally beneficial as itseemed from the original study. Most importantly, we suggestthat some of the extant computerized training protocolsmore-or-less successfully improve WMU (but even this isdebatable, see Chooi & Thompson; Redick et al.). However, thiscannot be said with any certainty concerning switching andinhibition abilities. Finding principles for successful training ofthese executive functions therefore remains an outstandingchallenge.

Two important limitations of the present study should beacknowledged. First, our near transfer effects were smaller(by about a third) from those in the original study, a factwhich might explain the lack of far transfer effects. Second,most of the measurements were administered both at pretestand at posttest and this is true for all the measurements inwhich we found training-related benefits. Thus, we cannotrule out the possibility that training has led to an improve-ment in the ability to gain from the pretest experience. Inother words, the gain is seen only if one has been pre-tested.This implies that what has been trained is not a generalability, but a task-specific skill. Future research should thusemploy (also) post-test-only conditions that are required inorder to rule out the aforementioned alternative account.

Acknowledgments

We would like to acknowledge Julia Karbach and JuttaKray for providing us their materials, and also Yifat Daniely,Ella Sharon and Tal Yatziv for their help in data collection.

478 M. Pereg et al. / Intelligence 41 (2013) 467–478

References

Altmann, E. M. (2007). Comparing switch costs: Alternating runs and explicitcuing. Journal of Experimental Psychology: Learning, Memory, andCognition, 33, 475–483.

Altmann, E. M., & Gray, W. D. (2002). Forgetting to remember: Thefunctional relationship of decay and interference. Psychological Science,13, 27–33.

Altmann, E. M., & Gray, W. D. (2008). An integrated model of cognitivecontrol in task switching. Psychological Review, 115, 602–639.

Armony-Shimoni, L. (2001). Generality of control processes in task switching.Ph.D. Thesis. Haifa, Israel: The Technion.

Baddeley, A. D. (1986). Working memory. New York: Oxford Univ. Press.Baddeley, A., Chincotta, D., & Adlam, A. (2001). Working memory and the

control of action: Evidence from task switching. Journal of ExperimentalPsychology: General, 130, 641–657.

Brown, R. G., & Marsden, C. D. (1988). Internal versus external cues and thecontrol of attention in Parkinson's disease. Brain, 111, 323–345.

Bryck, R. L., & Mayr, U. (2005). On the role of verbalization during task setselection: Switching or serial order control? Memory & Cognition, 33,611–623.

Buitenweg, J. I., Murre, J. M., & Ridderinkhof, K. R. (2012). Brain training inprogress: A review of trainability in healthy seniors. Frontiers in HumanNeuroscience, 6, 183. http://dx.doi.org/10.3389/fnhum.2012.00183.

Chooi, W. T., & Thompson, L. A. (2012). Working memory training does notimprove intelligence in healthy young adults. Intelligence, 40, 531–542.

Diamond, A., Barnett, W. S., Thomas, J., & Munro, S. (2007). Preschoolprogram improves cognitive control. Science, 318, 1387–1388.

Friedman, N. P., & Miyake, A. (2004). The relations among inhibition andinterference control functions: A latent-variable analysis. Journal ofExperimental Psychology: General, 133, 101–135.

Friedman, N. P., Miyake, A., Corley, R. P., Young, S. E., DeFries, J. C., & Hewitt, J.K. (2006). Not all executive functions are related to intelligence.Psychological Science, 17, 172–179.

Friedman, N. P., Miyake, A., Young, S. E., DeFries, J. C., Corley, R. P., & Hewitt, J.K. (2008). Individual differences in executive functions are almostentirely genetic in origin. Journal of Experimental Psychology: General,137, 201–225.

Garon, N., Bryson, S. E., & Smith, I. M. (2008). Executive function inpreschoolers: A review using an integrative framework. PsychologicalBulletin, 134, 31–60.

Gopher, D., Armony, L., & Greenshpan, Y. (2000). Switching tasks andattention policies. Journal of Experimental Psychology: General, 129,308–339.

Goschke, T. (2000). Intentional reconfiguration and involuntary persistencein task-set switching. In S. Monsell, & J. Driver (Eds.), Attention andperformance XVIII: Control of cognitive processes (pp. 331–355). Cambridge,MA: MIT Press.

Greenberg, J., Reiner, K., & Meiran, N. (2012). “Mind the trap”: Mindfulnesspractice reduces cognitive rigidity. PLoS One, 7, e36206. http://dx.doi.org/10.1371/journal.pone.0036206.

Greenberg, J., Reiner, K., & Meiran, N. (2013). “Off with the old”: mindfulnesspractice improves backward inhibition. Frontiers in Psychology, 3, 618.http://dx.doi.org/10.3389/fpsyg.2012.00618.

Jaeggi, S. M., Buschkuehl, M., Jonides, J., & Perrig, W. J. (2008). Improvingfluid intelligence with training on working memory. Proceedings of theNational Academy of Sciences, 105, 6829–6833.

Karbach, J., & Kray, J. (2009). How useful is executive control training? Agedifferences in near and far transfer of task‐switching training. Develop-mental Science, 12, 978–990.

Kashdan, T. B., & Rottenberg, J. (2010). Psychological flexibility as afundamental aspect of health. Clinical Psychology Review, 30, 865–878.

Kiesel, A., Steinhauser, M., Wendt, M., Falkenstein, M., Jost, K., Philipp, A. M.,et al. (2010). Control and interference in task switching: A review.Psychological Bulletin, 136, 849–874.

Klingberg, T., Fernell, E., Olesen, P. J., Johnson, M., Gustafsson, P., Dahlström,K., et al. (2005). Computerized training of working memory in childrenwith ADHD-a randomized, controlled trial. Journal of the AmericanAcademy of Child and Adolescent Psychiatry, 44, 177–186.

Lehto, J. (1996). Are executive function tests dependent on working memorycapacity? Quarterly Journal of Experimental Psychology, 49, 29–50.

Liefooghe, B., Vandierendonck, A., Muyllaert, I., Verbruggen, F., & Vanneste, S.(2005). The phonological loop in task alternation and task repetition.Memory, 13, 650–660.

Meiran, N. (2010). Task switching: Mechanisms underlying rigid vs.flexible self control. In R. Hassin, K. Ochsner, & Y. Trope (Eds.), Selfcontrol in society, mind and brain (pp. 202–220). NY: Oxford UniversityPress.

Minear, M., & Shah, P. (2008). Training and transfer effects in task switching.Memory & Cognition, 36, 1470–1483.

Miyake, A., Emerson, M. J., Padilla, F., & Ahn, J. C. (2004). Inner speech as aretrieval aid for task goals: The effects of cue type and articulatorysuppression in the random task cuing paradigm. Acta Psychologica, 115,123–142.

Miyake, A., Friedman, N. P., Emerson, M. J., Witzki, A. H., Howerter, A., &Wager, T. D. (2000). The unity and diversity of executive functions andtheir contributions to complex “frontal lobe” tasks: A latent variableanalysis. Cognitive Psychology, 41, 49–100.

Monsell, S. (2003). Task Switching. Trends in Cognitive Sciences, 7, 134–140.Monsell, S., Sumner, P., & Waters, H. (2003). Task-set reconfiguration with

predictable and unpredictable task switches. Memory & Cognition, 31,327–342.

Morgan, A. B., & Lilienfeld, S. O. (2000). A meta-analytic review of therelation between antisocial behavior and neuropsychological measuresof executive function. Clinical Psychology Review, 20, 113–136.

Owen, A. M., McMillan, K. M., Laird, A. R., & Bullmore, E. (2005). N-backworking memory paradigm: A meta-analysis of normative functionalneuroimaging studies. Human Brain Mapping, 25, 46–59.

Pennington, B. F., & Ozonoff, S. (1996). Executive functions and developmentalpsychopathology. Journal of Child Psychology and Psychiatry, 37, 51–87.

Ratcliff, R. (1993). Methods for dealing with reaction time outliers.Psychological Bulletin, 114, 510–532.

Redick, T. S., Shipstead, Z., Harrison, T. L., Hicks, K. L., Fried, D. E., Hambrick, D. Z.,et al. (2012). No evidence of intelligence improvement after workingmemory training: A randomized, placebo-controlled study. Journal ofExperimental Psychology: General, 1–21. http://dx.doi.org/10.1037/a0029082(Advance online publication).

Rogers, R. D., & Monsell, S. (1995). The cost of a predictable switch betweensimple cognitive tasks. Journal of Experimental Psychology: General, 124,207–231.

Rubinstein, J. S., Meyer, D. E., & Evans, J. E. (2001). Executive control ofcognitive processes in task switching. Journal of Experimental Psychology:Human Perception and Performance, 27, 763–797.

Saeki, E., Baddeley, A. D., Hitch, G. J., & Saito, S. (2013). Breaking a habit: Afurther role of the phonological loop in action control. online first.Memory & Cognition. http://dx.doi.org/10.3758/s13421-013-0320-y.

Saeki, E., & Saito, S. (2004). The role of the phonological loop in taskswitching performance: The effect of articulatory suppression in thealternating runs paradigm. Psychologia, 47, 35–43.

Schmiedek, F., Lövdén, M., & Lindenberger, U. (2010). Hundred days ofcognitive training enhance broad cognitive abilities in adulthood:Findings from the COGITO study. Frontiers in Aging Neuroscience 2, 27,1–10. http://dx.doi.org/10.3389/fnagi.2010.00027.

Shaffer, L. H. (1965). Choice reaction with variable S-R mapping. Journal ofExperimental Psychology, 70, 284–288.

Shipstead, Z., Redick, T. S., & Engle, R. W. (2012). Is working memory trainingeffective? Psychological Bulletin, 1–27. http://dx.doi.org/10.1037/a0027473(Advance online, publication).

Snodgrass, J. G., & Vanderwart, M. (1980). A standardized set of 260 pictures:Norms for name agreement, image agreement, familiarity, and visualcomplexity. Journal of Experimental Psychology: Human Learning andMemory, 6, 174–215.

Sosna, G. (2001). Practice and transfer in task-switching: the effect of thequantity and frequency of switching. MA Thesis, Beer-Sheva, Israel:Ben-Gurion University of the Negev.

Stroop, J. R. (1935). Studies of interference in serial verbal reactions. Journalof Experimental Psychology, 18, 643–662.

Tornay, F. J., & Milán, E. G. (2001). A more complete task-set reconfigurationin random than in predictable task switch. The Quarterly Journal ofExperimental Psychology, 54A, 785–803.

Unsworth, N., Heitz, R. P., Schrock, J. C., & Engle, R. W. (2005). An automatedversion of the operation span task. Behavior Research Methods, 37,498–505.

Vandierendonck, A., Liefooghe, B., & Verbruggen, F. (2010). Task switching:Interplay of reconfiguration and interference. Psychological Bulletin, 136,601–626.

von Bastian, C. C., & Oberauer, K. (2013). Distinct transfer effects of trainingdifferent facets of working memory capacity. Journal of Memory andLanguage, 69, 36–58.

Zelazo, P. D., Carlson, S. M., & Kesek, A. (2008). The development of executivefunction in childhood. In C. A. Nelson, & M. Luciana (Eds.), Handbook ofdevelopmental cognitive neuroscience (pp. 553–574). Cambridge, MA, US:MIT Press.

Zinke, K., Einert, M., Pfennig, L., & Kliegel, M. (2012). Plasticity of executivecontrol through task switching training in adolescents. Frontiers in HumanNeuroscience, 6, 41. http://dx.doi.org/10.3389/fbhum.2012.00041.