Supplementary Information
Cognitive Priming and Cognitive Training:
Immediate and Far Transfer to Academic Skills inChildren
Bruce E. Wexler1, Markus Iseli2, Seth Leon2, William Zaggle, Cynthia Rush3, Annette Goodman, A., Esat Imal1, Emily Bo2
1Department of Psychiatry Yale University School of Medicine, 2National Center for Research on Evaluation, Standards, and Student Testing, CRESST / UCLA,3Department of Statistics, Yale University
SREP-16-11950-T
1SupplementaryInformation
2
3SupplementaryData1
Pearson Math Assessment First grade,95% Free Lunch
120
100
80
60
40
20
0
Pro@icientAtRiskBelowPro@icient
4
5
120
Pearson Reading Assessment 3rd grade
100
80
60
40
20
0
Pro@icientAtRiskBelowPro@icient
6
7
120
Pearson Math Achievement 2nd grade
ClassRecievedSpecialReadingProgramClassRecievedActivate
100
80
60
40
20
0
FallWinterSpringFallWinterSpring
Pro@icientAtRiskBelowPro@icient
8
9Inthisschooloneclasswasgivenaspecialreadingenhancementprogramandthe
10otherwasgivenourbrain-‐trainingprogram.Theclassthatdidthebraintraining
11hadnon-‐significantlygreatergainsinreadingthandidtheclassthatgotthespecial
12readingprogram.Shownhere,theclassthatgotourbraintrainingshowedmuch
13greatergainsinmath.N.B.,thisiscomparisontoanactivecontrol. 14
15
16
17
18
19
20
21
22
23
24
25
26
27
28SupplementaryMethods1
29
30CCReadingGameResponseOutcomes
31WeuseSamejima’sgradedresponsemodel1.InIRTgradedresponsemodelsthe
32responsevariablesmustbecodedinsuchawaythateachitemhasaresponsethatfalls
33alonganorderedscalewitheachvalueonthescalebeingpresentforeachitem.Inthis
34studywetreatthelevelsofCCgame-playasiftheywereitemsonanassessment.For
35easeofinterpretationwekeepthenumberoforderedresponsecategorieswithineachof
36thetwoCCgamesconsistentforalllevels/items.Thedegreetowhichthevaluesofthe
37performanceconstructsvaryacrosslevelsisdifferentintheCCmathgameascompared
38totheCCreadinggame.Asaconsequence,therecodingofthecontinuousperformance
39constructsrequiresdifferingapproachesthataredescribedforeachgameinthe
40supplementalmaterials.
41
42IntheCCreadinggametherewere31levelsthroughwhichthestudentsplayed.Students
43receivedacontinuousscoreoneachofthethreeperformanceconstructs(speed,accuracy,
44andcombinedspeedandaccuracy).Thecontinuousconstructswerethenbinnedinto
45quintilesbasedonperformanceacrossallthelevels.Thiscategorizationresultedinscore
46coverageforall5categoriesineachofthe31gamelevels(seeTable1).Settingthesame
47cutpointsforall31itemshasthebenefitofeasingthecomparativeinterpretationofthe
48IRTlevelparameterswithrespecttoleveldifficultysincethesamescoringrubricapplies
49ateachlevel.
50
51
52Table1.Cut-PointsforCCReadingGameResponseVariables
Speed-CorrectMovesPerMinute / Accuracy-% CorrectMoves / StandardizedSpeed andAccuracyAllLevels
Cutpoint1 / <4.34 / <42% / <-0.69
Cutpoint2 / 4.34to 6.69 / 42%to 54% / -0.69to -0.22
Cutpoint3 / 6.70to 8.45 / 55%to 67% / -0.22to 0.19
Cutpoint4 / 8.45to 10.43 / 67%to 79% / 0.19to 0.66
Cutpoint5 / >10.43 / >79% / >0.66
53CCMathGameResponseOutcomes
54IntheCCmathgametherewere121levels.Themaximumnumberoflevelsengagedby
55anystudentwas115levels.InordertohaveadequatesamplesizecoverageforIRT
56analysisitwasnecessarytosubstantiallyreducethenumberoflevelsbycollapsinglevels
57withsimilargame-playcharacteristics.Forexampletheoriginalgamelevelsone,three,
58fourandfiveallsharedthesamegame-playconfigurationandproblemtype(addition
59problemwithasumlowerthanfive).Thesefouroriginallevelswererecodedintoa
60singlecollapsedlevel.Followingthisprocesswewereabletoreducetheoriginallevels
61toaworkablenumberoftwenty-fourcollapsedlevels.
62
63Intheprocessofattemptingtocreatebinnedquintilesfromthetwenty-fourcollapsed
64levelstherestillremainedunfilledcells.Thislackofadequatesamplecoveragesuggested
65thatthegame-playperformanceintheCCmathgamedecreasedmoresteeplyacross
66gamelevelsascomparedtotheCCreadinggame.Duetothisperformancedecreaseand
67inordertoensureadequatesamplecoverageacrossalltwenty-fourcollapsedlevelsitwas
68necessarytoemploythreeorderedresponsecategoriesintheCCmathgamein
69comparisontothefivecategoriesthatwereusedfortheCCreadinggame.
70
71Therewasalsoasharpdecreaseinperformancebeginningwithcollapsedlevelnumber
72elevenwhichhadthegame-playcharacteristicofaddingtentomultiplesoftenthrough
73ninety.Asaresultitwasnecessarytosetseparatecutpointsforcollapsedlevelsone
74through10andeleventhroughtwentyfour.Theresultingcutpointsforthethree
75responseconstructsareshowninTable2.InterpretationoftheIRTleveldifficulty
76parametersshouldtakethesecutpointsintoaccount.
77
78Table2.Cut-PointsforCCMathGameResponseVariables
Speed-CorrectMovesPerMinute / Accuracy-% CorrectMoves / StandardizedSpeed andAccuracyLevels1-10
Cutpoint1 / <3.44 / <31% / <0.15
Cutpoint2 / 3.44to 5.51 / 31%to 48% / 0.15to 1.07
Cutpoint3 / >5.52 / >48% / >1.07
Levels11-24
Cutpoint1 / <1.92 / <10% / <-0.75
Cutpoint2 / 1.92to 2.99 / 10%to 16% / -0.75to -0.36
Cutpoint3 / >2.99 / >16% / >-0.36
79
80Thestructureofourdatasetisonethatiscommonlyseeninmulti-levelgrowthmodels.
81Eachstudentparticipatedinmultiplegame-playsessionsresultinginbasiclevels(or
82dimensions)inthedataset.ModellevelIrepresentstherepeatedsessionsthatarenested
83withinstudentsthuscapturingthewithin-studentvariation,andmodellevelIIcaptures
84thebetween-studentvariation.WeusetheflexMIRTsoftware2,whichimplementsthe
85Metropolis-HastingsRobbins-Monro(MH-RM)algorithm3-5whichallowsforthe
86estimationofhigherdimensionalmodelsandwhichprovidesthecapacitytocombine
87IRTwithmulti-levelmodelsthatincludecovariates.Inadditiontogeneratinggame-play
88levelparameters,theIRTmodelalsogeneratesalatentfactorforeachlevelofthemodel.
89Wefixthevarianceofourbetween-studentfactortooneandallowthewithin-student
90factortovaryfreely.Wealsoincludeeightcovariatesinourmodel,sevenofwhich
91predictthewithin-student(modellevelI)factorofourmodelandonewhichpredictsthe
92between-student(modellevelII)factor.
93
94TheprimaryresearchquestioninvolvestheimmediateeffectoffourBTgamesonthe
95within-studentCCgameperformancevariation.Thesefourtreatment-relatedcovariates
96arecodedasdummyvariablesandtheconditionwithnoBTgamesservesasareference
97group.Thebetween-studentlatentvariablevariancewasfixedatoneandhasameanof
980,sotheestimateoftheBTgamedummyvariableisintheformatofaneffectsize.Three
99othercovariatesareincludedinthemodeltopredictthewithin-studentfactor(model
100
levelI):1)asimplelinearindicatorofthesessionday,whichrepresentsthecurrent
101
sessiondayforastudent;2)thelengthoftimeofeachsession;and3),avariablethat
102
indicatesifthestudenthadpreviouslyexperiencedanyofthelevelstheyengagedina
103
currentsession.Finally,wealsoincludegenderasacovariatetopredictthebetween-
104
studentfactor(modellevelII).
105
1.Kim,S.Camilli,G.Anitemresponsetheoryapproachtolongitudinalanalysis
106
withapplicationtosummersetbackinpreschoollanguage/literacy.Large-scale
107
assessmentsinEducation.2,doi:10.1186/2196-0739-2-1(2014).
108
2.Samejima,F.EstimationofLatentAbilityUsingaResponsePatternofGraded
109
Scores.PsychometricMonographNo.17,PsychometricSociety,Richmond,VA.
110
(1969).
111
3.Houts,C.R.,Cai,L.flexMIRT:FlexibleMultilevelItemFactorAnalysisand
112
TestScoringUser’sManual(2012).
113
4.Cai,L.High-dimensionalexploratoryitemfactoranalysisbyaMetropolis-
114
HastingsRobbins-Monroalgorithm.Psychometrika,75,33–57(2010).
115
5.Cai,L.Metropolis-HastingsRobbins-Monroalgorithmforconfirmatoryitem
116
factoranalysis.JournalofEducationalandBehavioralStatistics,35,307–335
117
118
(2010).