Upload
lethu
View
218
Download
0
Embed Size (px)
Citation preview
How study pre-registration can benefit neuroscience
Pia RotshteinSchool of Psychology, University of Birmingham
1
AcknowledgmentCortex RR editorial team
2
ChristopherChambersCardiffUniversity
ZoltanDienesUniversityofSussex
RobMcIntoshUniversityofEdinburgh
Talk Overview
• Problemswitheverydayresearchpracticesinpsychologyandcognitiveneuroscience
• Howstudypre-registrationcanhelp,includingdetailsoftheRegisteredReportsformatandexamples
• Whatcountasanhypothesisinaneuroimgaingstudy?
• ResponsetoFAQsandcriticisms
QUESTIONS
Intheinterestsofdoinggoodscience,whichpartofaresearchstudyshouldbebeyond yourcontrol?
Intheinterestsofadvancingyourcareer,whichpartofaresearchstudyismostimportantforpublishingin‘topjournals’?
Theresults
Theresults
Couldthisbewhyaliensflyrightby?
What’sbestforscience
Transparentandhighqualityresearch,
regardlessofoutcome
What’sbestforscientists
Producingalotof“goodresults”
Sciencehasanincentiveproblem
seeNosek,Spies&Motyl(2012).PerspectivesonPsychologicalScience,7(6):615–631
Whatisa“goodresult”?
“novel”“striking”“clear”“…foundasignificanteffect”“worked”“resultsshowconvincingly…”“majoradvance”“definitive”“beautiful”“breakthrough”“highimpact”
What happens when researchers are pressured to get “good results”?
Publicationbias– suppressionofnegativeorcomplexfindings
©DarioBattisti
Significancechasing– “p-hacking”,selectivereporting
HARKing – hypothesizingafterresultsareknown
Lowstatisticalpower–quantityofpapersoverquality
Lackofreplication – seenasboring,lackinginintellectualprowess
Lackofdatasharing– notime,toohard,noincentive
Generate and specify hypotheses
Design study
Collect dataAnalyse data & test hypotheses
Interpret data
Publish or conduct next experiment
Hypothetico-deductivescientificmethod
Generate and specify hypotheses
Design study
Collect dataAnalyze data & test hypotheses
Interpret data
Publish or conduct next experiment
Hypothetico-deductivescientificmethod
Publication biasLack of data sharing
Low statistical power
Significance chasing
Lack of replication
Inpsychology:1of100papersMakel etal(2012)
~50%chancetodetectmediumeffectsCohen(1962);Sedlmeier andGigerenzer (1989);BezeauandGraves(2001)
~50-100%prevalenceJohnetal(2012)
~50-90%prevalenceJohnetal(2012)Kerr(1998)
~92%positiveFanelli(2010)
~70%failureWicherts etal(2006)
Cyril Pernet
Jeanette Mumford
Whyisthishappening?
Becauseweplacetoomuchimportanceontheresults ofexperimentsandnotenoughontheprocessesthatproducethem
Resultsmakescienceexcitingbutjudgingthequalityofscience(andscientists)accordingtotheresultscondemnsustobeinga“soft”science
Can we fix this? Yes
Philosophy:Whatgiveshypothesis-testingitsscientificvalueis:• theQUESTIONitasks• theQUALITYofthemethodituses• nottheRESULTitproduces
Ifweacceptthisphilosophytheneditorialdecisionsatjournalsshouldbeblind toresults
RegisteredReports
FourcentralaspectsoftheRegisteredReports model:
• Partofthepeerreviewprocesstakesplacebeforeexperimentsareconducted
• Passingthisstageofreviewvirtuallyguaranteespublication
• Originalstudiesandhigh-valuereplicationsarewelcome
• Researchersdecidehypotheses,experimentalprocedures,andmainanalysesbefore datacollection
AuthorssubmitSTAGE1manuscriptwithIntroduction,ProposedMethods&
Analyses,andPilotData(ifapplicable)
Stage1peerreview
Ifreviewsarepositivethenjournaloffers in-principleacceptance(IPA),
regardlessofstudyoutcome(protocolnotpublishedyet)
Howitworks
Arethehypotheseswellfounded?
Arethemethodsandproposedanalysesfeasibleandsufficientlydetailed?
Isthestudywellpowered?(≥90%)
Havetheauthorsincludedsufficientpositivecontrolstoconfirmthatthestudywillprovideafairtest?
Editorialtriagereview
Total submission (44): 70% reject & invite to resubmit
1/18 rejected after stage 1Theoretically not sounds10 awarded IPA
Howitworks
Stage2peerreview Didtheauthorsfollowtheapprovedprotocol?
Didpositivecontrolssucceed?
Aretheconclusionsjustifiedbythedata?Manuscriptpublished!
Authorsdotheresearch
AuthorsresubmitcompletedSTAGE2manuscript:• Introduction andMethods (virtuallyunchanged)• Results(new):Registeredconfirmatoryanalyses
+unregisteredexploratoryanalyses• Discussion(new)• Datadepositedinapublicarchive
Around 8 have been published
What we learned?
◈ It was a “high risk & costly” study involving drug intervention and MEG
◈ The theoretical question was straight forward:◈ Are gamma bands oscillation associated with glutamatergic
principal cells
◈ Issues: how much drug to give? What should be the outcome neutral criteria?
◈ Reviewer was not happy with the results (suggesting non registered analyses)
18
What we learned?
◈ The theoretical question was complex:◈ Feature vs. category based organization within two regions of
interest anterior temporal, temporal-parietal junction◈ Issues: how to define region of interest◈ How specific the hypothesis need to be◈ What would be a neutral-outcome to assess data quality◈ Not ignoring null results◈ How to report results from whole brain analysis
20
Whatdoesmatter
THEHYPOTHESESARE CLEAR
THE STUDYIS POWERED
THEQUALITY OFTHE DATA IS
GOOD
THEMETHOD ISSOUND
BUT,whatisanacceptablehypothesisforimagingstudies?
• Myexperimentwillleadtochangeinactivation(/connectivity/amplitudechangeatanytimewindowetc.)somewhereinthebrain
• Myexperimentwillaffectaspecificfunctional/anatomicalstructures
23
25
1.“IsRegisteredReportssuitableforallsciences?”
• Applicabletoanyareaengagedindeductive,hypothesis-drivenresearchwhereoneormoreofthefollowingproblemsapply:
• Publicationbias• Significancechasing• HARKing – “hypothesizingafterresultsareknown”• Lowstatisticalpower• Lackofdirectreplication• Lackofdatasharing
3.“What’stostopRegisteredReportsfrombecomingadumpinggroundforinconclusivenullresults?”
4.“Won’tthislimitthereportingofserendipitousfindings?”
• aprioripowerrequirements(≥90%)increasereproducibilityofallfindings• Bayesianmethodswelcomed(B<0.33orB>3forsubstantialevidence).A
specialistBayeseditorhasbeenappointedatCortex (ZoltanDienes)
2.“What’stostopresearchersfrom‘pre-registering’astudythattheyhavealreadyconducted?”
• Time-stampedrawdatafilesmustbesubmittedatStage2withbasiclablogandcertificationfromallauthorsthatdatawascollectedafter provisionalacceptance
• SubmittingacompletedstudyatStage1wouldthereforebefraud• Strategywouldbackfireanywaywhenreviewersaskforprotocolamendments• RegisteredReportsaren’tdesignedtopreventfraudbuttoincentivizegoodpractice
• Thearenorestrictionsonthereportingofunregisteredexploratoryanalyses.• Confirmatoryandexploratoryanalyseswillsimplybelabeledcorrectly
5.“Thisistoomuchworkforauthors.Weusuallydon’tdecidehowtoanalyseourdatauntilafterwe’velookedatit”
6.“Thiswillbetoomuchworkforreviewers,whoarealreadyoverstretched”
• It’sasimilaramountofwork,justdoneatadifferenttime– andprovisionalacceptancevirtuallyguaranteesapublication,withoutthepressuretoobtain“goodresults”
• Notnecessary.RRscouldreducesequentialsubmissionofmanuscripts‘downthechain’ofjournals
• RRdoesrequiresindepthreviewofprospectivemethodsandtheoreticalgrounding.
7.“Reviewerscouldstealmyideasatthepre-registrationstageandscoopme”
• OnlyahandfulofpeopleknowabouteachStage1submission• Onceprotocolisaccepted,thejournalcan’trejectyourpaperbecausesomething
similarwaspublished(noveltyisirrelevant)• ManuscriptreceiveddateonpublishedRRwillbethedateofStage1submission• Howdifferentfromgrantapplications,conferencepresentations,seminars?
9.“What’stostopauthorswithprovisionalacceptancepullingtheirmanuscriptaftergettingstunningresultsandsubmittingittoNature/Science/PNAS?”
8.“Doesprovisionalacceptanceabsolutelyguarantee publication?”• No,butitgivespeaceofmindtoauthorsthattheirpaperswon’tberejectedbecause
ofnegativefindings,perceivednovelty/importanceofoutcomes.Nofiledrawer
• Nothing.InthatcasethejournalwillsimplypublishaWithdrawnRegistration,whichwillincludetheabstractandabriefexplanationforthewithdrawalofthepaper
28
10.“Ihavenoideaofwhateffectsizetoexpectinmyexperiment,sohowcanIdoapoweranalysisaspartofStage1?”
• Usuallythereisatleastsomeliterature.Butevenifnot,aminimaleffectsizeoftheoreticalinterestcanbespecified
• Ifminimaleffectsizeisuncertain,optionsareanorthodoxstatisticalapproachwithcorrectedpeeking(Strube etal.,2006,Beh ResMeth,38,24-27)orBayesianmethodstospecifydistributionofpossibleeffectsizes
• PilotresultstoestablisheffectsizearewelcomedinStage1submissions
11.“Sometimesadesignissound,butthedataisgarbagebecauseresearchersruntheexperimentpoorly.Howwillyoudistinguishnegativefindings/weirdresultsduetopoorpracticefromthosethataregenuine?”
• Authorsmustincludeoutcome-neutralconditionsforensuringthattheexperimentsarecapableoftestingthestatedhypotheses(e.g.positivecontrols,manipulationchecks)
12.“Ifpublicationisguaranteedinadvance,whywouldresearchersbotherrunningthestudycarefully?Thisschemewillincentivizefalsenegatives.”
• Runningthestudycarelesslywouldalsoderailtheoutcome-neutralteststhatarenecessaryforfinalacceptance
13.“Whathappensifweneedtochangesomethingaboutourexperimentalproceduresaftertheyareprovisionallyaccepted?”
• Minorchanges(e.g.replacingequipment)canbefootnotedinStage2manuscriptasprotocoldeviations
• Majorchanges(e.g.changingexclusioncriteria)requirewithdrawalandresubmission• Editorialteamdecideswhetherdeviationissufficientlyminortocontinue
15.“IhaveaccesstoanexistingdatasetthatIhaven’tyetanalysed.CanIsubmitthisproposedanalysisasaRegisteredReport?”
• NotatCortex,butotherjournalsareallowingthis…
14.“Someofmyanalyseswilldependontheresults,sohowcanIpre-registereachstepindetail?”(e.g.outlierexclusion,ROIselection)
• Pre-registrationdoesn’trequireeachdecisiontobespecified,onlythedecisiontree• Authorscanpre-registerthecontingencies/rulesforfuturedecisions
16.“Pre-registrationwilldenigrateexploratoryresearch”
• No,whatdenigratesexploratoryresearchisHARKing:presentingexplorationasconfirmation(shoehorningKuhnintoPopper)
• Exploratoryanalysesarewelcometobereported,butunderanexplicitheading.• Exploratoryresearchissimplynotvaluedinitsnativeform,soletsstartdoingso!
17.“HowwillRegisteredReportsincentivizereplicationstudies?”
• Conspiracyofcircumstancestellsusnottobotherdoingdirect(exact)replications
• Methodsectionsareoftentoovaguetoallowprecisereplication• Chroniclackofpowerinnovelresearchmeansthatreplicationsoften
requireverylargesamplessizes• Attemptingtoexactlyrepeatapreviousexperimentcanbeseen(in
psychology)asanactofaggression(cf.physics)• Mostpsych/neurojournalswantnoveltyandseereplicationsas(usually)
unpublishable
• RRs:haveproposedreplicationexperimentreviewedandprovisionallyacceptedbefore youinvestsubstantialresourcesintodoingit;potentiallyinvolveoriginalauthorsinpeerreviewoftheprotocol
RegisteredReportsinformationhubontheOpenScienceFramework
https://osf.io/8mpji/wiki/home/
Google“registeredreports”– tophit