Fazelb,c,* 6 a Oxford Uehiro Centre for Practical Ethics, Faculty of Philosophy, University of Oxford, Suite 8, Littlegate House, St Ebbes Street, Oxford OX1 1PT, 7 United Kingdom Depart
Trang 11 Original article
5 Q1T Douglasa, J Pugha, I Singha,b, J Savulescua, S Fazelb,c,*
6 a Oxford Uehiro Centre for Practical Ethics, Faculty of Philosophy, University of Oxford, Suite 8, Littlegate House, St Ebbes Street, Oxford OX1 1PT,
7 United Kingdom
Department of Psychiatry, University of Oxford, Warneford Hospital, Oxford OX3 7JX, United Kingdom
Oxford Health NHS Foundation Trust, Warneford Hospital, Oxford OX3 7JX, United Kingdom
10
11 Therearecurrentlymorethan200structuredtoolsavailablefor
12 assessing risk of violence in forensic psychiatry and criminal
13 justice[1].Thesearewidelydeployedtoinforminitialsentencing,
14 paroledecisions,anddecisionsregardingpost-releasemonitoring
15 andrehabilitation.Insomejurisdictions,includingCanada,New
16 Zealand,anduntil2012theUnitedKingdom,riskassessmenttools
17 are or were also used to justify indeterminate post-sentence
18 detention.Inaddition,violenceriskassessmenttoolsareusedto
19 inform decisions regarding detention, discharge, and patient
20 managementinforensicand,increasingly,generalpsychiatry
21 Thisarticlehighlightssomepotentialethicalproblemsposedby
22 riskassessment toolsand arguesthat betterdataon predictive
23 accuracyareneededtomitigatethese.Itfocusesontheuseofrisk
24 assessment tools in forensic psychiatric and criminal justice
25 settings
26 1 Professionalobligationsandcompetingvalues
27 Inthepsychiatricliterature,criticismof riskassessment has
28 focusedonthepossibilitythat,indeployingriskassessmenttools,
29 mental health professionalsmay fail tofulfil their professional
30 obligations to their patients [2,3] Health professionals are
31 expectedto make thecareof their patients their first concern,
32
to build trust and to respect patient preferences, and this is
33 reflectedinprofessionalguidelines[4].Somearguethattheuseof
34 riskassessmenttoolsisunjustifiedwhenitisintendedtorealise
35 othervalues, suchasjusticeor publicprotection,and does not
36 benefittheassessedindividual[5–8].BuchananandGroundshold
37 that‘‘itisinappropriatetocommentonadefendant’sriskunless
38 psychiatricinterventionisproposedorotherbenefitwillresult’’
39 [9].Similarly, Mullenclaims that‘‘[r]isk assessments arethe
40 proper concern of health professionals to theextent that they
41 initiateremedialinterventionsthatdirectlyorindirectlybenefit
42 thepersonassessed’’[10]
43 Theuseofriskassessmenttoolsisperhapsmostclearlyatodds
44 withtheinterestsoftheassessedindividualwherethetoolisused
45
to inform decisions regarding post-sentence detention In this
46 context,thedefaultpositionisthatthepersonwillbereleased;
47 however,ifthetoolindicatesahighriskofviolence,detentionmay
48
beextended.Itcouldbearguedthatdeployingthetoolthusruns
49 against the individual’s interest in being released as soon as
50 possible
51
Inothercases,however,theapplicationofarisk assessment
52 tool willbenefittheindividual.There areatleastthreewaysin
European Psychiatry xxx (2016) xxx–xxx
A R T I C L E I N F O
Article history:
Received 17 September 2016
Received in revised form 4 December 2016
Accepted 11 December 2016
Available online xxx
Keywords:
Violence
Forensic psychiatry
Ethics and human rights
Risk assessment
Crime prediction
Racial profiling
Discrimination
A B S T R A C T
Q2
* Corresponding author at: Department of Psychiatry, Medical Sciences Division,
University of Oxford, Warneford Hospital, Oxford OX3 7JX, United Kingdom.
E-mail address: seena.fazel@psych.ox.ac.uk (S Fazel).
ContentslistsavailableatScienceDirect
j our na l ho me p a ge : ht t p: / / w ww e ur opsy -j ou rna l c om
http://dx.doi.org/10.1016/j.eurpsy.2016.12.009
0924-9338/ C 2016 The Author(s) Published by Elsevier Masson SAS This is an open access article under the CC BY license ( http://creativecommons.org/licenses/by/4.0/ ).
Trang 253 whichitcouldconfersuchabenefit.First,theriskassessmentmay
54 beusedtoidentifybeneficialtreatments.Second,theuseofarisk
55 assessment tool may facilitate an earlier release or discharge
56 Supposeanindividualisbeingconsideredforparoleordischarge
57 fromasecurepsychiatricinstitution,butthisislikelytoberefused
58 onthebasisthat thereisinsufficientevidenceforalow riskof
59 violence.Inthissituation,applicationofariskassessmenttoolmay
60 providethe evidencenecessary to secure an end to detention
61 Third,evenwhenariskassessmentresultsinfurtherdetention,it
62 mightneverthelessconferabenefitbecauseextendeddetentionis
63 itselfintheindividual’sbestinterests.Forexample,itmayprevent
64 re-offendingandanevenlongerperiodofdetentioninthefuture
65 Moreover,evenwhenmentalhealthprofessionalsadminister
66 riskassessments thatare against theassessedindividual’s best
67 interests, it is not clear they thereby violate a professional
68 obligation,forthe viewthat medicalprofessionals ought never
69 tocompromiseapatient’sbestinterestscanbecontested.Inthe
70 settingofinfectiousdiseasecontrolitwouldbewidelyaccepted
71 that physicians may sometimes compromise a patient’s best
72 interestsinordertopromoteothervalues,suchasthehealthof
73 familymembers and thewider public[11,12] Similarly, many
74 wouldholdthatanobstetricianmaysometimesacttoprotecta
75 futurechild,evenifthiscomesatsomecosttothepatient—thatis,
76 theprospectivemother[13].Itcanbearguedthataparallelpoint
77 holdsinrelationtoforensicpsychiatry:professionalsinthisfield
78 maysometimesgiveprecedencetovaluesbesidesthewelfareof
79 theirownpatients[14].Thosewhoholdthatriskassessmenttools
80 shouldbeusedonlywhentheybenefitthepatientmaythusbe
81 overstatingtheethicaldifficultiescreatedbysuchtools
82 Nevertheless, the presence of competing values in risk
83 assessmentdoescreateapotentialethicalproblem:itispossible
84 thatsomevalues willbeunjustifiablysacrificed for thesake of
85 others.Forexample,thereisariskthattheinterestsofindividual
86 patientsorprisonerswillbeunjustifiablycompromisedinthename
87 ofpublicprotection,orthereverse.Wewillarguethatalackof
88 highqualitydataonpredictiveaccuracycompoundsthisethical
89 risk
90 2 Predictiveaccuracy
91 Existingdatasuggestthatmostriskassessmenttoolshavepoor
92 tomoderateaccuracyinmostapplications.Typically,morethan
93 half of individuals classified by tools as high risk are false
94 positives—theywillnotgoontooffend[15].Thesepersonsmaybe
95 detainedunnecessarily.Falsepositivesmaybeespeciallycommon
96 inminorityethnicgroups[16,17]
97 Ratesoffalsenegativesareusuallymuchlower.Nevertheless,
98 intypicalcasesaround9%ofthoseclassedaslowriskwillgoon
99 tooffend[15].Theseindividualsmaybereleasedordischarged
100 tooearly, posingexcessive risk tothe public Suchfailures of
101 negative prediction are frequently associated with significant
102 controversyandoutrage,asreactionstorecenthighprofilecases
103 demonstrate[18]
104 Theprevalenceofpredictionerrorsdoesnotentirelyundermine
105 therationalefordeployingriskassessmenttools.Tobalanceriskto
106 thepublicagainsttheinterestsoftheassessedindividual,some
107 methodforassessingriskisrequired,andriskassessmenttools,
108 even if limited in accuracy, may be the best option available
109 However,tomitigatetheriskofinadequateorexcessivedetention,
110 thelimitationsofriskassessmenttoolsneedtobewellunderstood
111 andfactoredintoclinicalandcriminaljusticeresponses
112 Unfortunately, published validation findings for the most
113 widely used tools, which allow for predictive accuracy to be
114 estimated in advance, frequently present a misleading picture
115 [19].First,thoughthereareexceptions,mosttoolshavenotbeen
116 externallyvalidatedoutsideoftheirderivationsample[20,21].Of
117 particularconcern,fewvalidationstudieshavebeenconductedin
118 women,ethnicminoritypopulations,andindividualsmotivatedby
119 religious or political extremism [16,17,19] Consequently, it is
120 unclearhowfarreportedaccuracyfindingscanbeextrapolatedto
121 new settings and populations [22,23] Second, there is strong
122 evidencethatconflictsofinterestareoftennotdisclosedinthis
123 field,andsomeevidenceofpublicationandauthorshipbias[24]
124 (Authorship bias occurs when research on tools tends to be
125 publishedbytheauthorsofthosetools,whotypicallyfindbetter
126 performance.)Third,publishedstudiesfrequentlypresentonlya
127 smallnumberofperformancemeasuresthatdonotprovideafull
128 pictureofpredictiveaccuracy[25]
129 Thus, not onlyis the predictiveaccuracy ofrisk assessment
130 toolsimperfect,itisalsoimperfectlypresentedintheliterature
131 Thelimitedandskewedevidencebasecreatesariskthatdecision
132 makerswillrelymoreheavilyonriskassessmentscoresthantheir
133 accuracywarrants.Tomitigatethisrisk,thereisaneedforbetter
134 quality data covering more subpopulations Validation studies
135 shouldincludemorethanjustoneortwoperformancestatistics,
136 anddataonthenumbersoftrueandfalsepositivesandnegatives
137 should be clearly presented Conflict of interests need to be
138 disclosed, and reviews by authors with financial conflict of
139 interestsshouldbetreatedwithcaution
140
Inadditiontoriskingover-relianceonriskassessmentscores,
141 deficienciesintheevidencebasealsogenerateatleastthreemore
142 specific problems, which we explain below: they (i) thwart
143 attemptstomatchriskassessmenttoolstodifferentcontextsof
144 application, (ii) complicate efforts to determine whether risk
145 assessmenttoolsareunjustifiablydiscriminatoryorstigmatising,
146 andthereby(iii)contributetoariskthatcontentiousdemographic
147 variableswillbeprematurelyeliminatedfromassessmenttools
148
3 Therighttoolforthecontext
149 Selecting the optimal risk assessment tool for a given
150 applicationrequirestrade-offstobemadebetweenfalsenegatives
151 and false positives; attempts to reduce the number of false
152 positiveswillincreasethenumberoffalse negatives[26].Tools
153 withalowrateoffalsenegatives(duetohighsensitivity)willbe
154 most effective at protecting the public, and may garner most
155 politicalsupport,whiletoolswithalowrateoffalsepositives(due
156
tohighspecificity) willbest protecttherights and interests of
157 prisonersandpsychiatricpatients
158 Theoptimalbalancebetweenfalsepositivesandfalsenegatives
159
isanethicaldecisionandwilldependonthesocialandpolitical
160 contextinwhichthetoolistobeused[27].Forexample,avoidance
161
offalsepositivesmaybemoreimportantinjurisdictionswithless
162 humane detention practices than in jurisdictions with more
163 humane practices, since the less humane the conditions of
164 detention,thegreatertheharmfalsepositiveswilltendtoimpose
165
ontheassessedindividual[28]
166 The appropriate balance between false positives and false
167 negatives will alsodependon the stagein the criminaljustice
168 processorpatientpathwayatwhichthetoolwillbedeployed.For
169 instance,supposethat a riskassessment tool isused toinform
170 decisions about post-sentence detention in a setting where an
171 individual’sinitialsentenceisdesignedtobeproportionatetotheir
172 degreeofresponsibilityandtheseriousnessofthecrime.Inthis
173 case, detaining the individual beyond the end of the initial
174 sentenceinvolves imposing a disproportionatelylong period of
175 detention.In thiscontext,specialcareshouldbetakentoavoid
176 falsepositives,andtheremaybegroundstopreferatoolwitha
177 verylowfalsepositiveratetoonethatisoverallmoreaccurate
178 However,thesituationisdifferentwhenatoolisusedtoinform
179 paroledecisions.Inthiscontext,falsepositivesmayleadtorefusal
180
ofparoleandanunnecessarilylongperiodofincarcerationfrom
T Douglas et al / European Psychiatry xxx (2016) xxx–xxx 2
Trang 3181 thepointofviewofpublicprotection.Yetifweassumethatthe
182 initialsentencesare themselvesproportionate, thentheoverall
183 period of detention for ‘false positive’ individuals will remain
184 withinthelimitsrequiredbyproportionality.Inthiscontextitmay
185 bemoreimportanttoavoidfalsenegatives
186 Matching risk assessment tools to different contexts of
187 applicationthusrequirestrade-offsbetweenpositiveandnegative
188 predictiveaccuracy.Foreachcontext,wemustfirstdecidewhich
189 typeofaccuracytoprioritisetowhichdegree,andthenselectatool
190 thatreflectsthispriority.Unfortunately,intheabsenceofreliable
191 data,it is notpossible tomake the latterdecision confidently
192 There is a need for studies using representative samples for
193 relevant subpopulations, avoiding highly selected samples,and
194 presentingperformancemeasuresthat allowfalse negativeand
195 false positive rates to be reliably estimated for a particular
196 application
197 4 Discriminationandstigmatisation
198 Some argue that singling out individuals for unfavourable
199 treatment on the basis of their demographic characteristics
200 amounts to unjustified discrimination This criticism is often
201 levelledatracialprofilingbypoliceandairportsecurity[29].A
202 similarconcernmightberaisedregardingriskassessmenttools
203 that take into account an individual’s demographic
characte-204 risticssuchasethnicity,age,immigrationstatusandgender.It
205 hasbeen suggested thatrisk assessmenttools should employ
206 only ‘individualised’ information, such as information about
207 declared plans and desires based on face to face interviews
208 [30,19],though,eventhen,judgmentsmaybesubjecttoimplicit
209 biasesbasedonthedemographiccharacteristicsoftheindividual
210 beingassessed[31]
211 However, the requirement to utilise only individualised
212 informationis overlyrestrictive.Some wouldargue that
demo-213 graphic profiling is discriminatory, or problematically so, only
214 whenthedemographicvariablesusedarerecognisedsocialgroups
215 (such as ethnic or gender groups) [32], or certain kinds of
216 recognisedsocialgroups,forinstance,thosewhosemembershipis
217 unchosen[33],orthathavehistoricallybeensubjecttooppression
218 [34] Risk assessment tools could theoretically exclude such
219 variables
220 Inreply,itmightbearguedthatexclusionofsuchvariablesis
221 insufficienttoavoidmoralconcerns.First,eveniftheproblematic
222 demographicvariables areformallyexcludedfromtheanalysis,
223 they may continue to exert an influence; there remains the
224 potentialfor implicitbiasin theapplication ofrisk assessment
225 toolsandinterpretationofriskscores[16,19,17].Second,evenif
226 theproblematicdemographicvariablesareformallyexcludedfrom
227 theanalysisandthereisnoimplicitbiasinapplyingthetools,there
228 may still be a correlation between membership of certain
229 demographicgroupsand riskscore.Forexample,membersofa
230 particularethnicgroupmaybemorelikelythanaveragetoreceive
231 highriskscores.Some mayholdthatsuchacorrelationisitself
232 problematic, especially if it is due to past wrongdoing against
233 membersofthedemographicgroupinquestion(e.g.,membersof
234 theethnicgroupareindeedmorelikelytooffend,butonlybecause
235 theyarevictimsofunjustsocialexclusion),ifthecorrelationdoes
236 notreflectatruedifferenceinrisk(e.g.,falsepositivesoccurmore
237 frequentlythan averagein theminority ethnicgroup), orifthe
238 correlationislikelytoleadtostigmatisationofthegroupdeemed
239 tobehigherrisk
240 However,eveniftheuseofriskassessmenttoolsdoesinvolvea
241 problematic form of discrimination or stigmatisation, it could
242 nevertheless be justified if the case in favour of using the
243 informationispowerfulenough.Theparallelwithracialprofiling
244 in airport screening is instructive here Airport screening is a
245 limitedresourceandtherearereasonstodeployittodetectthe
246 maximumnumberofwould-beterrorists.Ifprofilingenablesafar
247 greater number of terrorist attacks to be prevented with the
248 resourcesavailablethananyotherpolicy,andifthecosttothose
249 profiled is low, then it is arguably justified even if somewhat
250 problematic,forexample,becausediscriminatoryorstigmatising
251 Similarly,theresourcesavailableforthepreventionofviolenceare
252 limited,andifdeployingariskassessmenttoolpreventsfarmore
253 violencethan couldotherwisebepreventedwiththeresources
254 available,itmightbejustifiedevenifitdoesraisesomeconcerns
255 aboutdiscriminationandstigmatisation
256 Nevertheless,itisimportantthatriskassessmenttoolsdeploy
257 themostspecificpredictiveinformationavailable.Arguably,what
258
ismostobjectionableaboutsomeformsofracialprofilingisthat
259 theydeployracialappearanceasapredictorwhenmorespecific
260 predictors of security threat are available and, were these
261 predictorsused,racialappearancewouldaddnofurtherpredictive
262 value [35,36] In such circumstances, use of racial appearance
263 seemsunnecessary
264 Similarly,itmaybeproblematictousedemographicpredictors
265
inriskassessmenttoolswhenmorespecificpredictorsoffuture
266 offendingareavailableandthesepredictorswouldrendertheuse
267
ofdemographiccategoriesredundant
268 Unfortunately,thelackofgoodevidenceonaccuracymakesit
269 difficult to ascertain whether existing tools do use the most
270 specificpredictorsavailable.Todeterminethis,wewouldneedto
271
beabletocomparetheaccuracyofmorespecificandlessspecific
272 tools using relevant, reliable and unbiased data on accuracy
273 Currentlydeployedtoolsfrequentlydo usedemographicfactors
274 suchasageandimmigrationstatusaspredictors,and although
275 recentevidencesuggeststhatincludingsuchdemographicfactors
276 improvespredictiveaccuracy[37,38],furtherdataareneededto
277 confirmthis
278
Intheabsenceofthesedata,therearetworisks.Ontheone
279 hand, mental health professionals may continue to employ
280 coarse demographic variables that result in unnecessary
281 discrimination or stigmatisation On the other, given growing
282 public concern regarding the use of such variables [39,40],
283 professionals orpolicy makersmayprematurelyremovethem
284 from riskassessmenttools [41].Before variables areremoved
285 becausetheyarepotentiallycontentious,highqualityresearch
286 that uses transparent methods and presents all relevant
287 outcomesshould investigatewhether thedemographicfactors
288 included in current tools add incremental validity to tool
289 performance
290 Funding
291 ThisworkwassupportedbygrantsfromtheWellcomeTrustQ3
292 [100705/Z/12/Z] (WT086041/Z/08/Z, #095806, WT104848/Z/14/
293 Z),andtheUehiroFoundationonEthicsandEducation
294 Disclosureofinterest
295
SFhaspublishedresearchonriskassessment,includingaspart
296
ofateamthathasderivedandvalidatedonetoolforprisonerswith
297 psychiatricdisorders
298
299
[1] Singh JP, Desmarais SL, Hurducas C, Arbach-Lucioni K, Condemarin C, Dean K,
300
et al International perspectives on the practical application of violence risk
301
assessment: a global survey of 44 countries Int J Forensic Mental Health
302
2014;13:193–206.
303
[2] Eastman N, et al ‘‘Professional duties and personal integrity’’ and ‘‘conflictingQ5
304
ethical values’’ In: Forensic psychiatry Oxford University Press; 2012.
305
[3] Appelbaum PS Dangerous severe personality disorders: England’s experiment
306
in using psychiatry for public protection Psychiatr Serv 2005;56(4):397–9.
Trang 4307 [4] General Medical Council (GMC) Good medical practice: guidance for doctors –
308 the duties of a doctor; 2013, http://www.gmc-uk.org/guidance/
309 good_medical_ practice/duties_of_a_doctor.asp
310 [5] Eastman N The ethics of clinical risk assessment and management: developing
311 law and the role of mental health professionals In: Gray N, Laing J, Noaks L,
312 editors Criminal justice, mental health and the politics of risk Cavendish
313 Publishing; 2002.
314 [6] Buchanan A, Grounds A Forensic psychiatry and public protection Br J
315 Psychiatry 2011;198(6):420–3.
316 [7] Gunn J Future directions for treatment in forensic psychiatry Br J Psychiatry
317 2000;176(4):332–8.
318 [8] Mullen PE Forensic mental health Br J Psychiatry 2000;176(4):307–11.
319 [9] Buchanan A, Grounds A Forensic psychiatry and public protection Br J
320 Psychiatry 2011;198(6):420–3 http://dx.doi.org/10.1192/bjp.bp.111.095471
321 at p 421.
322 [10] Mullen PE Forensic mental health Br J Psychiatry 2000;176(4):307–11 http://
323 dx.doi.org/10.1192/bjp.176.4.307 at p 308.
324 [11] Gostin LO Public health law: power, duty, restraint University of California
325 Press; 2001.
326 [12] Wilkinson TM Contagious disease and self-defence Res Publica 2007;13:339–
328 [13] Savulescu J Future people, involuntary medical treatment in pregnancy and
329 the duty of easy rescue Utilitas 2007;19:1–20.
330 [14] Appelbaum PS A theory of ethics for forensic psychiatry J Am Acad Psychiatry
331 Law 1997;25:233–47.
332 [15] Fazel S, Singh JP, Doll H, Grann M Use of risk assessment instruments to
333 predict violence and antisocial behaviour in 73 samples involving 24 827
peo-334 ple: systematic review and meta-analysis BMJ 2012;345:e4692.
335 [16] Shepherd SM Violence risk instruments may be culturally unsafe for use with
336 indigenous patients Australas Psychiatry 2016;24:565–7.
337 [17] Shepherd SM, Lewis-Fernandez R Forensic risk assessment and cultural
338 diversity – contemporary challenges and future directions Psychol Public
339 Policy Law 2016;22:427–38.
340 [18] Parry H Rapist released halfway through sentence went on to attack three
341 more while on parole including a schoolgirl who was raped in front of her
342 boyfriend Daily Mail 2015
http://www.dailymail.co.uk/news/article-343
3131895/Rapist-released-halfway-sentence-went-attack-three-parole-344 including-two-schoolgirls-raped-boys-with.html
345 [19] Shepherd SM, Sullivan D Covert and implicit influences on the interpretation
346 of violence risk instruments Psychiatry Psychol Law 2016 http://dx.doi.org/
348 [20] Siontis GCM, Tzoulaki I, Castaldi PJ, Ioannidis JPA External validation of new
349 risk prediction models is infrequent and reveals worse prognostic
discrimi-350 nation J Clin Epidemiol 2015;68:25–34.
351 [21] Siontis and collaborators found that only 16% of new prediction models are
352 validated by different authors within five years of their first use, and when
353 external validation of tools did occur, predictive accuracy was generally
354 reduced Siontis GCM, et al External validation of new risk prediction models
355 is infrequent and reveals worse prognostic discrimination J Clin Epidemiol
356 2015;68(1):25–34 http://dx.doi.org/10.1016/j.jclinepi.2014.09.007
357 [22] Zhou J, Witt K, Xiang Y, Zhu X, Wang X, Fazel S Violence risk assessment in
358 psychiatric patients in China: a systematic review Aust N Z J Psychiatry
359 2015;50:33–45.
360
[23] A recent Chinese review found that instruments developed in the West had
361
significantly lower predictive accuracy when used in China compared to that
362
reported for Western populations See Western A, et al Violence risk
assess-363
ment in psychiatric patients in China: a systematic review Aust N Z J
364
Psychiatry 2015;50:33–45.
365
[24] Singh JP, Grann M, Fazel S Authorship bias in violence risk assessment? A
366
systematic review and meta-analysis PLoS ONE 2013;8:e72484.
367
[25] Singh JP Predictive validity performance indicators in violence risk
assess-368
ment: a methodological primer Behav Sci Law 2013;31:8–22.
369
[26] Walker N Dangerousness and mental disorder R Inst Philos Suppl
370
1994;37:179–90 http://dx.doi.org/10.1017/S1358246100010055 at p 182.
371
[27] Sinnott-Armstrong W Neural lie detection in courts In: Buzzi E, Hyman SE,
372
Raichle ME, Kanwisher N, Phelps EA, Morse SJ, et al., editors Using imaging to
373
identify deceit: scientific and ethical questions American Academy of Arts and
374
Sciences; 2009 p 35–9.
375
[28] Ryberg J Racial profiling and criminal justice J Ethics 2010;15:79–88.
376
[29] Lever A Why racial profiling is hard to justify: a response to Risse and
377
Zeckhauser Philos Public Affairs 2005;33:94–110.
378
[30] Nadelhoffer T, Bibas S, Grafton S, Kiehl KA, Mansfield A, Sinnott-Armstrong W,
379
et al Neuroprediction, violence, and the law: setting the stage Neuroethics
380
2012;5:67–99.
381
[31] Amodio DM, Devine PG Stereotyping and evaluation in implicit race bias:
382
evidence for independent constructs and unique effects on behavior J Pers Soc
383
Psychol 2006;91:652–61.
384
[32] Lippert-Rasmussen K Born free and equal? A philosophical inquiry into the
385
nature of discrimination Oxford University Press; 2013 p 336.
386
[33] Kahlenberg RD The remedy: class, race, and affirmative action Basic Books;
387
1997 p 386.
388
[34] Scanlon T Moral dimensions: permissibility, meaning Blame: Belknap Press;
389
2008 p 247.
390
[35] Lippert-Rasmussen K Nothing personal: on statistical discrimination J Polit
391
Philos 2007;15:385–403.
392
[36] Lippert-Rasmussen K ‘‘We are all different’’: statistical discrimination and the
393
right to be treated as an individual J Ethics 2010;15:47–59.
394
[37] Fazel S, Chang Z, Fanshawe T, La˚ngstro¨m N, Lichtenstein P, Larsson H, et al.
395
Prediction of violent reoffending on release from prison: derivation and
396
external validation of a scalable tool Lancet Psychiatry 2016;3(6):535–43.
397
[38] Skeem JL, Lowenkamp CT Risk, race, and recidivism: predictive bias and
398
disparate impact Soc Sci Res Netw 2016 http://dx.doi.org/10.2139/
399
ssrn.2687339
400
[39] Angwin J, Larson J, Mattu S, Kirchner L Machine Bias: there’s software used
401
across the country to predict future criminals And it’s biased against blacks.
402
ProPublica 2016
https://www.propublica.org/article/machine-bias-risk-403
assessments-in-criminal-sentencing [accessed 02.12.16].
404
[40] Barry-Jester AM, Casselman B, Goldstein D Should prison sentences be
405
based on crimes that haven’t been committed yet? FiveThirtyEight 2015.
406
http://fivethirtyeight.com/features/prison-reform-risk-assessment/
[acces-407
sed 02.12.16].
408
[41] Monahan J, Skeem JL Risk assessment in criminal sentencing Annu Rev Clin
409
Psychol 2016;12:489–513.
T Douglas et al / European Psychiatry xxx (2016) xxx–xxx 4