PROBABILITY AND STATISTICS IN THE LAW

42
PROBABILITY AND STATISTICS IN THE LAW Philip Dawid University College London

description

PROBABILITY AND STATISTICS IN THE LAW. Philip Dawid University College London. STATISTICS = LAW. Interpretation of evidence Hypothesis testing Decision-making under uncertainty. Prosecution Hypothesis. INGREDIENTS. Defence Hypothesis. Evidence. BAYESIAN APPROACH. - PowerPoint PPT Presentation

Transcript of PROBABILITY AND STATISTICS IN THE LAW

Page 1: PROBABILITY AND STATISTICS IN THE LAW

PROBABILITY AND STATISTICS IN THE LAW

PROBABILITY AND STATISTICS IN THE LAW

Philip Dawid

University College London

Page 2: PROBABILITY AND STATISTICS IN THE LAW

STATISTICS = LAWSTATISTICS = LAW

• Interpretation of evidence

• Hypothesis testing

• Decision-making under uncertainty

Page 3: PROBABILITY AND STATISTICS IN THE LAW

INGREDIENTSINGREDIENTS

Prosecution Hypothesis G

Defence Hypothesis G

Evidence E

Page 4: PROBABILITY AND STATISTICS IN THE LAW

– or posterior odds:

)|( EGP

)|(

)|(

E

E

GP

GP

• BAYESIAN APPROACH• BAYESIAN APPROACH

• FREQUENTIST APPROACH• FREQUENTIST APPROACH

and

)|( GP E

)|( GP E

Find posterior probability of guilt:

Look at & effect on

decision rules

Page 5: PROBABILITY AND STATISTICS IN THE LAW

SALLY CLARKSALLY CLARK

• Sally and Stephen Clark’s sons Christopher and Harry died suddenly at ages 11 and 8 weeks, in Sally’s care

• The Clarks claimed that their children had died from natural causes (SIDS??)

• Contested prosecution medical evidence of maltreatment

–SALLY CONVICTED OF MURDER

Page 6: PROBABILITY AND STATISTICS IN THE LAW

• A paediatrician testified that, for a family like the Clarks, the probability of one child dying from SIDS is 1 in 8,543

At Trial:At Trial:

• He was asked if the report calculated “the risk of two infants dying in that family by chance.”

• Answer: Yes, you have to multiply 1 in 8,543 times 1 in 8,543 …. [the CESDI study] points out that it’s approximately a chance of

1 in 73 million

Page 7: PROBABILITY AND STATISTICS IN THE LAW

WHAT TO THINK?WHAT TO THINK?

• Clear intuitive argument against independence (and thus calculation of “1 in 73 million”)

• BUT probability of 2 natural deaths remains very small

HOW TO CONSIDER?

Page 8: PROBABILITY AND STATISTICS IN THE LAW

Prosecutor’s FallacyProsecutor’s Fallacy

)|( EGP

)|( GP E• = 1 in 73 million

• Probability of deaths arising from natural causes is 1 in 73 million

• = 1 in 73 million

• Probability of innocence is 1 in 73 million

Page 9: PROBABILITY AND STATISTICS IN THE LAW

Alternatively…Alternatively…

• P(2 babies die of SIDS) = 1/73 million

• P(2 babies die of murder) = 1/2000 million

BOTH figures are equally relevant to the decision between the two possible causes

Page 10: PROBABILITY AND STATISTICS IN THE LAW

BAYES:BAYES:

POSTERIOR

ODDS

)(

)(

)(

)(

)|(

)|(

GP

GP

GP

GP

GP

GP

|E

|E

E

E

=LIKELIHOOD

RATIO PRIOR

ODDS

If prior odds = 1/2000 million posterior odds = 0.0365

%5.3)|( EGP

73m ??

Page 11: PROBABILITY AND STATISTICS IN THE LAW

IDENTIFICATION EVIDENCEIDENTIFICATION EVIDENCE

:

:

:

S

C

I i

Assume

million10/1)(])[,|(

1])[,|(

xIPxIGP

xIGP

CS

S

E

E

“match probability”

),(: xIxI SC E

Individual i

Criminal

Suspect Evidence:

Match

Page 12: PROBABILITY AND STATISTICS IN THE LAW

PROSECUTOR’S ARGUMENTPROSECUTOR’S ARGUMENT

The probability of a match having arisen by innocent means is 1/10 million.

So )|( EGP = 1/10 million

– i.e. )|( EGP is overwhelmingly close to 1

– CONVICT

Page 13: PROBABILITY AND STATISTICS IN THE LAW

DEFENCE ARGUMENTDEFENCE ARGUMENT

• Absent other evidence, there are 30 million potential culprits

• 1 is GUILTY (and matches)

• ~3 are INNOCENT and match

• Knowing only that the suspect matches, he could be any one of these 4 individuals

• So 41)|( EGP

–ACQUIT

Page 14: PROBABILITY AND STATISTICS IN THE LAW

BAYESBAYES POSTERIOR ODDS = (10 MILLION) “PRIOR” ODDS

)|(

)|(

BGP

BGP

PROSECUTOR’S argument OK if

Only BAYES allows for explicit incorporation of B

2/1)|( BGP

DEFENCE argument OK if million 1/30)|( BGP

MPLR /1

Page 15: PROBABILITY AND STATISTICS IN THE LAW

The Island ProblemThe Island Problem

• N+1 on island: N (100) innocent, 1 guilty

• Match, probability = P (0.004)

• Prosecution:

• Defence:

PGP 1)|( E

)1/(1)|( NPGP E

(0.996)

(0.714)

Page 16: PROBABILITY AND STATISTICS IN THE LAW

Other ArgumentsOther Arguments

Let number of individuals i having Ii = x be M

)|()|( 1 EE MEGP

– need distribution of M given

Note: Initially

1),|( MMGP E

So

),(: xIxI SC E

):1(Bin~ PNM

Page 17: PROBABILITY AND STATISTICS IN THE LAW

Argument 1Argument 1

• Evidence tells us

• So

1M

)1);;1(Bin~|()|( 1 MPNMMEGP E

(0.902)

Page 18: PROBABILITY AND STATISTICS IN THE LAW

Argument 2Argument 2

• Evidence tells us 1 (guilty) individual has x

• Our of remaining N innocents, number with x is ; while

• So

):(Bin~ PNM

(0.824)

MM 1

PN

P

PNMMEGPN

)1(

)1(1

));(Bin1~|()|(1

1

E

Page 19: PROBABILITY AND STATISTICS IN THE LAW

Argument 3Argument 3

• Evidence E is equivalent to 2 successes on 2 Bernoulli trials with replacement

• So

• So

• Then (0.714

– as for defence)

2

1)|(

N

mmMP E

mNm PPm

NmmMP

12 )1(

1)|( E

)1/(1

)|()|( 1

NP

MEGP

EE

Page 20: PROBABILITY AND STATISTICS IN THE LAW

DENIS ADAMSDENIS ADAMS

– Match probability = 1/200 million

1/20 million

1/2 million

Doesn’t fit descriptionVictim: “not him”Unshaken alibiNo other evidence to link to crime

• Sexual assault• DNA match

Page 21: PROBABILITY AND STATISTICS IN THE LAW

BAYES’S THEOREMBAYES’S THEOREM

POSTERIOR ODDS on guilt

= LIKELIHOOD RATIO PRIOR ODDS

= 2 million (1 / 200,000)

= 10 (10:1)

Posterior probability of guilt = 10/11

= 91%

Reasonable doubt – ACQUIT

Page 22: PROBABILITY AND STATISTICS IN THE LAW

WHAT ABOUT OTHER EVIDENCE?WHAT ABOUT OTHER EVIDENCE?

• Didn’t fit description• Victim: “not him”• Unshaken alibi

LR = 0.1 / 0.9 = 1/9

LR = 0.25 / 0.5 = 1/2

Apply Bayes’s Theorem again:Final odds on guilt = 10 1/9 1/2

}

= 5/9 (5:9) (probability of guilt = 5/14 = 35%)

Page 23: PROBABILITY AND STATISTICS IN THE LAW

Dependence on Match Probability

Match probability 1/200m 1/20m 1/2m

Posterior probability of guilt

98% 85% 35%

– number of noughts does matter!

Page 24: PROBABILITY AND STATISTICS IN THE LAW

DATABASE SEARCHDATABASE SEARCH

• Crime trace, frequency (match probability) 1 in 1 million

• Search Police DNA database (D) of size 10,000

• Find unique match: “John Smith” (S)

• No other evidence

Page 25: PROBABILITY AND STATISTICS IN THE LAW

Defence CaseDefence Case

• Probability of finding a match in database if innocent ~ 10,000 (1/1,000,000) = 1/100

• Match probability of 1/100 is not convincing evidence

• Evidence against John Smith is (significantly) weakened by virtue of database search

– ACQUIT

Page 26: PROBABILITY AND STATISTICS IN THE LAW

Prosecution CaseProsecution Case

• We have examined 10,000 individuals

• Of these, 9,999 found not to match

• This has reduced the pool of potential alternative culprits

• Evidence against John Smith is (marginally) strengthened by virtue of database search

– CONVICT

Page 27: PROBABILITY AND STATISTICS IN THE LAW

Which likelihood ratio?Which likelihood ratio?• Hypothesis HS: “John Smith did it” is data-

dependent• Replace by hypothesis HD: “Someone in

database D did it”– equivalent after search identifies S (but not before)

• LR = 1/(match probability) is now only 100– weak evidence?

• But HD is a priori 10,000 times more probable than HS

– posterior odds the same! – agrees with prosecution argument

Page 28: PROBABILITY AND STATISTICS IN THE LAW

Multiple StainsMultiple Stains

• 2 DNA stains– 1 on sheet, 1 on pillow

– assume 2 perpetrators, 1 stain from each

• John Smith (S) matches pillow stain– associated “match probability” P

• What are appropriate hypotheses, likelihoods, inferences?

Page 29: PROBABILITY AND STATISTICS IN THE LAW

HypothesesHypotheses• S left one of 2 stains

• S left pillow stain

• S left pillow stain

• S left neither stain

• S left neither stain

• S didn’t leave pillow stain

2/PLR

PLR

)1(2/)2( PLR

( = prior probability S is guilty)

Page 30: PROBABILITY AND STATISTICS IN THE LAW

What to present in Court?

• Hypotheses equivalent (only) after data

• Different prior odds

• Identical posterior odds

Page 31: PROBABILITY AND STATISTICS IN THE LAW

Mixed StainsMixed Stains

• Crime trace containing DNA from more than 1 contributor–Rape

–Scuffleetc

Page 32: PROBABILITY AND STATISTICS IN THE LAW

O. J. SIMPSONO. J. SIMPSON

Crime

OJS

RG

A

B

C

Marker DQ-Frequency

13%

20%

28%

“MATCH” to OJS

Allele

Page 33: PROBABILITY AND STATISTICS IN THE LAW

MATCH PROBABILITY?MATCH PROBABILITY?• PROSECUTION:

Frequency of OJS type

AB: 5%• DEFENCE:

Combined frequency of all matching types

AA, AB, AC, BB, BC, CC: 39%

• LR approach assuming Goldman (AC) in mixture:

AB, BB, BC: 21%• LR approach not assuming Goldman in mixture:

(more complex calculation) ~ 21%

Page 34: PROBABILITY AND STATISTICS IN THE LAW

MISSING DNA DATAMISSING DNA DATA

• What if we can not obtain DNA from the suspect ? (or other relevant individual?)

• Sometimes we can obtain indirect information by DNA profiling of relatives

• But analysis is complex and subtle…

Page 35: PROBABILITY AND STATISTICS IN THE LAW

HANRATTYHANRATTY

• James Hanratty convicted and executed in 1962

• DNA profile from crime items analysed in 1998

• Population frequency less than 1 in 2.5 million

• DNA profiles from mother and brother – “consistent with” crime DNA being from Hanratty

(“A6” murder and rape, 1961)

Page 36: PROBABILITY AND STATISTICS IN THE LAW

PRESS REPORTSPRESS REPORTS

• “There is a 1 in 2.5 million chance that Hanratty was not the A6 killer”

• “The DNA is 2.5 million times more likely to belong to Hanratty than anyone else”

Likelihood Ratio based on profiles of mother and brother (complex calculation):

440

– even though no direct match to Hanratty!

Page 37: PROBABILITY AND STATISTICS IN THE LAW

DISPUTED PATERNITYDISPUTED PATERNITY

• MOTHER (m1) of CHILD (c1) claims that PUTATIVE FATHER (pf) is its TRUE FATHER (tf)

But DO have DNA profiles from:

• Two full BROTHERS (b1, b2) of PUTATIVE FATHER

undisputedchild

disputedchild

brothers

• His UNDISPUTED CHILD (c2) and its MOTHER (m2)

• DNA profiles from MOTHER and CHILD No profile from PUTATIVE FATHER

Page 38: PROBABILITY AND STATISTICS IN THE LAW

DECISION AIDDECISION AID“PROBABILISTIC EXPERT SYSTEM”

– embodies probabilistic relationships (between inherited genes)

Page 39: PROBABILITY AND STATISTICS IN THE LAW

ANALYSISANALYSIS

• Measurements for 12 DNA markers on all 6 individuals

• Enter data, “propagate” through system

• Overall Likelihood Ratio in favour of paternity:

~1300

Page 40: PROBABILITY AND STATISTICS IN THE LAW

FURTHER COMPLEX DNA CASES

FURTHER COMPLEX DNA CASES

• Contamination

• Laboratory errors, mix-up, fraud

• Relatives

– …

Page 41: PROBABILITY AND STATISTICS IN THE LAW

• Statistics

• Law

• Crime Science

• Psychology

• Economics• Philosophy of

Science

• Geography• Medicine• Ancient History• Computer Science• Education• …

EVIDENCE, INFERENCE AND ENQUIRY

EVIDENCE, INFERENCE AND ENQUIRY

www.evidencescience.org

Page 42: PROBABILITY AND STATISTICS IN THE LAW

EVIDENCE SCIENCEEVIDENCE SCIENCE

• Subject- and substance-blind approach• Inference, explanation, causality• Recurrent patterns of evidence• Narrative, argumentation, analysis, synthesis• Cognitive biases• Formal rules• Decision aids• Interdisciplinary studies• …