Is there a world-wide epidemic of "health care serial killers" (killer nurses?). Or is there an epidemic of falsely accused health care serial killers? Analysis of the case of Lucia de Berk together with discussion of the role of statistics - in that case, and in forensic statistics in general

1. Lessons from Lucia
gill@math.leidenuniv.nl
http://www.math.leidenuniv.nl/~gill

2. Overview of Lecture
• Background
• Theory (statistical paradigms)
• Lucia
• Conclusions

3. Background
• Serial killer nurses: is there an epidemic?
• Victorino Chua (UK)
• Daniela Poggiali (It)
• Nils H. (Germany)
• Ben Geen (UK)
• …
• But perhaps also an epidemic of falsely
convicted innocent nurses !

4. Academic (?) research
• Katherine Ramsland (2007) Inside the minds
of health care serial killers: why they kill
• ElizabethYardley and David Wilson (2014)
In Search of the ‘Angels of Death’:
Conceptualising the Contemporary Nurse
Healthcare Serial Killer
“Red flag check-list”: inspired (in part) by the case of Lucia de Berk …
from the time when everyone knew she was guilty
Sources: newspaper reports and prosecution documents

5. was necessary. For example, through discussion, we interpreted regularly as meaning a pat-
tern of employment that was out of the ordinary, and which might have seen the nurse set-
tling at a job for only a few months and then moving on. However, it is also clear that at
some later stage this checklist should be further refined and scrutinised, especially as it
Charles Cullen 11
Kimberley Saenz 9
Kristen Gilbert 8
Robert Rubane Diaz 8
Sonia Caleffi 8
Beverley Allitt 7
Cecile Bombeek 7
Vickie Dawn Jackson 7
Aino Nykopp-Koski 6
Orville Lynn Majors 6
Benjamin Geen 5
Petr Zalenka 5
Christine Malevre 4
Irene Becker 4
Stephan Letter 4
Colin Norris 2
Table 15. Prevalence of items on ‘red flag’ checklist
Item N
% of cases in which this
item was present
1. Moves from one hospital to another 6 38
2. Secretive/difficult personal relationships 7 43
3. History of mental instability/depression 10 63
4. Predicts when someone will die 4 25
5. Makes odd comments/claims to be ‘jinxed’ 3 19
6. Likes to talk about death/odd behaviours when someone dies 1 6
7. Higher incidences of death on his/her shift 15 94
8. Seems inordinately enthused about his/her skills 6 38
9. Makes inconsistent statements when challenged about deaths 3 19
10. Prefers nightshifts—fewer colleagues about 6 38
11. Associated with incidents at other hospitals 6 38
12. Been involved with other criminal activities 7 43
13. Makes colleagues anxious/suspicious 9 56
14. Craves attention 7 43
15. Tries to prevent others checking on his/her patients 0 –
16. Hangs around during investigations of deaths 0 –
17. In possession of drugs at home/in locker 8 50
18. Lied about personal information 0 –
19. In possession of books about poison/serial murder 1 6
20. Has had disciplinary problems 4 25
21. Appears to have a personality disorder 8 50
22. Has a substance abuse problem 3 19
Note: % ≠ 100 as all cases had at least one checklist item
Copyright © 2014 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Profil. (2014)
DOI: 10.1002/jip
In Search of the ‘Angels of Death’: Conceptualising the
Contemporary Nurse Healthcare Serial Killer
ELIZABETH YARDLEY* and DAVID WILSON
Birmingham City University, Centre for Applied Criminology, City North Campus, Franchise Street,
Perry Barr, Birmingham, B42 2SU, United Kingdom
Abstract
Focusing specifically upon nurses who commit serial murder within a hospital setting, this
paper aims to establish insights into this particular subcategory of healthcare serial killer.
In addition, the paper aims to test the usefulness of an existing checklist of behaviours
among this group of serial murderers. Drawing upon existing lists of healthcare serial
killers produced by other scholars as well as legal records and an online news archive,
we identified and researched healthcare serial killer nurses, collating socio demographic
and criminological data and applying the aforementioned checklist to each case. Our find-
ings suggest that to date, the label ‘healthcare serial killer’ has been applied in too loose a
manner, making the understanding of this phenomenon problematic. In further refining the
definition and identifying the socio-demographic and criminological characteristics of the
victims, perpetrators and crimes, we have developed more specific and therefore useful
insights for practitioners and identified a potentially useful checklist which, with revisions,
could contribute towards preventative strategies and interventions. Copyright © 2014
John Wiley & Sons, Ltd.
Key words: healthcare serial killers; nurses; hospitals
INTRODUCTION
Healthcare serial killers (HSKs), sometimes also known as medical murderers (Hickey, 2010),
have emerged as a common concern within popular culture and true crime, with various books,
documentaries and films all eager to understand what might have motivated, for example,
Beverly Allitt, Harold Shipman or Charles Cullen to have taken the lives of their patients
(Davies, 1993; Graeber, 2013; Peters, 2005). However, whilst the HSK has been emerging
into popular consciousness, academic, criminological research about this type of offender
remains somewhat underdeveloped. Even so, a relatively small body of work has begun to iden-
tify, map and describe cases of healthcare serial murder (see, for example, Ramsland, 2007;
*Correspondence to: Elizabeth Yardley, Birmingham City University, Centre for Applied Criminology, City
North Campus, Franchise Street, Perry Barr, Birmingham B42 2SU, UK.
E-mail: elizabeth.yardley@bcu.ac.uk
Journal of Investigative Psychology and Offender Profiling
J. Investig. Psych. Offender Profil. (2014)
Published online in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/jip.1434
Copyright © 2014 John Wiley & Sons, Ltd.

6. Theory:
statistical paradigms
• Bayes (one person statistics)
• Frequentist (two person, collaborative statistics)
• Likelihood (avoiding the issue)
NB two paradigms of probability: “subjective” (Bayesian), “objective” (frequentist)

7. Bayes’ rule
• Posterior odds
= prior odds * likelihood ratio
• Likelihood ratio
= Prob( data | HP ) : Prob( data | HD )
Bayesian/frequentist peaceful coexistence theorem:
{Decision theoretic admissible} = {Bayesian (for some prior)}

8. Current research:
use of Bayes net
(aka graphical model)
• Bayesian model of (probabilistic) causality
• Bayesian means “probability as degree of
belief” (epistemological, not ontological)
• Statistical correlations “explained” by causal
dependence on past events
• Some of those events are known, others unknown
• Computations: with GeNIe, HUGIN Lite, or in R

9. wid and Evett (1997) consider a fictitious burglary case, described
ows:
An unknown number of o↵enders entered commercial premises
late at night through a hole which they cut in a metal grille. In-
side, they were confronted by a security guard who was able to set
o↵ an alarm before one of the intruders punched him in the face,
causing his nose to bleed.
The intruders left from the front of the building just as a police
patrol car was arriving and they dispersed on foot, their getaway
car having made o↵ at the first sound of the alarm. The security
guard said that there were four men but the light was too poor for
him to describe them and he was confused because of the blow he
had received. The police in the patrol car saw the o↵enders only
from a considerable distance away. They searched the surrounding
area and, about 10 minutes later, one of them found the suspect
trying to “hot wire” a car in an alley about a quarter of a mile
from the incident.
Example: Dawid and Evett (1997)

10. Example (ctd) : Dawid and Evett (1997)
At the scene, a tuft of red fibres was found on the jagged end of
one of the cut edges of the grille. Blood samples were taken from
the guard and the suspect. The suspect denied having anything to
do with the o↵ence. He was wearing a jumper and jeans which
were taken for examination.
A spray pattern of blood was found on the front and right sleeve
of the suspect’s jumper. The blood type was di↵erent from that of
the suspect, but the same as that from the security guard. The tuft
from the scene was found to be red acrylic. The suspect’s jumper
was red acrylic. The tuft was indistinguishable from the fibres of
the jumper by eye, microspectrofluorimetry and thin layer chro-
matography (TLC). The jumper was well worn and had several
holes, though none could clearly be said to be a possible origin for
the tuft.
In this example there are three general kinds of evidence: eye-witne
J. Mortera A. P. Dawid (2006), Probability and Evidence, Research Report No. 264,
Department of Statistical Science, University College London.

11. Example: Dawid and Evett (1997)
Squares = observed = evidence; circles = not observed; C = hypothesis of interest

12. Dawid and Evett (1997)
blood, and fibre; and for each kind a variety of individual evidential items.
We can summarise the salient features of the evidence against the suspect as
follows:
EYEWITNESS
G : The evidence of the security guard
W : The evidence of the police o cer who arrested the suspect
BLOOD
R : The bloodstain in the form of a spray on the suspect’s jumper
X1: Suspect’s blood type
X2: Guard’s blood type
Y2: Blood type of blood spray on jumper
FIBRES
X3: Properties of the suspect’s jumper
Y1: Properties of fibre tuft
The uncertain hypotheses and variables that enter are:
HYPOTHESES
C: Whether the suspect was or was not one of the o↵enders
11
We can summarise the salient features of the evidence against the suspect as
follows:
EYEWITNESS
G : The evidence of the security guard
W : The evidence of the police o cer who arrested the suspect
BLOOD
R : The bloodstain in the form of a spray on the suspect’s jumper
X1: Suspect’s blood type
X2: Guard’s blood type
Y2: Blood type of blood spray on jumper
FIBRES
X3: Properties of the suspect’s jumper
Y1: Properties of fibre tuft
The uncertain hypotheses and variables that enter are:
HYPOTHESES
C: Whether the suspect was or was not one of the o↵enders
11
A: The identity of the person who left the fibres on the grille
B: The identity of the person who punched the guard
N: The number of o↵enders

13. Of these the specific charge before the court is C = true; the others are
included to provide a complete account of the problem.
Figure 1: Bayesian network for burglary example
Dawid and Evett (1997)
J. Mortera A. P. Dawid (2006), Probability and Evidence, Research Report No. 264,
Department of Statistical Science, University College London.

14. Example: prior to
entering evidence

15. Example: posterior to
entering evidence

16. Lucia

17. Shifts Court dataCourt dataCourt data CorrectedCorrectedCorrectedCorrected
JKZ MCU-1 incidentincident incidentincident
Oct ’00 – Sept ’01 with without with with
Lucia
with 9 b133 b7 b13
Lucia
without 0 b887 b4 b88
RKZ-42
Aug – Nov ’97
Lucia
with b6 b52 b5 b5
Lucia
without b9 272 10 27
RKZ-41
Aug – Nov ’97
Lucia
with 1 bb0 1 bb
Lucia
without 4 361 4 35
Lucia: the data

18. LuciadeB.
Reconstructie van
een gerechtelijke
dwaling
LuciadeB.
Reconstructievaneengerechtelijkedwaling
n Magazines Ton Derksen
TonDerksen

19. Lucia: time-line
• Sept. 4, 2001,“unexpected” death of Amber
• 2003: life sentence for 4 murders and 2 attempts;
proof: statistical
• 2004: life sentence of 7 murders and 3 attempts;
proof: medical
• 2006: confirmed by supreme court
• 2006: publication of book by Ton Derksen
(philosopher of science)
• 2006: case submitted to special committee for
review of exceptional possibly unsafe convictions

20. • 2008: CEAS reports death of Amber natural,
recommends reopening
• 2008:“advocate-general” to supreme court admits
there is no “novum”, commissions further
investigations
• 2009:AG recommends case is reopened (with
“novum” if required: former key pathologist agrees
with new findings – he had less information at his
disposal
• 2009: supreme court accepts, case is reopened
• 2010: not-guilty verdict (all deaths natural; nurses
behaviour exemplary; medical errors)

21. Lucia: likelihood ratio
• Hypothesis of the prosecution: (most of
the) Lucia incidents are murders or
attempted murders
• Hypothesis of the defence: the events are
natural and would have happened anyway
• Prob(data|HP):Prob(data|HD)=1:1

22. Lucia: the original
statistical analysis
• Frequentist approach; hypothesis test;
null hypothesis = “balls in vases”
• For each of three data sets, court’s statistician
computed the “p-value” P(as extreme as Lucia or
more | balls in vases model)
• For JKZ MCU-I, he multiplied by 26
(= # nurses worked on the ward that year)
• Product of three p-values = 1 in 342 million

23. Lucia: the defense
• Judge:“what is the probability the coincidence is due
to chance?”
• Defence 1.There are so many different probability
models, you cannot compute a probability
• Defence 2. Multiplying p-values is wrong (reductio
ad absurdam)
• Judges:“we are not here to do thought experiments,
but to determine facts”
• Judges:“The verdict of the court does not depend
on a statistical computation of probabilities”

24. Lucia: the defense
• Judge:“what is the probability the coincidence is due
to chance?”
• Defence 1.There are so many different probability
models, you cannot compute a probability
• Defence 2. Multiplying p-values is wrong (reductio
ad absurdam)
• Judges:“we are not here to do thought experiments,
but to determine facts”
• Judges:“The verdict of the court does not depend
on a statistical computation of probabilities”

25. No one checked the
data!
• Three children responsible for multiple identical
events, some in Lucia’s shifts, some not
• No consistent definition of “incident”
• No consistent definition of “time of incident”
• The data suggested the hypothesis
• No-one studied the “normal” situation (clusters
of events, clusters of shifts are normal)

26. Shifts Court dataCourt dataCourt data Corrected dataCorrected dataCorrected dataCorrected data
JKZ MCU-1 incidentincident incidentincident
Oct ’00 – Sept ’01 with without with without
Lucia
with 9 b133 b7 b135
Lucia
without 0 b887 b4 b883
RKZ-42
Aug – Nov ’97
Lucia
with b6 b52 b5 b53
Lucia
without b9 272 10 273
RKZ-41
Aug – Nov ’97
Lucia
with 1 bb0 1 bb2
Lucia
without 4 361 4 359
1

27. Some p-values
• Cochran-Mantel-Haenszel test & Elffers’
post-hoc correction
1 in 916
• Ultimate stratification
11 days at JKZ with both incident & Lucia
on duty
1 in 25
• Gamma(1) heterogeneity over Poisson
intensity JKZ, RKZ pooled
1 in 25

28. Aftermath
• Since 2010, no more media interest
• The legal system got the blame,
the taxpayer paid the bill
• There have been reforms, improvements,
communication between legal and scientific
communities
• Medical community is silent

29. Interview with president
Council for Justice
• “The system worked fine”
• “Murderers who escape conviction usually
confess on their deathbed”

30. What really happened?
• In Dutch hospitals: 2000 deaths per year due to
avoidable medical errors; culture of denial; frequent
communication failures
• During 9 months up to 4 Sept. 2001, there was gossip
about Lucia among nurses and specialists
• Medical errors by specialists were being associated
with Lucia
• Director and top medical staff (but not all),
under oath: there was no suspicion till 4 Sept. 2001

31. • No suspicion at all till 4 September, 2001?
Director Paul Smits reported 10 unnatural deaths
and suspicious reanimations, over last year, within
15 minutes of being informed of death of Amber,
and on the very same day
• Strange fact: these 10 “incidents” were also
reported to Health Inspectorate. Conclusion:
nothing wrong.
• 4 medical specialists, it appears, have lied to police
and to courts (and to one another) concerning the
treatment of their own patients
What really happened?

32. Key case: baby Amber
• Baby Amber did not die of digoxin poisoning
• In fact the circumstances of her death are completely
consistent with a “natural” process
• Lucia did not have opportunity (doctors were with
the baby at the time when the court had determined
she must have acted)
• It seems there might have been digoxin in the body,
but it did not play any role in her death, and there
are many innocent explanations for how it got there
… if it was there at all

33. HCSK’s
• Once a hospital has “identified” a HCSK,
the suspect has no chance any more
• Lucia got accused through a combination of
unlucky coincidences
• She got exonerated through another
combination of lucky coincidences … and a
lot of very hard work of very many
“outsiders”

34. gill@math.leidenuniv.nl
http://www.math.leidenuniv.nl/~gill
Conclusions
• Forensic statistics is in its infancy
• It requires non-standard paradigms and will need
new methodology
• Multiparty statistics
• Nuisance parameters
• Model the forensic investigation process
• Communication of statistical ideas to non-
statisticians is going to be the bottle-neck