Formal tools for handling evidence

1
Formal tools for handling evidence Dr Valentina
Leucari

• Discussion by Dr Mike Joffe

2
Strengths of the paper

• comparing and contrasting different types of
  graphical method is useful Bayesian networks
  and Wigmore charts here
• decomposing large models into smaller component
  ones is very useful
• the idea of having modular components what the
  paper calls recurrent structures of evidence
  could be extremely valuable

3
Fundamental attributes

• Schum
• relevance
• (2) credibility
• (3) strength
• p18 credibility (2) is represented by
  likelihood ratios (3)
• p33 no conditional independence (2) in Wigmore
  charts, but still notion of relevance (1)
• does this mean that strength (3) corresponds to
  value of information (1)?

• Bayesian networks
• value of information
• (2) conditional probability tables
• (3) likelihood ratio

4
Explaining away where knowledge of one being
true lowers the probability of the other being
true
X3
X2
X4
X3 Sacco was involved in other crimes X2 Sacco
was involved in this crime X4 Sacco intended to
escape from the police when they arrested him
5
Potential problems I

• there could be other reasons for X4 (apart from
  the policeman making this up), e.g. a
  generalisation everyone fears the police, or
  immigrants are treated as suspect or Saccos
  political activity caused him to fear arrest, or
  Sacco was paranoid
• if Sacco were involved in other crimes (X3), this
  might increase not decrease the level of
  suspicion re this crime (X2).

6
Potential problems II

• also, its not clear from the charts what the
  process is e.g. the chart for explaining away
  (fig 3.11) looks similar to that for the filter
  fragment (fig 3.9) and to that for
  Event/Competence/Sensation (fig 3.7) etc
• and how would this be interpreted in the case of
  a causal model? interpretation is clear when
  its a case of a belief making another less
  likely (p6 A model can be causal)

7
Different languages I

• conditional probabilities (or more generally,
  joint distributions) are in themselves
  non-directional a direction (arrow-head) is
  only present if imposed (e.g. blocks in graphical
  models)
• are we dealing with objective causal relations
  here, or subjective belief systems? or both?
  Bayesian networks are a process model
  intended to capture a complex process by which
  some series of events could have been generated
  (Schum 2005 see p33)

8
A typology of graphical methods

• THOUGHT PROCESSES generalisability requires
  justification of structure, links, etc
• BAYES NETS correct subjective beliefs about
  objective/quve causal relationships
• CAUSAL RELATIONSHIPS arrows that represent
  causation in the world combines a priori
  specification and empirical testing
• STATISTICAL ASSOCIATIONS represent joint
  distributions only links non-directional

9
Different languages II

• different languages (here verbal reasoning and
  the laws of probability) the questions are
• (a) when and how to use each of them
• (b) how to interrelate them to manage boundary
  compatibility

10
Grouping of items of evidence

• the suggestion in section 6.2 (page 37) that
  items of evidence could be grouped into Witness
  evidence, Physical evidence and Consciousness of
  guilt evidence is problematic it would be
  better to group items according to sub-stories,
  e.g. whether the suspect was present, who hit the
  guard, etc
• more broadly, I would welcome the idea of
  alternative stories being made more explicit

11
How special is Law?

• in this programme, we have focused a great deal
  on legal examples what special considerations
  does this introduce?
• we should consider using journalism as a focus
  it is more rooted in commonsense science,
  undistorted by arbitrary rules of evidence
  (although in practice distorted by
  commercialism), and not constrained to the guilt
  or not of particular people
• responsibility not culpability