Transcript 256kb

Two Solutions to the Main
Problem of Probabilistic
Causality: a Comparison
by Donald Gillies, University College
London
Contents
1. Introduction
2. Causality – Generic and Indeterminate
3. Probabilistic Causality and its Main
Problem
4. A Suggested Solution to the Main Problem
5. Pearl's Alternative Approach
Generic Causality
A causes B is said to be generic if it can be
instantiated on different occasions.
Example:
Throwing a stone over the cliff edge causes it to
fall into the sea.
(1)
Single-case Causality
A causes B is said to be single-case if it applies to
only one instance.
Example:
A heart attack caused Mr Smith's death.
(2)
Determinate Causality
A causes B is said to be determinate, if, whenever
A occurs, it is followed by B.
Example:
Throwing a stone over the cliff edge causes it to
fall into the sea.
(1)
Indeterminate Causality
A causes B is said to be indeterminate, if A can
occur without always being followed by B.
Example:
Smoking causes lung cancer.
(3)
(Only about 5% of smokers get lung cancer.)
Annual Death Rate for Lung
Cancer per 100,000 Men,
standardised for age
Non-smokers
10
Smokers
104
1-14 gms tobacco per day
52
14-24 gms tobacco per day
106
25 gms tobacco per day or more
224
(A cigarette is roughly equivalent to 1 gm of
tobacco)
Source: Doll and Peto, 1976, p. 1527.
Causality Probability Connection
Principle (CPCP)
If A causes B, the P(B | A) > P(B | ¬A)
(CPCP)
A Multi-Causal Fork
CPCP in the case of a MultiCausal Fork with Binary Variables
If X causes Z, then
P(Z = 1 | X = 1) > P(Z = 1 | X = 0)
(*)
Von Mises' Principle
FIRST THE COLLECTIVE – THEN THE
PROBABILITY
Two Reference Class Principles
FIRST THE REFERENCE CLASS – THEN THE
PROBABILITY
FIRST THE REFERENCE CLASS – THEN THE
CAUSALITY
Proposal for Solving Hesslow's
Problem
Divide S into two disjoint reference classes
S & (Y = 0) and S & (Y = 1).
Claim that (*) holds for each of these two
reference classes but not necessarily for S
itself.
Pearl on Probabilistic Causality
(PC)
“ ... the PC program is known mainly for the
difficulties it has encountered, rather than its
achievements. This section explains the main
obstacle that has kept PC at bay for over half a
century, and demonstrates how the structural
theory of causation clarifies relationships between
probabilities and causes” (2011, p. 714)
Need for the do-Calculus
“The way philosophers tried to capture this
relationship, using inequalities such as
P(E | C) > P(E)
was misguided from the start – counterfactual
'raising' cannot be reduced to evidential 'raising' or
'raising by conditioning'. The correct inequality,
according to the structural theory ..., should read:
P(E | do(C)) > P(E)
where do(C) stands for an external intervention
that compels the truth of C. The conditional
probability P(E | C) ... represents a probability
resulting from a passive observation of C, and
rarely coincides with P(E | do(C))
(Pearl, 2011, p. 715)
CPCP using do-Calculus
If X causes Z, then
P(Z = 1 | do(X = 1)) > P(Z = 1 | do(X = 0))
(**)
Latent Structural Causal Models
“If a set PAi in a model is too narrow, there will be
disturbance terms that influence several variables
simultaneously and the Markov property will be
lost. Such disturbances will be treated explicitly
as 'latent' variables ... . Once we acknowledge
the existence of latent variables and represent
their existence explicitly as nodes in a graph, the
Markov property is restored.”
(Pearl, 2000, p. 44)
Pearl on non-Markovian Models
“ ... we confess our preparedness to miss the
discovery of non-Markovian causal models that
cannot be described as latent structures. I do not
consider this loss to be very serious, because
such models – even if any exist in the
macroscopic world – would have limited utility as
guides to decisions.”
(2000, p. 62)
Pearl on Critics of the Markov
Assumption
“ ... they propose no alternative non-Markovian
models from which one could predict the effects of
actions and action combinations.”
(2000, p. 62)
Suggested Answer to this Challenge:
Multi-Causal Forks + Sudbury's Theorems