A Bayesian and Toulminian Critique of the IARC Carcinogen Classification

1
A Bayesian and Toulminian Critique of the IARC
Carcinogen Classification

• Raymond Richard Neutra MD Dr.PH

2
A Basic Policy Question

• All other things being equal how certain must we
  be of how much ill-health before we would opt for
  cheap or expensive protective alternatives?

3
GENERIC DECISION TREE
Degree of Certainty
Remed.
results
results
Do Risk Eval.
No Remed
Problem presents
results
Remed.
Dont Do Risk Evaluation
results
No Remed.
4
How California EMF Risk Assessment Process
Differed From IARCs

• A risk evaluation guideline was prepared,
  submitted to public comment, revised and approved
  by an external advisory committee
• Risk Assessment done by government scientists
  submitted to public comment and approved (with
  reservations) by external advisory committee
• All meetings open to public

5
Transparency Used Degree of Certainty, not
Yes/No
6
Probability, Certainty or Willingness to Certify?

• Probability implies some Platonic quality of
  the real world.
• Certainty has a taste of arbitrary opinion.
• Virtually certain that one can certify that..,
  prone to certify that.. Implies some orderly
  agreed upon process behind the certifying claim.
• I might be certain that God exists but would
  not certify that he does.

7
PROBLEM WITH IARC CLASSIFICATION

• How probable is Class 1 B probable
  carcinogen?
• What does Class 2 B possible carcinogen mean?
  Not impossible? Just short of more likely than
  not? Coffee and Fiberglass are both in 2B.
• Committee diversity suppressed
• Hinges on adjectives about likelihood of bias and
  confounding in Epidemiology

8
A Nature of Evidence Classification
Masquerading as a Probability of Hazard
Classification

• I Clear Tox Clear Epi
• II a Clear Tox Unclear or Absent Epi
• II b Unclear or Absent Tox Mod. Epi
• III Insuf. Volume or Conflicting Evid.
• IV Both Tox Epi clear null result

9
GuidelinesPolicy Neutral Language to Describe
patterns of Evidence in the Grounds

• An effect near the resolving power of the study
  not a weak effect
• An effect well above the resolving power of the
  study not a big, or robust effect
• An absent evidentiary base not there is no
  evidence that

10
Assumes 11 Link Between Evidence Pattern and
Probability of Hazard

• Are there some agents for which a null or pos.
  tox result is less persuasive than for other
  agents?
• a) wrong animal model
• b) dose response not linear at hi dose
• c) testing wrong ingredient of mixture

11
So What Warrants the Leap from Patterns of
Evidence to a Willingnes to Certify Causality?

• 17th century hope for mathematical a-contextual
  deductions of certainty do not work in biology
• Toulmins argument theory and probability theory
  help distinguish reasonable and unreasonable
  general inferential rules that warrant
  conclusions.

12
                           
 
BACKING Studies blood enzymes are higher in MI
patients than in those with other causes of chest
pain
Warrant If a patients enzymes are high and he
has chest pain, one can increase ones
certainty that he has (MI)
Grounds On the basis of age and weight one
suspects MI a priori.   The patients blood
enzymes are elevated.   There is chest pain
Claim   One should increase ones degree of
certainty that patients chest pain is caused by
MI as opposed to other causes.
Rebuttal All other evidentiary tests being equal
13
Bayes Theorem, the Universal Warrant

• Posterior Odds
• Prior Odds Relative Likelihood of Evidence

14
Possible Test Results for Relevant Streams of
Evidence

• Streams Biophysics, in vitro, physiology,
  toxicology and epidemiology
• Say each stream could yield a clear effect, weak
  effect or no effect of EMFs (3 outcomes)
• 33333243 possibilities

15
A Bayesian Analysis of the Conservative Warrant
Only believe if all streams yield clear effect.

• Relative Likelihood conveyed by meeting this
  standard is a very small number divided by an
  extreeeeemly small number. Ratio is very large.
• Relative Likelihood conveyed by missing this
  standard is a large number divided by another
  large number. The ratio is close to 1.00
• This warrant is assymetric, prone to falsely
  exonerate but not prone to falsely incriminate.

16
Promising Too Much

• Both Kochs Postulates and Bradford Hills
  Criteria have come, over time, to be presented
  as dichotomous tests to distinguish causal from
  non-causal agents.
• I will re-examine them in the light of Toulmins
  argument theory and Bayes probabilistic
  induction and argue that they should be viewed as
  good rule-in tests but poor rule-out tests.

17
Kochs Postulates

• He never listed them as such, others made lists.
• He was following earlier recommendations by Henle
  and Klebs in 1875 when he identified B Anthracis
  in diseased animals, obtained pure cultures in
  bovine Aqueous Humor and produced illness and
  death and demonstrable infestation with B
  Anthracis in Rabbits and Mice.

18
Motivation of Kochs Steps

• Demonstrate specific agent in tissues of all
  cases of the disease. (High Likelihood of a
  specific agent among cases)
• This agent has zero likelihood in healthy animals
  (not an incidental finding)
• This agent has zero likelihood in cases of other
  disease ( this species of bacteria causes a
  specific disease because of its unique physiology
  and is not an opportunistic invader after other
  disease processes are under way.)

19
Koch himself did not require all steps to be
satisfied.

• Cholera injected into animals did not cause all
  of them to be sick.
• Pettenkofer drank a preparation of cholera and
  was not sick though his stools had commas
• But Cholera was present in all patients with the
  syndrome and not in healthy persons or in other
  diseases. It could be cultured.
• Altona with sand filtration had lower rates of
  Cholera than Hamburg without filtration.(Accepted
  an ecological study as proof)

20
Hills Aspects of Association

• Strength, (yes, no)
• Consistency (yes, no)
• Specificity (yes, no)
• Temporality (yes, no)
• Dose Resp. (yes, no)

• Plausibility (yes, no)
• Experim. Ev. (yes, no)
• Coherence (yes, no)
• Analogy (yes, no)

21
Initial Graded Exposure Occurs
Plausible intermediate Steps
Associated with Epoch
Not Associated With Unexpected Disease
Subsequent Disease is Related To dose and
Coherent With Epoch
22
Tests for an Agent/Syndrome Association

• Tests of Causal Model
• Expected dose/response
• Temporality
• Plausibility
• Specificity
• Coherence
• Analogy

• Tests of the Pedigree of the Evidence
• Consistency
• Strength of Assoc.
• Homogeneity
• Kochs Pure Culture
• Kochs animal model experiment
• Cause disease
• Prevent disease

23
Incomplete Warrants

• Criterion, If OR is strongI am warranted to
  believe in causality is incomplete.
• What am I warranted to believe if OR is moderate
  or very near 1.0?
• What is the backing for these warrants?

24
A COMPLETE WARRANT FOR STRENGTH

• Test What relative risk is conveyed by the
  evidence for this hypothesis?
• If far larger than known effects of confounders
  and bias, am warranted to increase belief a lot
• If about the size of confounders and bias, am
  warranted to maintain prior belief
• If close to 1.00, am warranted to decrease belief
• Backing Size of bias/conf. in previous studies

25
A Complete Warrant for Kochs Produce Disease by
Injection

• Test What happens when I inject a pure culture
  into a test animal?
• I am warranted to increase my degree of certainty
  of causality if the animal falls ill or dies and
  its diseased tissues are swarming with the
  agent.
• My degree of certainty of causality decreases
  only slightly if the animal remains healthy.
• Backing Previous experience of likelihood of
  post injection disease in recognized human
  pathogens and recognized agents that are not
  human pathogens.

26
COHERENCE A COMPLETE NON-DICHOTOMOUS WARRANT

• Coherence Test To what degree does the evidence
  support predicted subsidiary hypotheses of this
  vs. other hypotheses?
• Warrant Increase or decrease your willingness to
  certify each hypothesis according to the relative
  likelihood of the evidence for their subsidiary
  hypotheses.

27
Display Arguments of Defense, Prosecutor and Judge
28
Consistent Evidentiary Tests Etiological
Balancing

29
What is the relative likelihood conveyed by the
combined test results?

• Not unlike the problem of medical diagnosis What
  is the relative likelihood conveyed by all the
  evidence in the history, physical and battery of
  lab tests?
• Not yet possible to computerize diagnosis or risk
  assessment to provide posterior odds free of
  differences in professional judgment.

30
Where professional judgement enters IARC
classification

• In judging how to use these phrases
• bias and confounding can be reasonably excluded
• limited
• sufficient
• Combining judgments on different streams of
  evidence into an overall judgment.

31
Display Disagreements on Prior and Posterior
Certainties
32
A Willingness to Certify Scale