Accueil > Rubrique de services > Archive Equipe > Minh Ha Duong > Opinions > On the IPCC 4AR risk and uncertainty guidelines

On the IPCC 4AR risk and uncertainty guidelines

lundi 25 octobre 2004, par Minh Ha Duong

This is a comment on Guidance Notes for Lead Authors of the IPCC Fourth Assessment Report on Addressing Uncertainties as it appears in IPCC WG III (oct. 2004) Documents in support of the AR4 writing process, pp. 77-80. It is addressed to Martin Manning, Michel Petit and WGIII co-anchors Rob Swart at RIVM and H.-Holger Rogner at IAEA. There is a forum attached to this article if you want to reply in public.

Presently the difference between Table 3 and Table 4 expresses the opposition between objective and subjective probabilities. I believe it’s more important to express the difference between probabilities and possibility. This is why I would suggest to make Table 3 specify lower bounds on probabilities
 [1], instead of assuming implicitely that there is a scalar probability level called by another name. Experts would state "More than 2/3 chance" rather than "About 2/3 chance".

The quotes around "odds" in Table 3 could be removed by explaining more clearly Bruno de Finetti’s definition of subjective probabilities. The point of that definition is to rigorously define experts confidence using observable betting behavior rather than psychological interior feelings. Defining high confidence as there is more than 2/3 chance of X happening is not using odds. Using odds means to define high confidence as : the expert engage its credit saying that X will happen, and to shows the strength of that opinion he declares himself ready to take a bet on that at unfavorable odds. He would pay 2:1 if X does not happen.. [2] [3]

Something should be said beyond Table 3 and 4 to emphasis the distinction between about as likely as not, and no probability available. Assessing the result of a well balanced coin tossup is not exactly like assessing a chess match between two unknown players [4]. Is there a calibrated langage to describe situations of structural uncertainty ?

In paragraph 10 I have problems with the subjective caption ’level of confidence’ is for Table 2 because the table is objective. Both table’s axis, the number of independant sources of information and their degree of conflict, are objective characteristics of an information. The words defined by the table entries are appropriately objective, too. Clarification could be achieved by changing Table 2’s caption to something objective like "Factors of understanding", and then reordering the accompanying text paragraph 10.
- First describe Table 2 as a vocabulary to describe objectively on which grounds the assessment is based
- Then remind that the expert’s subjective confidence levels are to "consider the amount of evidence and the degree of consensus".
- Conclude that it’s up to the experts to find the balance between precision and confidence.

Table 3 desserves a numbered paragraph on its own.

Numbers in table 3 and 4 should be regarded as approximate (fuzzy numbers) because the meaning of words is vague. This is an additional layer of uncertainty. Imprecision in the communication channel is in addition to the imprecision of the message being communicated.

The "calibrated" words and numbers proposed by guideline 9 would look more attractive to my eyes if it was made clear that they come from scientific experimental studies (some are cited in the paper).

Voir en ligne : The IPCC Maynooth meeting

[1In Possibility theory, you would call that lower bound a level of necessity level. But I think the guidelines can’t introduce that kind of technical vocabulary. Statements that "the probability of X happening is greater than n percent" are easier to make for people educated only in probability theory.

[2Assume there is an unknown probability p that X happens. The expert receives 1 if X happens (probability p) but looses 2 if it does not (probability 1-p). The expected value of the bet is 3p - 2. Therefore, saying that p >= 2/3 is the same as saying that the expected value of the bet is positive. This is why, assuming that the experts are rational and accept only bets with positive expected value the statement in terms of betting odds is equivalent to the statement in terms of probability constraint.

[3Note that experts here are not asked to determine a fair rate at which they would be ready to bet for or against X happening, but only one-sided betting odds. This is because I suggested to use necessity levels in Table 3.

[4Some might argue that the two situations are the same from the decision making point of view. IPCC guidelines should not obfuscate the debate on precaution by presuming that policymakers are always neutral to ambiguity.


  • If likelihood is understood as objective probabilities, i.e. degrees of truth, then the highest degree of likelihood should not read "Virtually certain", but "True". The lowest is "False".

    As I explained elsewhere, the objective/subjective dimension is orthogonal to the precise/imprecise dimension (=probability/possibility theories). In the Guidelines, the possibilistic approach (bound on probability) is associated with the objective degree of likelihood scale, while the probabilistic approach goes with the degrees of certainty scale. I believe this is counter intuitive : a comitee was there !