Our latest paper was just accepted by GRL. I think the contents may be of general interest, so I'll be blogging about it some time soon when I'm less tired and emotional. Those who are interested in the meantime will find it on my work web page.
> OK, so now I know why you are so passionately Bayesian.
What did you thinkk this was all about?
And here you have Bayesian reasoning in all its glory:
i) We have a subjective estimate for climate sensitivity of 1.5 - 4.5 (whihc we got by asking "the experts")
ii) observations do not agree with it, they give rather ...
iii) we combine i) and ii) to come up with a new estimate
If I would have made such an example James would have accused me of providing a caricature of the Bayesian approach.
PS: Now James can put me in his category "skeptic" and William can call me "septic". I would emphasize that my comment is NOT about the issue of climate change; And that I am all for understanding the physics of the atmosphere (and the system atmosphere + oceans + ... + economy) better.
Wolfgang - Are you sure you read the same paper I did? I didn't think your characterization was correct or fair. I admit being a bit baffled by the word subjective, but I thought the kind of evidence being brought to bear seemed real and substantial.
I read the first paper on James' list. He states clearly his problem: i) "Experts" know the value x to be in a certain range. ii) Observations give a different range for x. iii) Use Bayesian statistics to improve this range ii) so that it fits into the range i)
In all cases, the range of errors is large and we have no real understanding why the experts disagree. But we use the data produced by those experts anyway to improve our estimate.
The following is just one example:
Volcanic Eruptions (sect 3.2) Although *it might appear* that this information is already implicit in the 20th century reconstructions, *those papers generally did not consider* the short-term temperature changes in detail [..] Therefore we consider it *reasonable* to treat this constraint as a physically and observationally independent one.
In my example:
Although the two experts CIP and KingKong disagree about the factorization of 12 (CIP produced 12 = 3*4, KingKong produced 12 = 78*1234), it seems reasonable to consider them as observationally independent (because CIP used a pen, while KingKong ate a banana).
But I know that you are much more familiar with the topic and I am not. It is certainly possible that I misunderstood something. And I certainly do not want to get into a climate change debate. I leave this one for you and Lumo et al.
Yes you certainly do misunderstand something rather fundamental.
The comment that ii) Observations give a different range for x. iii) Use Bayesian statistics to improve this range ii) is not valid or meaningful. It is only through a bayesian viewpoint of uncertainty that observations provide any "range" (ie probabilistic estimate) in the first place.
What we have done is show how this analysis should be performed, in particular we have illustrated how deliberately ignoring some evidence when forming the estimate will generate an unrealistically high confidence interval which does not really describe our uncertainty. Of course in general theoretical terms this fact is well known, but it's not been properly considered or quantified in this context before.
> The comment that i)[..] ii)[..] is not valid or meaningful.
James,
what you say is not valid or meaningful is an accurate if not verbatim representation of your own introduction! I was careful to condens your own statements into my 3 points i) - iii)
I quote from your own paper (read it to fill in the [..]), the first sentences of 1. Introduction:
i) A subjective estimate [..] is likely to lie in the range of 1.5 - 4.4C [..].
ii) [..] there has been an increasing focus on the potential of observationally-derived constraints [..] they have concluded that the upper limit is difficult to constrain, [..] as high as 6C and many reaching even higher levels.
iii) In this paper, we show that when this (Bayesian analysis) is performed, the uncertainty [..] can be greatly reduced.
If you disagree with me about what your own paper states, we should end this discussion right here.
PS: My own conclusion iii) would of of course be something like:
The large range of expert opinion [whether it is "subjective" (i) or "objective" (i) ] shows clearly that we are dealing with inconsistent scenarios. Rather than to contemplate the meaning of the word "uncertainty", we should identify the root cause for these discrepancies and establish a consistent theory of climate change.
The problem with your point (ii) and how it compares to my version, is that even when "observationally-derived" constraints are generated, they still also necessarily depend on a prior which is necessarily subjective. That's simply how it works, and this is why I've spent the last week or two trying to persuade you to work yourself through a real example raher than just waffling.
For further clues, note that in particular, the 20th century warming constraint is presented as a probability distribution based on a uniform prior, and the other constraints are merely likelihood functions.
12 comments:
Now you can stop teasing people at sci.env!
Congratulations :)
It seems pretty important to me.
crandles
OK, so now I know why you are so passionately Bayesian. Maybe I will understand BP after I read your paper.
> OK, so now I know why you are so passionately Bayesian.
What did you thinkk this was all about?
And here you have Bayesian reasoning in all its glory:
i) We have a subjective estimate for climate sensitivity of 1.5 - 4.5 (whihc we got by asking "the experts")
ii) observations do not agree with it, they give rather ...
iii) we combine i) and ii) to come up with a new estimate
If I would have made such an example James would have accused me of providing a caricature of the Bayesian approach.
PS: Now James can put me in his category "skeptic" and William can call me "septic".
I would emphasize that my comment is NOT about the issue of climate change; And that I am all for understanding the physics of the atmosphere (and the system atmosphere + oceans + ... + economy) better.
And I apologize again for my typos.
Wolfgang - Are you sure you read the same paper I did? I didn't think your characterization was correct or fair. I admit being a bit baffled by the word subjective, but I thought the kind of evidence being brought to bear seemed real and substantial.
CIP,
I read the first paper on James' list.
He states clearly his problem:
i) "Experts" know the value x to be in a certain range.
ii) Observations give a different range for x.
iii) Use Bayesian statistics to improve this range ii) so that it fits into the range i)
In all cases, the range of errors is large and we have no real understanding why the experts disagree. But we use the data produced by those experts anyway to improve our estimate.
The following is just one example:
Volcanic Eruptions (sect 3.2)
Although *it might appear* that this information is already implicit in the 20th century reconstructions, *those papers generally did not consider* the short-term temperature changes in detail [..]
Therefore we consider it *reasonable* to treat this constraint as a physically and observationally independent one.
In my example:
Although the two experts CIP and KingKong disagree about the factorization of 12 (CIP produced 12 = 3*4, KingKong produced 12 = 78*1234),
it seems reasonable to consider them as observationally independent (because CIP used a pen, while KingKong ate a banana).
But I know that you are much more familiar with the topic and I am not. It is certainly possible that I misunderstood something.
And I certainly do not want to get into a climate change debate. I leave this one for you and Lumo et al.
Wolfgang,
Yes you certainly do misunderstand something rather fundamental.
The comment that
ii) Observations give a different range for x.
iii) Use Bayesian statistics to improve this range ii)
is not valid or meaningful. It is only through a bayesian viewpoint of uncertainty that observations provide any "range" (ie probabilistic estimate) in the first place.
What we have done is show how this analysis should be performed, in particular we have illustrated how deliberately ignoring some evidence when forming the estimate will generate an unrealistically high confidence interval which does not really describe our uncertainty. Of course in general theoretical terms this fact is well known, but it's not been properly considered or quantified in this context before.
> The comment that
i)[..]
ii)[..]
is not valid or meaningful.
James,
what you say is not valid or meaningful is an accurate if not verbatim representation of your own introduction! I was careful to condens your own statements into my 3 points i) - iii)
I quote from your own paper (read it to fill in the [..]), the first sentences of 1. Introduction:
i) A subjective estimate [..] is
likely to lie in the range of 1.5 - 4.4C [..].
ii) [..] there has been an increasing focus on the potential of observationally-derived constraints [..] they have concluded that the upper limit is difficult to constrain,
[..] as high as 6C and many reaching even higher levels.
iii) In this paper, we show that when this (Bayesian analysis) is performed, the uncertainty [..] can be greatly reduced.
If you disagree with me about what your own paper states, we should end this discussion right here.
PS: My own conclusion iii) would of of course be something like:
The large range of expert opinion [whether it is "subjective" (i) or "objective" (i) ] shows clearly that we are dealing with inconsistent scenarios.
Rather than to contemplate the meaning of the word "uncertainty", we should identify the root cause for these discrepancies and establish a consistent theory of climate change.
Wolfgang,
You really still dont get this do you?
The problem with your point (ii) and how it compares to my version, is that even when "observationally-derived" constraints are generated, they still also necessarily depend on a prior which is necessarily subjective. That's simply how it works, and this is why I've spent the last week or two trying to persuade you to work yourself through a real example raher than just waffling.
For further clues, note that in particular, the 20th century warming constraint is presented as a probability distribution based on a uniform prior, and the other constraints are merely likelihood functions.
> You really still dont get this do you?
I think I do.
How is the espresso at Starbucks ?
Post a Comment