Arctic sea ice decline; are we sure it's anthropogenic?

8 views
Skip to first unread message

Fergus

unread,
Sep 10, 2007, 8:06:05 AM9/10/07
to globalchange
Not that I'm suggesting for amoment that it isn't, you understand.
This is more of a technical question.

Given all of the uncertainties relating to natural variability,
decadal cycles and the like, the ACIA in 2004 went no further than
saying (to paraphrase) 'This looks like an anthropogenically-induced
decline from GW, but we can't say so for certain'.

Things have moved fast since then, in terms of the rate of summer
decline. Has the process gone far enough yet for us to say,
definitively, that this must be an effect of AGW? Is there a numerical/
statistical analysis which places recent losses beyond the possible
bounds of natural variability + error?

Finally, is there any known reasonable alternative hypothesis for the
rate of change in the summer sea ice extent/area?

If it is clear that the uncertainty about the causes of Arctic sea ice
decline has diminished to the point of near certainty, we would then
be strongly placed to shout loudly on all blogs, ours and those of
skeptics; 'Where's the ice?'

William M Connolley

unread,
Sep 10, 2007, 8:32:41 AM9/10/07
to globalchange

On Mon, 10 Sep 2007, Fergus wrote:
> Things have moved fast since then, in terms of the rate of summer
> decline. Has the process gone far enough yet for us to say,
> definitively, that this must be an effect of AGW? Is there a numerical/
> statistical analysis which places recent losses beyond the possible
> bounds of natural variability + error?

Quite a few people seem to be pushing "the ice has declined faster than the
models predict" line. So that would appear to rule out anthropogenic factors as
the cause :-)

More seriously, I don't think "attribution" of ice decline is done in the way
that T changes are done. It seems to be more of the "look at this and look at
what we predicted" kind of thing.

-W.

William M Connolley | w...@bas.ac.uk | http://www.antarctica.ac.uk/met/wmc/
Climate Modeller, British Antarctic Survey | 07985 935400

--
This message (and any attachments) is for the recipient only. NERC is subject
to the Freedom of Information Act 2000 and the contents of this email and any
reply you make may be disclosed by NERC unless it is exempt from release under
the Act. Any material supplied to NERC may be stored in an electronic
records management system.

Fergus

unread,
Sep 10, 2007, 8:53:20 AM9/10/07
to globalchange
I get the impression that the line tends to be: 'look, we told you
this would happen if AGW kicks in...' , which sets up the sea ice
decline as an indicator of the credibility of the hypothesis and a
verification of model projections (but as you note...)

The problem I have is working out how we get from the 'what' to the
'why'. In one sense, it really is self-evident; warmer water, less
ice. It seems to make the Polar amplification idea look fairly solid,
too. But there must be a way of applying some kind of statistical tool
to the numbers to work out the significance of the change. There
should also be a way of calculating the ration of feedback to forcing,
with the information that is now available.

I know what you mean, though.

William M Connolley

unread,
Sep 10, 2007, 9:20:00 AM9/10/07
to globalchange

On Mon, 10 Sep 2007, Fergus wrote:
> I get the impression that the line tends to be: 'look, we told you
> this would happen if AGW kicks in...' , which sets up the sea ice
> decline as an indicator of the credibility of the hypothesis and a
> verification of model projections (but as you note...)

sure. in which case the models have failed, as the obs are outside their range
:-) or does only predicting too little count?

> The problem I have is working out how we get from the 'what' to the
> 'why'. In one sense, it really is self-evident; warmer water, less
> ice. It seems to make the Polar amplification idea look fairly solid,
> too. But there must be a way of applying some kind of statistical tool
> to the numbers to work out the significance of the change. There
> should also be a way of calculating the ration of feedback to forcing,
> with the information that is now available.

you can do sig tests to show that the change is not likely due to chance (cue
JA...). at which point you know immeadiately its due to... solar variation!

-w.

Eric Swanson

unread,
Sep 10, 2007, 9:45:20 AM9/10/07
to globalchange

On Sep 10, 8:32 am, William M Connolley <w...@bas.ac.uk> wrote:
> On Mon, 10 Sep 2007, Fergus wrote:
> > Things have moved fast since then, in terms of the rate of summer
> > decline. Has the process gone far enough yet for us to say,
> > definitively, that this must be an effect of AGW? Is there a numerical/
> > statistical analysis which places recent losses beyond the possible
> > bounds of natural variability + error?
>
> Quite a few people seem to be pushing "the ice has declined faster than the
> models predict" line. So that would appear to rule out anthropogenic factors as
> the cause :-)
>
> More seriously, I don't think "attribution" of ice decline is done in the way
> that T changes are done. It seems to be more of the "look at this and look at
> what we predicted" kind of thing.

I suppose you are referring to the paper which was published in the
GRL just last
Saturday and was given some recent press coverage. Here's the
citation:

Overland, J. E., and M. Wang (2007), Future regional Arctic sea ice
declines,
Geophys. Res. Lett., 34, L17705, doi:10.1029/2007GL030808.

Given that this year's sea-ice extent (and also area) is exhibiting a
very strong
decline, it would be easy to conclude that the models appear to be
understating
the problem. Perhaps there is some other mechanism involved with this
year's
decline which may not be included in the various models. If this
year's minimum
is just due to natural variability, we would see a return of more sea-
ice next
year, right? No worries, everything's OK, just keep moving!!

E. S.


Fergus

unread,
Sep 10, 2007, 9:56:47 AM9/10/07
to globalchange

On 10 Sep, 14:20, William M Connolley <w...@bas.ac.uk> wrote:
>
> you can do sig tests to show that the change is not likely due to chance (cue
> JA...). at which point you know immeadiately its due to... solar variation!
>

Can you also do sig tests to eliminate the solar component? IOW: do we
have enough data yet to match solar variability to sea ice
variability? (I am sure this actually involves far too many other
contingent variables, btw).

Eric: The Overland paper (abstract) reads a bit oddly in the light of
this year's decline (in that the estimates look too conservative).
Then you notice it was submitted back in May, so written before that.
It looks to me like a generally 'low end' estimate is still the
models' best guess, compared to observations, and the ensemble run has
'upped' the estimates of decline somewhat, but still appear to suffer
from some deficiency in their process; perhaps an underestimate of
feedbacks?

James Annan

unread,
Sep 10, 2007, 10:28:57 PM9/10/07
to global...@googlegroups.com
William M Connolley wrote:

> you can do sig tests to show that the change is not likely due to chance (cue
> JA...

...who will observe that significance tests can never answer questions
like "was it due to chance" (or not), but only questions like "how
likely is it that observations as extreme as these would arise in a
hypothetical world with no forcing"?

Which is not the same thing at all.

Here closeth the parenthesis.

:-)

James

Michael Tobis

unread,
Sep 10, 2007, 10:39:01 PM9/10/07
to global...@googlegroups.com
I'm at a conference with some statisticians at this moment. One talk
today discussed a recent paper in Nature which alleged that the
Chixculub meteor had a certain origin with probablity 90%. The gist of
the talk was that the professional statistician was unable to ascribe
enough meaning to the 90% claim to refute it.

Admittedly he was a "frequentist"; I had no idea how seriously people
take this division, but as a marginally statitistically educated
person with Bayesian sympathies, he left me unable to ascribe much
meaning to it myself.

Statistical attribution was always a red herring.

Global change isn't a drug trial and we can't round up 500 planets to
give half of them CO2 and half a placebo to get a 99% refutation of
the null hypothesis. We actually have to think, not just apply
formulas.

mt

James Annan

unread,
Sep 10, 2007, 11:56:02 PM9/10/07
to global...@googlegroups.com
Michael Tobis wrote:
> I'm at a conference with some statisticians at this moment. One talk
> today discussed a recent paper in Nature which alleged that the
> Chixculub meteor had a certain origin with probablity 90%. The gist of
> the talk was that the professional statistician was unable to ascribe
> enough meaning to the 90% claim to refute it.

http://www.nature.com/nature/journal/v449/n7158/full/nature06070.html

The paper itself is actually clear enough. If a particular event took
place as described, then it would have produced impacts at a much
greater rate than the background, such that any impact would with 90%
probability come from this event. I don't think it is unreasonable to
think about a single impact as a random sample from an "urn" of rocks
floating around in space. But on the face of it the research does not
justify the claims made in the press (or your phrasing above).

> Global change isn't a drug trial and we can't round up 500 planets to
> give half of them CO2 and half a placebo to get a 99% refutation of
> the null hypothesis. We actually have to think, not just apply
> formulas.

What I find interesting about it all is how little people care. It's not
as if I am the first person to think about it, indeed I am doing nothing
more than following a well-worn path (there are rants aplenty on this
general topic on the web). And yet...as Nature put it, "the concerns you
have raised apply more generally to a widespread methodological
approach" and therefore can safely be ignored.

James

Michael Tobis

unread,
Sep 11, 2007, 12:46:25 AM9/11/07
to global...@googlegroups.com
Yes, that's the one, thanks.

Since this isn't a public talk I won't identify the frequentist in
question, but he was uncomfortable with the very idea of assigning a
probability to an event that "either happened or didn't". Something
about babies and bathwater comes to mind.

That said, he also described a very long and involved set of
calculations that went into the figure, and pointed out that no effort
was made to assign confidence bounds to any of it.

I don't know of any claims about this paper in the press.

I have two very trusted independent sources who haven't the slightest
doubt that the Chicxulub impact was the dinosaur killer; I am not sure
they care very much which deep space event sent that gift our way. I
certainly can't get all that worked up about it. Although I'm a
worrier by nature I have a hard time getting too alarmed by potential
harm from Chicxulub Jr.

While I am mentioning Chicxulub, it is amazing that life survived the
event at all. That's based on the description I heard from
gbeophysicist Sean Gulick of Texas recently, who is working this up as
an outreach talk. Likely all remaining life descends from a few
survivors deep within caves which were immune to the huge temperature
swings.

mt

On 9/10/07, James Annan <james...@gmail.com> wrote:
>

James Annan

unread,
Sep 11, 2007, 1:23:32 AM9/11/07
to global...@googlegroups.com
Michael Tobis wrote:
> Yes, that's the one, thanks.
>
> Since this isn't a public talk I won't identify the frequentist in
> question, but he was uncomfortable with the very idea of assigning a
> probability to an event that "either happened or didn't". Something
> about babies and bathwater comes to mind.

I would be interested to know if he listens to (and acts upon) the
weather forecast :-) Tomorrow's weather is not a random repeatable
sample, merely an unknown deterministic event. Of course people
(including me) do talk about frequentist notions such as reliable
probabilities ("reliable" meaning that eg an event has historically
happened on p% of the occasions that it was forecast to happen with p%
probability), but I would hope that most if not all researchers would
agree if they thought about it carefully that in fact the probabilities
can only be Bayesian in nature.

>
> That said, he also described a very long and involved set of
> calculations that went into the figure, and pointed out that no effort
> was made to assign confidence bounds to any of it.
>
> I don't know of any claims about this paper in the press.

http://www.sciencedaily.com/releases/2007/09/070906135629.htm

"the team found a 90 percent probability that the object that formed the
Chicxulub crater was a refugee from the Baptistina family" is a rather
typical example. But it was unfair of me to criticise the press as the
claim appears in the paper itself. Of course my comments don't mean that
the 90% figure is unreasonable, only that it is not directly supported
by the research.

James

Tom Adams

unread,
Sep 11, 2007, 8:22:31 AM9/11/07
to globalchange
On Sep 11, 12:46 am, "Michael Tobis" <mto...@gmail.com> wrote:
> Yes, that's the one, thanks.
>
> Since this isn't a public talk I won't identify the frequentist in
> question, but he was uncomfortable with the very idea of assigning a
> probability to an event that "either happened or didn't".

(Opps!) Unconfortable with the very idea of assigning a 5/6
probability to your survival while playing Russian Roulette once with
a six-shooter, are you?

(I'll let you know when I feel another counter-example coming on.)

Probably, the unconfort you are feeling has to do with accepting the
guy's the model, not with assigning probability to a one-off when the
model is correct.

> Something
> about babies and bathwater comes to mind.
>
> That said, he also described a very long and involved set of
> calculations that went into the figure, and pointed out that no effort
> was made to assign confidence bounds to any of it.
>
> I don't know of any claims about this paper in the press.
>
> I have two very trusted independent sources who haven't the slightest
> doubt that the Chicxulub impact was the dinosaur killer; I am not sure
> they care very much which deep space event sent that gift our way. I
> certainly can't get all that worked up about it. Although I'm a
> worrier by nature I have a hard time getting too alarmed by potential
> harm from Chicxulub Jr.
>
> While I am mentioning Chicxulub, it is amazing that life survived the
> event at all. That's based on the description I heard from
> gbeophysicist Sean Gulick of Texas recently, who is working this up as
> an outreach talk. Likely all remaining life descends from a few
> survivors deep within caves which were immune to the huge temperature
> swings.

The End Permian Extinction was much bigger by all accounts. Only a
handful of proto-mammalian species survived. One of the proto-mammals
had such success afterward that it covered the continent (there was
only one back then) in big herds, a greater mammalian mono-culture
than our modern day herds of farm animals.

But I was not there 250 million years ago, I was just a gleam in the
eye of a Lystrosaurus.

>
> mt

> > James- Hide quoted text -
>
> - Show quoted text -

Tom Adams

unread,
Sep 11, 2007, 8:51:38 AM9/11/07
to globalchange

On Sep 11, 1:23 am, James Annan <james.an...@gmail.com> wrote:
> Michael Tobis wrote:
> > Yes, that's the one, thanks.
>
> > Since this isn't a public talk I won't identify the frequentist in
> > question, but he was uncomfortable with the very idea of assigning a
> > probability to an event that "either happened or didn't". Something
> > about babies and bathwater comes to mind.
>
> I would be interested to know if he listens to (and acts upon) the
> weather forecast :-) Tomorrow's weather is not a random repeatable
> sample, merely an unknown deterministic event. Of course people
> (including me) do talk about frequentist notions such as reliable
> probabilities ("reliable" meaning that eg an event has historically
> happened on p% of the occasions that it was forecast to happen with p%
> probability), but I would hope that most if not all researchers would
> agree if they thought about it carefully that in fact the probabilities
> can only be Bayesian in nature.

Sometimes these probabilities come from equivalencing the situation to
a model that can be understood via probability theory. No one
actually takes real samples for a real representation of the model,
but one already knows what the frequencies would be if one did.

But maybe there are some examples that can't be interpreted in this
manner?

Tom Adams

unread,
Sep 11, 2007, 10:18:36 AM9/11/07
to globalchange

On Sep 11, 1:23 am, James Annan <james.an...@gmail.com> wrote:

> Michael Tobis wrote:
> > Yes, that's the one, thanks.
>
> > Since this isn't a public talk I won't identify the frequentist in
> > question, but he was uncomfortable with the very idea of assigning a
> > probability to an event that "either happened or didn't". Something
> > about babies and bathwater comes to mind.
>
> I would be interested to know if he listens to (and acts upon) the
> weather forecast :-) Tomorrow's weather is not a random repeatable
> sample, merely an unknown deterministic event. Of course people
> (including me) do talk about frequentist notions such as reliable
> probabilities ("reliable" meaning that eg an event has historically
> happened on p% of the occasions that it was forecast to happen with p%
> probability), but I would hope that most if not all researchers would
> agree if they thought about it carefully that in fact the probabilities
> can only be Bayesian in nature.

I disagree to some degree. The weather prediction probabilities can
be (and are) model-based frequentist probabilities.

Now there is a hidden assumption: "The model fits reality". The
weatherman is basically acting as if he has a 100% degree of belief in
the model. The degree of belief in the model is perhaps Bayesian in
nature.

Sometimes I have heard local weathermen make a prediction different
from the national predicition. They have some understanding that
makes them doubt the local applicability of the general prediction.
Perhaps that's a lower degree of belief in the model.

I think this is the way it often works. The weather prediction is
obviously not purely Bayesian. It not like the weatherman (or some
committee) measures their psyche to determine a degree of belief.
They just commit to a model.

If this is not the way its done, then it should be done this way. Use
a mixed Bayesian/frequentist method. One thing that you should demand
is that the link between the model and probability be cut and dried:
nothing but pure math and (if real or simulated sampling is needed)
sound sampling procedures. All the fuzzy "degree of belief" stuff
should be confined to the "Does the model fit reality?" issue.

If the weatherman is allowing fuzziness to infect the model-
probability connection part, then he is not being a Bayesian, he is
just making a blunder. I have no doubt that this happen in various
applications, but its just a mistake, not a valid use of Bayesian
probability.

James Annan

unread,
Sep 11, 2007, 8:44:35 PM9/11/07
to global...@googlegroups.com
Tom Adams wrote:
>
>
> On Sep 11, 1:23 am, James Annan <james.an...@gmail.com> wrote:
>> Michael Tobis wrote:
>>> Yes, that's the one, thanks.
>>> Since this isn't a public talk I won't identify the frequentist in
>>> question, but he was uncomfortable with the very idea of assigning a
>>> probability to an event that "either happened or didn't". Something
>>> about babies and bathwater comes to mind.
>> I would be interested to know if he listens to (and acts upon) the
>> weather forecast :-) Tomorrow's weather is not a random repeatable
>> sample, merely an unknown deterministic event. Of course people
>> (including me) do talk about frequentist notions such as reliable
>> probabilities ("reliable" meaning that eg an event has historically
>> happened on p% of the occasions that it was forecast to happen with p%
>> probability), but I would hope that most if not all researchers would
>> agree if they thought about it carefully that in fact the probabilities
>> can only be Bayesian in nature.
>
> I disagree to some degree. The weather prediction probabilities can
> be (and are) model-based frequentist probabilities.

No.

Anyone can (and indeed frequently does) dress up a Bayesian probability
by generating an ensemble of outcomes to describe their posterior pdf.
But that doesn't make the underlying problem frequentist, it is just a
computationally and intuitively convenient method.

The standard paradigm of numerical weather prediction is that the
atmosphere is a deteministic system, which is imperfectly observed. Even
in the case of a perfect model, there is no such thing as the correct
probabilistic forecast (except perhaps pedants may point out the
degenerate case: the correct forecast is the [deterministic] output from
the perfect model run with perfect initial conditions, but we can never
hope to achieve this in reality).

The very best we could ever hope for, if there was a widely available
and agreed set of observations, is that all forecasters would generate
the same ("intersubjective") probabilities. Even this requires not only
a perfect model but also a universally agreed interpretation of all
observations, which is rather unlikely. It also requires a perfect, or
at least universally agreed, method for calculating probabilities, which
is whole other can of worms in itself.

The probability cannot be a function of the atmospheric state itself,
since the forecast will change if different observations are made. In
practice it is quite reasonable for different forecasters to give
different forecasts on any given day - and both can be "right" in the
sense of giving reliable forecasts in the long run.

James

James Annan

unread,
Sep 11, 2007, 8:52:40 PM9/11/07
to global...@googlegroups.com
Tom Adams wrote:
> On Sep 11, 12:46 am, "Michael Tobis" <mto...@gmail.com> wrote:
>> Yes, that's the one, thanks.
>>
>> Since this isn't a public talk I won't identify the frequentist in
>> question, but he was uncomfortable with the very idea of assigning a
>> probability to an event that "either happened or didn't".
>
> (Opps!) Unconfortable with the very idea of assigning a 5/6
> probability to your survival while playing Russian Roulette once with
> a six-shooter, are you?

Of course a frequentist would be uncomfortable with that idea: their
interpretation of probability does not apply to single events, only as
the limiting frequency of an infinite number of "identical" experiments.
Even this concept is rather hard to define, since in a deterministic
world identical experiments should give identical results.

(note that Michael was reporting someone else's views, not his own).

James

Tom Adams

unread,
Sep 12, 2007, 8:17:30 AM9/12/07
to globalchange

Seems pragmatic to interpret the Russian Roulette case as 5/6 based on
a frequentist thought experiment. No?

James Annan

unread,
Sep 12, 2007, 8:35:11 AM9/12/07
to global...@googlegroups.com

I certainly think so, but I'm not sure what your point is.

James

Tom Adams

unread,
Sep 12, 2007, 8:39:09 AM9/12/07
to globalchange
On Sep 11, 8:52 pm, James Annan <james.an...@gmail.com> wrote:
> Tom Adams wrote:
> > On Sep 11, 12:46 am, "Michael Tobis" <mto...@gmail.com> wrote:
> >> Yes, that's the one, thanks.
>
> >> Since this isn't a public talk I won't identify the frequentist in
> >> question, but he was uncomfortable with the very idea of assigning a
> >> probability to an event that "either happened or didn't".
>
> > (Opps!) Unconfortable with the very idea of assigning a 5/6
> > probability to your survival while playing Russian Roulette once with
> > a six-shooter, are you?
>
> Of course a frequentist would be uncomfortable with that idea: their
> interpretation of probability does not apply to single events, only as
> the limiting frequency of an infinite number of "identical" experiments.

So, a frequentist would be be just as bothered by a billion events as
by a single event. An infinite number of identical experiments is
always impossible.

Seems to me that it is a conceptual blunder to get hung up on this.

*All* applications of the frequentist interpretation involves
conterfactual conditions as does Newton's first law of motion and many
other useful concepts.

Michael Tobis

unread,
Sep 12, 2007, 8:58:26 AM9/12/07
to global...@googlegroups.com
Hey, I'm just reporting.

The guy was obviously mathematically sphisticated and a professional
statistician, but I couldn't really see how he could get any results
that don't violate his philosophy.

I was amazed to see that this is a real controversy in some circles.

What probability actually means (outside the purely mathematical
measure theory ideas without any connection to realisty) may be a bit
hard to pin down but it's obviously useful in cases other than picking
marbles out of an urn, if even that's permissble.

mt

Tom Adams

unread,
Sep 12, 2007, 9:44:15 AM9/12/07
to globalchange

On Sep 12, 8:58 am, "Michael Tobis" <mto...@gmail.com> wrote:
> Hey, I'm just reporting.
>
> The guy was obviously mathematically sphisticated and a professional
> statistician, but I couldn't really see how he could get any results
> that don't violate his philosophy.
>
> I was amazed to see that this is a real controversy in some circles.
>
> What probability actually means (outside the purely mathematical
> measure theory ideas without any connection to realisty) may be a bit
> hard to pin down but it's obviously useful in cases other than picking
> marbles out of an urn, if even that's permissble.

We build abstract models and we can prove everything about the models
using formal logic and mathematics. But do the models correspond to
reality? Answering this question is harder, can't be solved with just
logic and math.

This is a general problem, I am not sure why probability gets singled
out so often for special mention.

>
> mt


>
> On 9/12/07, Tom Adams <tadams...@yahoo.com> wrote:
>
>
>
>
>
> > On Sep 11, 8:52 pm, James Annan <james.an...@gmail.com> wrote:
> > > Tom Adams wrote:
> > > > On Sep 11, 12:46 am, "Michael Tobis" <mto...@gmail.com> wrote:
> > > >> Yes, that's the one, thanks.
>
> > > >> Since this isn't a public talk I won't identify the frequentist in
> > > >> question, but he was uncomfortable with the very idea of assigning a
> > > >> probability to an event that "either happened or didn't".
>
> > > > (Opps!) Unconfortable with the very idea of assigning a 5/6
> > > > probability to your survival while playing Russian Roulette once with
> > > > a six-shooter, are you?
>
> > > Of course a frequentist would be uncomfortable with that idea: their
> > > interpretation of probability does not apply to single events, only as
> > > the limiting frequency of an infinite number of "identical" experiments.
>
> > So, a frequentist would be be just as bothered by a billion events as
> > by a single event. An infinite number of identical experiments is
> > always impossible.
>
> > Seems to me that it is a conceptual blunder to get hung up on this.
>
> > *All* applications of the frequentist interpretation involves
> > conterfactual conditions as does Newton's first law of motion and many
> > other useful concepts.
>
> > > Even this concept is rather hard to define, since in a deterministic
> > > world identical experiments should give identical results.
>
> > > (note that Michael was reporting someone else's views, not his own).
>

James Annan

unread,
Sep 12, 2007, 10:56:03 PM9/12/07
to global...@googlegroups.com
Tom Adams wrote:
>
> We build abstract models and we can prove everything about the models
> using formal logic and mathematics. But do the models correspond to
> reality? Answering this question is harder, can't be solved with just
> logic and math.
>
> This is a general problem, I am not sure why probability gets singled
> out so often for special mention.

It's because, in a nutshell, it is specifically through the mechanisms
of probability that we connect these abstract models to practical
decision making (at least in the standard utility-maximising "rational"
paradigm).

James

James Annan

unread,
Sep 13, 2007, 3:19:27 AM9/13/07
to global...@googlegroups.com
Michael Tobis wrote:
> Hey, I'm just reporting.
>
> The guy was obviously mathematically sphisticated and a professional
> statistician, but I couldn't really see how he could get any results
> that don't violate his philosophy.

Well getting back to the origins of this sub-thread there would be no
problem in saying that for their particular experiment, 90% of the large
impacts originated from the event they simulated (with the rest coming
from the natural background). That's a perfectly ordinary frequentist
approach, and in fact this is precisely what they calculated.

> I was amazed to see that this is a real controversy in some circles.

Well, a lot of scientists (myself included) were brought up on a diet of
purely frequentist probability, and many of them also seem to be
uncomfortable with the idea of subjectivity in scientific judgements.
Hence uniform priors being defined as "ignorance" (where "ignorance" is
circularly defined as the state of mind described by a uniform
distribution...) and other such drivel.

But I hope this particular horse can be considered well thrashed now.
I'm going to a workshop on probability in climate science in a couple of
weeks, which several other climate scientists are slated to attend, so
it will be interesting to see if and how their thinking has progressed
in the 2 years(!) since I started talking about this sort of thing.

James

Janne Sinkkonen

unread,
Sep 13, 2007, 7:31:44 AM9/13/07
to globalchange
On Sep 12, 3:58 pm, "Michael Tobis" <mto...@gmail.com> wrote:
> I was amazed to see that this is a real controversy in some circles.

The bayesian vs. frequentist division is well alive, and will be until
the last frequentist is dead. :)

There are practicing frequentists who say that they are in principle
bayesian, but that because formulating one's prior in any more than
2-3 dimensions is impossible, bayesian methods are in practice
unusable. Pointing out the implicit priors behind frequentist methods
(confidence intervals, ridge regression), or suggesting that model
families are actually discrete priors does not influence their
opinion.

Meanwhile, bayesian methods are successfully used in machine learning
and in computational statistics with (kind of uninformative) priors
that are in practice confirmed empirically. And increasingly
hierarchical data forces bayesian thinking even on conservative fields
such as hypotheses testing in clinical research.

But even subjective probabilities are conditioned on some model. That
model needs to make sense and be explicit enough, if one wants the
probabilities to be taken seriously.

In the context of weather prediction etc., it is important to
understand that (1) bayesian techniques can be applied even when the
priors are not anyone's subjective beliefs; (2) bayesian models can be
tested and empirically validated just as any other models.

--
Janne

Michael Tobis

unread,
Sep 13, 2007, 11:53:04 AM9/13/07
to global...@googlegroups.com
The (in my opinion) valid criticism of the 90% number was that there
was a great deal of uncertainty in defining the urn. There was no
suggestion that the draw from the urn was illegitimate, only that the
urn was not a good model of reality.

The speaker went on to suggest that Bayesian methods overstate
confidence. I am not sure on what grounds he thought the asteroid
study was Bayesian, or whether he was simply pointing to a common
problem with statistical inference.

I challenged him afterward about decision-making under uncertainty.
It's my opinion that the hypothesis-testing view of statistics,
designed for climical experiments, is terribly misplaced in discussing
unavoidable decisions where the default is non-obvious.

If we can express only, say 3% confidence in X and say 0.08 %
confidence in not-X, we haven't used all the information available to
us. This hardly matters if the question X is purely theoretical. If we
are trying to decide whether to expend real resources on an asteroid
defense system, the 96.92 % probability, on this view, that
"statistics has nothing to say" is both ridiculous and unhelpful.
Surely we can conclude that X is more likely than not-X.

I understand that there is some philosophical problem with what
"likely" means, but it's hard for me to understand the idea that this
leads to a universal shrug and professed ignorance. At least you can
say these guys aren't cynically motivated in their attachment to the
idea that their work is very nearly useless.

mt

Michael Tobis

unread,
Sep 13, 2007, 4:55:23 PM9/13/07
to globalchange
Philip Stark of UC Berkeley is the frequentist in question. I pointed
him to this discussion and he replied as follows (forwarded with his
permission).

mt

---------- Forwarded message ----------
From: Philip B. Stark <...>
Date: Sep 13, 2007 1:24 PM
Subject: Re: frequentism discussion

Hi Michael--

Thanks for pointing me to this and for your interest.

My argument--applied to earthquake forecasts by the USGS--is in
this preprint:
http://statistics.berkeley.edu/~stark/Preprints/611.pdf

Slides from Monday's talk are here (in OpenOffice format):
http://statistics.berkeley.edu/~stark/Seminars/sandia07.odp

My points about the frequentist coverage probability of Bayesian
credible regions were not connected to the Chicxulub story...
Luis and I are planning to write something up for the conference
volume on that topic.

The gist of my argument about whether the chance the KT impactor
came from Baptistina Asteroid Family is that the probability
ultimately comes from stochastic assumptions about the collision
that formed the BAF (Gaussians for some things, truncated Gaussians
for some things, uniform for others, etc.), simplified simulations
about how orbits evolve to move objects into the resonances with
Jupiter and Mars, how objects are ejected from resonance, and on and
on, plus ad hoc corrections to astronomical catalogs, extrapolation
of empirical scaling laws well beyond the data, assumptions that
things like albedo, density and specific heat capacity are constant,
etc. There's a huge chain of physical and probabilistic assumptions
comprising an extremely complex--and tenuous--model. Buy the model,
buy the 90% figure. But why buy the model?

This is not like an urn model, and my skepticism is not because this
is a one-off event, or even that it either happened or didn't. (We
could rephrase it to be prospective, rather than retrospective,
and I'd have much the same problem with it.) My main complaint
is that the model is far fetched and only weakly tied to observation.

Best wishes,
Philip


On Thu, 2007-09-13 at 12:47 -0500, Michael Tobis wrote:
> There's an interesting discussion of reasoning under uncertainty on a
> public discussion list which I manage. I mentioned the Chixculub
> argument you presented at the Santa Fe workshop and that has proved an
> anchor of the discussion.
>
> However, I am doing a lousy job of defending your position because of
> some combination of not understanding it and/or not agreeing with it.
> So far nobody else has showed up to advocate for a frequentist
> perspective.
>
> You may at least be interested in reading the discussion so far. Your
> input would be most welcome.
>
> http://groups.google.com/group/globalchange/browse_thread/thread/cd703ecce94de4d4
>
> regards
> Michael Tobis
> http://www.ig.utexas.edu/people/staff/tobis/
--
Philip B. Stark | Professor of Statistics | University of California
Berkeley, CA 94720-3860 | 510-642-1430 | www.stat.berkeley.edu/~stark

Reply all
Reply to author
Forward
0 new messages