Back off man, I'm a scientist.. Well, ok, I'm not.
Well, I think I will throw my hat in the ring. I don't think I'm ideologically inclined to dismiss man-caused global warming, and I think science has lots of authoritative weight. Furthermore, I do think it is more likely than not that man-made forces are causing at least some global warming. That said, my subjective probability is only a little over 50% (In technical terms I am talking about the posterior distribution that is the product of Bayesian updating). One of the reasons my updating (the term for the change in probability caused by seeing the evidence ... in this case the authoritative consensus of scientists) isn't stronger is because of the way some combination of the press and the scientists interviewed by the press (at least the ones I've heard) talk about the issue. I haven't and I don't want to dig into the underlying literature (assuming I could even understand it properly); I just don't care enough about the issue. And while the discussion in this thread (from my skimming of it) is lucid and clear enough, I don't think anything in the thread cleared up why it is difficult for me to do some kind of Bayesian updating with regard to the probability of man-cause global warming.
It has been reported widely by the press, by scientists in interviews, and by Leb in this thread that there is overwhelming consensus for man caused global warming. I think that's important, and it's the primary reason (because of its authoritative weight) why I think man cause global warming is more likely than not. Yes, I am skeptical that the consensus number is 97%, but I would be shocked if it wasn't very high (my guess is the true number is at least 90%). Still, it's the wrong number. Yes, it is correlated with the right number (so it's certainly not irrelevant), but it's still the wrong number. And, other posters in the thread have noted that it is the wrong number: most notably SU and Creekster. it's not the percentage of scientist that believe in man caused global warming that's important, but rather their certainty with respect to the cause. If I'm a scientist and I'm asked whether I believe in man caused global warming, then I would answer yes, if I think the probability is greater than 50%. So what scientists should publish (as a consensus document) is a description of their probability distribution (in fact in science, there is a precise name for it: a posterior distribution ... in a dream world I would get a discussion of the prior distribution as well).* Furthermore, not all 'cause' is equal. Really, I want to know the probability that it's a meaningful or significant causer in terms of magnitude (not significant in a statistical sense). Of course, what is actually a big enough magnitude is a mushy boundary but conceptually it is not a hard concept.
So what I am looking for is something that describes the probability distribution of scientists after they have weighed the evidence. But, for whatever reason, some combination of the press and scientists have reduced that to a simple statement about binary consensus. The problem is I essentially have to infer the consensus probability distribution from the binary consensus statement. For example, does it make sense to assume a uniform distribution from 51%-99% among the 97% who believe in man-cause global warming and uniform distribution from 1%-49% for the 3% that don't? Probably not, but I don't have a great sense for what the distribution should look like so I am left to guess (so that's basically what I've done). Look, talking about the posterior distribution can get technical very fast; I understand that (which is way I don't view the collapse to only reporting a binary consensus as dodgy in any way). But I think scientists can give a far better sense of their posterior distributions without much additional complexity.
So when I start my updating using scientific consensus with high authoritative weight I don't update to a super high probability (that man-caused global warming is the truth). Also, I think there are additional things that further reduce my probability. For example, the dispersion in the model estimates of temperature trends over time have pretty wide bands. That suggests there is a significant amount of uncertainty (but I am guessing a bit). Furthermore, we are in a sub-field with limited ability to experimentally manipulate (the bread and better of causal inference). Well, maybe a better way to put it would be to say, that it appears the sub-field has some important limitations on how generalizable the inferences are from experimental manipulation. This doesn't mean I throw my hand up and proclaim it's all hokum. But I do lower my subjective probability some.
I think the peer review process is really quite noisy. Incentives of people are notorious mixed and often political (in the micro sense of field ... academic politics can be awful just like work place politics). Furthermore, recently we've seen problematic reports related to the number of experimental results that can be reproduced even in the best journals. I am not proposing a direction to the bias here, but noise undermines at least my probabilistic confidence (it's not a big hit but it does matter). I do agree that there is an incentive for scientist to try to overturn consensus or orthodoxy. But I don't think that's true for the mode or median scientist. I actually think that incentive is only strong for a relatively small part of the distribution. Yes, if you view yourself as a young elite (and ambitious) scientist, you might want to try to overturn consensus; but it is risky. A publishing strategy that deals with smaller issues may be much more likely to pay off in terms of gaining tenure and a solid reputation.
I view the reported data scandals in much the same way. They shouldn't cause people to throw their hands up and once again exclaim it's all hokum (they don't imply the probability is zero) but it strikes me as rather rational to lower our probabilistic confidence at least a little. Additionally, I think the median political beliefs of scientists enter in much the same way. I don't think scientists are generally driven by political ideology. But, I do think it matters on the margin at least a little. This causes me to update less strongly towards a man-made cause; but only a little less strongly.
(Also, there are some technical things related to the nature of my Bayesian prior that I won't go into to that leads to less updating.)
Given these issues, I think it is pretty defensible to be on the below 50% divide of the debate.
footnotes
----
**Note, one thing I don't want is the 'confidence intervals' from the models ... those are related but technically shouldn't be same as the probability of significant in magnitude man-caused global warming given the evidence.
Well, I think I will throw my hat in the ring. I don't think I'm ideologically inclined to dismiss man-caused global warming, and I think science has lots of authoritative weight. Furthermore, I do think it is more likely than not that man-made forces are causing at least some global warming. That said, my subjective probability is only a little over 50% (In technical terms I am talking about the posterior distribution that is the product of Bayesian updating). One of the reasons my updating (the term for the change in probability caused by seeing the evidence ... in this case the authoritative consensus of scientists) isn't stronger is because of the way some combination of the press and the scientists interviewed by the press (at least the ones I've heard) talk about the issue. I haven't and I don't want to dig into the underlying literature (assuming I could even understand it properly); I just don't care enough about the issue. And while the discussion in this thread (from my skimming of it) is lucid and clear enough, I don't think anything in the thread cleared up why it is difficult for me to do some kind of Bayesian updating with regard to the probability of man-cause global warming.
It has been reported widely by the press, by scientists in interviews, and by Leb in this thread that there is overwhelming consensus for man caused global warming. I think that's important, and it's the primary reason (because of its authoritative weight) why I think man cause global warming is more likely than not. Yes, I am skeptical that the consensus number is 97%, but I would be shocked if it wasn't very high (my guess is the true number is at least 90%). Still, it's the wrong number. Yes, it is correlated with the right number (so it's certainly not irrelevant), but it's still the wrong number. And, other posters in the thread have noted that it is the wrong number: most notably SU and Creekster. it's not the percentage of scientist that believe in man caused global warming that's important, but rather their certainty with respect to the cause. If I'm a scientist and I'm asked whether I believe in man caused global warming, then I would answer yes, if I think the probability is greater than 50%. So what scientists should publish (as a consensus document) is a description of their probability distribution (in fact in science, there is a precise name for it: a posterior distribution ... in a dream world I would get a discussion of the prior distribution as well).* Furthermore, not all 'cause' is equal. Really, I want to know the probability that it's a meaningful or significant causer in terms of magnitude (not significant in a statistical sense). Of course, what is actually a big enough magnitude is a mushy boundary but conceptually it is not a hard concept.
So what I am looking for is something that describes the probability distribution of scientists after they have weighed the evidence. But, for whatever reason, some combination of the press and scientists have reduced that to a simple statement about binary consensus. The problem is I essentially have to infer the consensus probability distribution from the binary consensus statement. For example, does it make sense to assume a uniform distribution from 51%-99% among the 97% who believe in man-cause global warming and uniform distribution from 1%-49% for the 3% that don't? Probably not, but I don't have a great sense for what the distribution should look like so I am left to guess (so that's basically what I've done). Look, talking about the posterior distribution can get technical very fast; I understand that (which is way I don't view the collapse to only reporting a binary consensus as dodgy in any way). But I think scientists can give a far better sense of their posterior distributions without much additional complexity.
So when I start my updating using scientific consensus with high authoritative weight I don't update to a super high probability (that man-caused global warming is the truth). Also, I think there are additional things that further reduce my probability. For example, the dispersion in the model estimates of temperature trends over time have pretty wide bands. That suggests there is a significant amount of uncertainty (but I am guessing a bit). Furthermore, we are in a sub-field with limited ability to experimentally manipulate (the bread and better of causal inference). Well, maybe a better way to put it would be to say, that it appears the sub-field has some important limitations on how generalizable the inferences are from experimental manipulation. This doesn't mean I throw my hand up and proclaim it's all hokum. But I do lower my subjective probability some.
I think the peer review process is really quite noisy. Incentives of people are notorious mixed and often political (in the micro sense of field ... academic politics can be awful just like work place politics). Furthermore, recently we've seen problematic reports related to the number of experimental results that can be reproduced even in the best journals. I am not proposing a direction to the bias here, but noise undermines at least my probabilistic confidence (it's not a big hit but it does matter). I do agree that there is an incentive for scientist to try to overturn consensus or orthodoxy. But I don't think that's true for the mode or median scientist. I actually think that incentive is only strong for a relatively small part of the distribution. Yes, if you view yourself as a young elite (and ambitious) scientist, you might want to try to overturn consensus; but it is risky. A publishing strategy that deals with smaller issues may be much more likely to pay off in terms of gaining tenure and a solid reputation.
I view the reported data scandals in much the same way. They shouldn't cause people to throw their hands up and once again exclaim it's all hokum (they don't imply the probability is zero) but it strikes me as rather rational to lower our probabilistic confidence at least a little. Additionally, I think the median political beliefs of scientists enter in much the same way. I don't think scientists are generally driven by political ideology. But, I do think it matters on the margin at least a little. This causes me to update less strongly towards a man-made cause; but only a little less strongly.
(Also, there are some technical things related to the nature of my Bayesian prior that I won't go into to that leads to less updating.)
Given these issues, I think it is pretty defensible to be on the below 50% divide of the debate.
footnotes
----
**Note, one thing I don't want is the 'confidence intervals' from the models ... those are related but technically shouldn't be same as the probability of significant in magnitude man-caused global warming given the evidence.
Comment