Friday, May 02, 2008

Re: Predictions and Apocalypse   [Jim Manzi]

Chris Horner raises some great issues in his last two posts.  Here are a few comments:

1. Projections and predictions are distinct in UN IPCC terminology. Think of a long-term climate prediction as requiring (in a grossly simplified illustration) two ingredients: (i) a scenario for carbon emissions, and (ii) a Global Climate Model (GCM) that can convert this emissions scenario to an estimate for future climate change (e.g., temperature increase). In this terminology a projection is a prediction for future climate change given a specific emissions scenario. The reason that this is a useful distinction is that the atmospheric scientists are expert in projecting future climate impacts of emissions, not in predicting future population growth, changing patterns of energy use, etc. When evaluating the accuracy of a model, it is sensible to keep track of these two different sources of forecast error (i.e., how much of our forecast error was because we missed our guess for how much more CO2 we pumped into the atmosphere vs. how much was because we missed our guess for how much the amount of CO2 actually pumped into the atmosphere would impact climate?). 

2. I take the spirit of your comment to be (as per Roger Pielke’s post) that the current global warming theory is non-falsifiable, since warming, cooling, or no temperature change over the next decade are all asserted to be consistent with the theory. As someone who has called for model validation on actual forward forecasts (not only "hindcasting") for some time, I have a high degree of sympathy for this view. Non-falsifiable = non-scientific is a really useful rule-of-thumb. However, I think that you need to keep a couple of things in mind.  First, one needs to match the time period of the falsification test to the underlying physical theory. I have often been presented with the assertion by climate scientists that we require something like a 30-year period to distinguish signal from noise (i.e., the proper test period is at least 30 years), so one could see the events described in the paper, and still have not falsified the predictive model. Second, I don't really think that a binary "data is consistent or inconsistent with theory and model predictions" is the most productive way to think about the results of such tests. Instead, it's really the distribution of predicted-to-actual results for a series of predictions that we care about.

3. It’s a joke that the climate modeling community has not had to date — and despite the paper that you reference, doesn’t look in any danger of starting anytime soon — a disciplined program of making formal climate predictions for future years, escrowing the code used to make the predictions, and then each year applying actual emissions and other forcings data over the period since the model was built to the exact model code used to make the prediction in order to create a true distribution of model accuracy. All predictive modeling communities resist this (as all humans resist real accountability if they can get away with it) — it’s management’s job to force this issue. One known problem of not doing this is that it leads any predictive modeling community to grossly over-estimate its accuracy. Another is that by not highlighting model error, its slows the rate of model improvement.

4. I have been pointing out the technical limitations to model accuracy for some time. However, I think it’s a mistake to believe that “models are useless” = “we know AGW is not a huge problem.” Uncertainty is not our friend. We know from replicated lab experiments that CO2 is a greenhouse gas (i.e., it absorbs and redirects infrared radiation but not shorter-wavelength radiation). Therefore, all else equal, increasing CO2 in the atmosphere will increase global temperatures. The relevant scaling parameter for this rate of increase is “climate sensitivity,” which is defined as the increase in global mean surface temperature that would result from a doubling of the atmospheric concentration of CO2. In the absence of feedbacks, it is an essentially undisputed fact that climate sensitivity is about 1C. The purpose of GCMs is to model climate feedbacks.  Current models estimate climate sensitivity to be about 3C, with a (loosely defined) range of 1.5 – 4.5C.  Estimates of this range have been consistent for more than 30 years, from the Charney Commission of 1979 through the most recent UN IPCC Assessment last year. As I have reviewed in detail elsewhere, even if we assume that GCMs are exactly correct (that is, that climate sensitivity is exactly 3C), or even if we use current estimates for the odds-weighted distribution of possible climate sensitivities between 1.5C and 4.5C, the economic benefits created by CO2 emissions abatement don’t justify the costs of lost consumption that would be created by various cap-and-trade or carbon tax proposals.  The only legitimate argument for such abatement proposals is that the non-zero probability of very large climate sensitivities, and therefore massive climate change, justify these losses in consumption today in order to avoid a low-odds disaster scenario in the distant future. I’ve argued against this point of view at length as well. But we should recognize that if we had high confidence in our climate models, it would undercut, not support, the argument for rapid, aggressive emissions abatement today, since it could eliminate this tail of the climate sensitivity — and hence economic-cost — probability distribution.

05/02 10:00 AMShare