Here is Gavin’s response from the link Roger provided:

Viento, If you restrict the simulations you want to those that have surface trends within the obs uncertainty as defined by Douglas et al (+0.12 +/- 0.04), then you only retain 9 of the models. For the trends at 300mb, the 2 sigma range for those models is then: 0.23+/-0.13 degC/dec (compared to 0.31+/-0.25 in the full set). So you do get a restriction on the uncertainty, and in each case the RAOBCORE v1.4 is clearly within the range. At 500mb, you get 0.17+/-0.076 (from 0.23+/-0.22), which puts the obs just outside the range – but with overlapping error bars. So I don’t think the situation changes – you still cannot find a significant difference between the obs and the models.

It appears to me that Gavin has missed the point of the Douglas et al study.

The issue, as I read it, is not which models to exclude — and Douglas et al does not exclude any of the 22. The issue is how to calculate the uncertainty of the models. Douglas is arguing that using the range of the data is misleading due to the existence of outliers. From page 7 of Douglas et al:

Santer et al. (2005) have argued that the model results are consistent with observations and that the disparity between the models and observations is ‘removed’ because their ranges of uncertainties overlap. They define ‘range’ as the region between the minimum and maximum of the simulations among the various models. However, ‘range’ is not statistically meaningful. Further – and more importantly – it is not robust; a single bad outlier can increase the range value of model results to include the observations. A more robust estimate of model variability is the uncertainty of the mean of a sufficiently large sample.

So that is a fundamental premise of the Douglas et al paper: that the uncertainty of the mean is a more reasonable estimate of model uncertainty than using the range of model outputs

The range of model outputs includes one model that predicts the following temperature trends:

Altitude(hPa) – Trend, degC/decade

Surface: .028
1000: .024
925: .046
850: .073
700: .027
600: -.026
500: -.026
400: -.001
300: .020
250: .024
200: .032
150: -.001
100: -.136

I see no way to support the claim that “the science is settled” if model uncertainty encompasses predictions of trivial to no heating.

I’d be very interested to hear what our resident statisticians think about the best way to calculate the uncertainty of the models.

I think you may have misread my post. I stated, “when the warming is SAID to be ‘natural’.” I was not offering an opinion on that contention (made by others) one way or another.

http://www.realclimate.org/index.php/archives/2007/12/tropical-troposphere-trends#comment-75596

Here is part of my response:

“Again, it is not right vs. wrong. There is a probability that the obs come from a distribution represented by the models. Using the full suite this seems to be a reasonably high probability (which was not calculated by either RC or Douglass et al.). Restricting that test to those models with the best accuracy with respect to surface trends appears to decrease this probability.”

Of course I’m a political scientist, so what do I know?

More generally, more closely to my own area of expertise (politics and science), the subject of comparing models to obs in the tropics seems to be so thoroughly tribal (good guys vs. bad guys) that I doubt that much will be able to be said on this other than “large uncertainties” meaning that it will be quite easy to confirm one’s priors by asserting the uncertainties, or trying to restrict them selectively.

Of course, the tribal nature of arguing academics is not new or unique to climate change, it is just far more visible due to the nature of the issue and the visibility of blogs.

My head exploded most likely 80 times. But I don’t have any margin of error on that, so it could have been 0 or 1,000,000.

Prepare to have the top of your head explode.

Anyway, as far as this discussion here: See my comments about the GMTAT on the BB.

Model runs with results that are unreasonably low are discarded.

Model runs with results that are unreasonably high are used to base policy on?