Anyhow, I do not want to debate this with you here b/c it is OT. Go to Tamino’s blog a pontificate there.

[DC: I think Girma has been banned most places, for obvious reasons.]

[DC: Off topic - also your linked chart connects a smoothed curve up to 2005 to actual observations for 2006-2008. That's highly misleading.]

Thanks.

[DC: I just meant I'm not in the same league as Tamino. But I think we can all agree on 10 yr temp trends.]

2. Change topic: Could you do a second degree polynomial fit to the full series UAH data? or RSS? I just want to see how much the “recent flattening” turns the record with that degree of play (not saying its justified, but just to look at).

3. How do I download the monthly data to Excel so I can do regressions?

[DC:
1. If your "biggie" correspondent likes me, he'll love Tamino. There was a recent post there (within the last month or so) on the correct calculation of short-term uncertainties. I'll try and dig that out when I have a chance.

Later .. gotta run]

A) 2008 outside confidence limits.
These are 90% confidence limits derived from the smoothed projections in AR4 Fig. 10.26 (as explained above). The smoothed observation trends are within these projections. Some individual years are outside smoothed projected/hindcast trend confidence limits (e.g. 1998 and 2008), but that is to be expected.

B) Linear trends 2000-2010
For the period shown in my final graph, I have A1B at 0.207 deg/decade and A2 at 0.176 deg/decade. The average between the two is very close to 0.19 deg/decade, so the generic 0.2 deg/decade is a tad high, if anything.

C) Smoothed projection trend
The smoothed projections in fig. 10.26 are very close to linear for the 2000-2010. I could have shown the actual smoothed trends for A1B and A2, and perhaps I will in part 2. This will slightly complicate the graph, but not change the essential conclusion.

Deep-
A) Do you really think your dashed line indicates those are 5%-95% confidence intervals? If yes, that means 2008 was outside the 90% confidence intervals.

B) Figure 10.4 represents the IPCC projections and its based on the underlying models. The trend over the period you show is higher than 2 C/century.

C) If you are going to compare smoothed data to IPCC projections, it would be best to apply the same smoothing to the actual, honest to goodness IPCC projections. You should avoid apples to coconuts comparisons even if the apple to coconuts comparisons give you the answers you prefer.

The other thing that you fail to mention is that the fall in the 8 year trend in the early 1990s was due to the Pinatubo eruption in 1991. There is no analogous eruption to explain the current falling trend.

One can plausibly argue that observed temperature trends are below projections, albeit without statistical significance. But there is absolutely no evidence that “trends are falling” in any meaningful sense (i.e. at the multi-decadal level).

The recent inter-annual trough appears to have resulted from the confluence of a fairly strong with La Nina episode with an unusually extended inter-cycle solar minimum (both of which have now ended). It is almost certain that the eight year trend from 2003-2010 will be well above that from 2001-2008. At that point will you then say the “currently falling” trends are rising again?

“In order to undertake a similar analysis comparing current observations to the latest IPCC projection, we first derive a smoothed temperature record to the end of 2008, using a 21-point lowess smooth (data used goes back to 1970 and is baselined to 1980-1999).”

Your graph says LOESS which is a different algorithm. That being said, using a 21 point loess (span = 0.7? in R) is far too coarse and is not the intended purpose of that statistical method. Loess and Lowess were designed as noise reduction algorithms and adapted to statistical method for other noisy data types. A span of 0.7 in almost any series is next to useless for the derivation of signal from noise. For your dataset a span of 0.25 or 0.3 would be far more suitable for fitting to the data. The span should not reflect the length of the base period but the minimum length of the span that will give an accurate signal from the noise.

My point is that you are using lowess or loess outside of it’s intended purpose and with a smoothing factor that most statisticians (I referred this before to a statistics expert) would regard as not giving any useful information.

[DC: Some authorities treat lowess and loess as equivalent, while others treat lowess as a particular form (namely local linear fitting) of the more general loess smoothing (which can be quadratic). I've added a few references on the subject as an update above. I'll also relabel the graph to eliminate any confusion.

Of course, lowess smoothing is much superior to the higher-order polynomial trends used until recently by Roy Spencer, as discussed in a previous post.