As we all know, the GISS Temp for August 2008 rose during Sept 2008. After the update, the GISS Temp for Sept 2008 was lower than the August 2008 temperature. Are you wondering about the net effect of the increase in August temperatures and subsequent decrease in September temperature had on the trend since Jan 2001?

Interestingly enough, the trend dropped from -0.24 C/century to -0.26 C/century.

Why did the trend drop? Well, the current temperature fall below the trend line. Barring unpredictable revisions of past temperatures, if the October GISS anomaly is less than 0.518C, the trend will drop next month as well.

Yes, those who want to say there is “no statistically significant warming trend since 2001 ” and base that on GISSTemp are correct. That said, there is no statistically significant cooling trend either. So, you coolers– cool your jets!

Of course, this still leaves the more important “luke-warmer” question: How is the “about 2C/century” projected for the surface trends in the IPCC AR4 holding up compared to the data? Especially with the new-improved statistical model? 🙂

As discussed last month., I selected an “AR(1)+White noise” statistical model, and estimated the parameters based on years when the stratosphere was clear aerosols from large volcanic eruptions. Below, I compare the distribution of 93 month trends one would expect based on the historical “weather noise” observed in GISS during the “volcano-lite” period:

You’ll probably want to click on the graph to read the detail, but the main points are:

1. If “weather noise” is as observed by GISS during the past “volcano -lite” period, and the “true trend” is 2C/century masked by “weather noise” modeled as AR(1) plus noise, and with parameters estimated based on the period discussed here, we’d expect 3.2% of 93 month trends to be less than -0.26 C/century. Since in statistics, it’s fairest to use “two-tailed” tests, we say the likelihood of an excursion this large appears to be about 6.4%. So, the likelihood is less than 1 in 10. The IPCC AR4 term for our confidence that 2C/century is correct is “very low confidence”.
2. Under the assumptions above, the 2C/century trend is not “falsified” based on the more common 95% confidence intervals. That is, if we require a 1 in 20 chance of decreeing something incorrect, we

What are the caveats?

As usual, there are caveats. This is a blog, so I’m not listing every tiny one. The big ones I can think of this second are:

1. Weather (and measurement noise) is estimated using AR(1)+White noise. Statistical models are assumed in all statistical analyses, but it is rarely possible to prove that a particular model really applies. If you think the weather noise should be modeled using something else, you will get different answers.
2. The parameters are estimated during a particular 30 year historic period when the stratospheric aerosol loading was fairly low, and consequently not varying. This condition has also applied during the period I am analyzing. Since volcanic eruptions like Pinatubo are thought to result in variations in dramatic temperature variations; I believe estimating “weather noise” during periods without eruptions like Pinatubo likely results in a better estimate of the expected “weather noise” during since 2001 than estimating weather noise to include eruptions like Piinatubo and El Chicon. However, this assumption does affect the results. (Most recently, I compared the amount of noise in a “volcano -medium” to a “volcano-lite” period here.
3. The “noise” in the statistical model should be affected by both measurement uncertainty and natural variability of weather. It can be argued that “weather noise” might remain constant, but “measurement uncertainty” may be lower than it was between 1914 and 1944. If so, the “noise” in the statistical models is too large and could cause us to over estimate the spread of trends.
4. Using a finite time span to estimate the parameters results in biased estimates of the parameters. I have reason to believe the effect is fairly small, but it tends to cause me to under estimate the spread in trends. So, this error is in the opposite direction of the one discussed above. I’m looking into ways to correct the bias in the estimate.
5. As always: Outliers happen. Things that happen less than 1/10 times do happen. Things that happen less than 1/20 times happen. All we can ever do with statistics is explain our assumptions and report the results. 🙂

What’s in store for October?

As I mentioned before, I’m tempted to set up a “brownie -bet” pool where people bet on the future values of already posted GISS Temp’s. I doubt either phenomenological understanding or statistical knowledge is will ever help anyone predict this value. So, it could be a lot of fun. But, to do this properly, (and avoid baking and shipping too many brownies) I need to come up with an automated way to bet and a “points system”. That way, people can accumulate points, and the “brownie winner” can be whoever got the most points by March — or something like that.

If you have suggestions, post them! (If you want to guess the magnitdue of GISS’s September anomaly that will appear in November’s GISS Temp update, post away. This month, the prize for closest will be “bragging rights”!