Recently, I received multiple requests to discuss a paper, due to appear in Journal of Geophysical Research (JGR-atmosphere), that has been presented in the media just before the Bali conference and the Nobel Peace prize ceremony here in Oslo, Norway. The paper concludes that the warming measured over land is most likely exaggerated due to non-climatic effects, and it presents a regression analysis suggesting that the real (climatic) global mean temperature trend should be ~50% lower over land.
So, are the surface temperature trends inflated? This new paper by McKitrick & Michaels (henceforth ‘M&M2007‘) is a followup of an earlier paper they wrote in 2004 in Climate Research (MM2004a), which I discussed in my first RC post (Are Temperature Trends Affected by Economic Activity?) and in a commentary in Climate Research (Benestad, 2004).
So what’s new? Let’s backtrack a little and recount some of the previous arguments.
One of my main concerns then was that their analysis had not taken into consideration the dependency between the data points, as the temperature exhibits non-negligible spatial correlations. Furthermore, data from the same country were compared with the same national value in terms of economic indices. It was a bit like doing a poll by asking 10 people the same question 100 times and then claiming that it’s a survey with a sample size of 1000.
In 2004, M&M2004b said they were unaware of any paper in the refereed applied climatology literature that had performed a test where half the data was excluded when doing the model calibration and the rest was left for model validation. I guess that they were not really up-to-date then, because that has been a standard approach for a long time.
But this time they have split the data sample and used a part for validation, which I suggested in my comment in Climate Research. But they have not done it properly this time, and they still do not eliminate the effect of dependency. They split the data by randomly picking points which were either used for training the data or validating the model, thus data from adjacent sites which are related will end up in the different batches for training or validation.
The map above shows a simple estimate of the temperature change over the 1979-2002 period (here taken as the differences in the mean over two sub-periods and the National Centers for Environmental Prediction (NCEP) re-analysis have been used instead of the CRU data), and it’s easy to see that the warming varies smoothly from location to location. In other words, the trend estimates have significant spatial correlation.
The fact that they used sea-level pressure (SLP) data from (1974) because they could not find more recent data, suggest that they still are not up-to-date. Updated data, such as the National Center for Environmental Prediction SLP, have long been available from NOAA Climate Diagnostics Center. Furthermore, a wealth of up-to-date climate data are available from the KNMI (Dutch Meteorological Institute) ClimateExplorer.
Their regression analysis appears to suffer from over-fitting, since they have thrown in a lot of variables (both ‘meteorological’ and ‘economical’) for various vague reasons.
Not surprisingly, their analysis produces some strange results as a result of this shortcoming. They find that the greatest differences between measured and adjusted trends at Svalbard and other places in the Arctic and Antarctic (See marked sites in Figure below). This is not convincing. Thus, the results themselves provide examples of spurious values obtained by their analysis. Even if they were identified as ‘outliers’ (Svalbard was apparently not one), the fact that their analysis produced highest corrections for economic activity at these places suggest that their analysis is not very reliable.
The graphic below shows a Google Earth image of Svalbard, which is one of the sites marked in the map above with a large trend correction due to economic activities.
I have not examined the economic data, but it appears that M&M2007 maybe cannot win – either (i) the spatial distribution of the economic indices are equally smooth and M&M2007′s attempt to account for dependencies within each country fails to resolve the problem of dependency between the countries, or (ii) the economic indices vary abruptly from country to country and thus have very different spatial scales and structures to those seen in the warming trends. Either option suggest that their analysis may lead to spurious results, over-fit, or suffer from inter-dependencies.
I also think that M&M2007 is biased and gives an incorrect picture, as they do not discuss the fact that also the world oceans are warming up, and whether any economic activity can take the blame for that. I think it is difficult to argue that factors such as the urban heat island effect plays an important role here.
They do not mention my criticisms raised in Benestad (2004) either, which discussed a number serious concerns about their previous study; They merely state, as if it were a matter of fact, that urbanisation and economic activity has been shown to affect local and regional temperature measurements – citing their old criticised paper.
Their analysis relies on University of Alabama-Huntsville (UAH) satellite data (Microwave Sounding Unit, MSU) with a weaker global trend than others, and neglect to examine or even mention other products such as the Remote Sensing System (RSS) data. The difference between these data sets are discussed in previous RC posts (here). They reckoned that any of their results would not be contingent on the choice of MSU product, but did not test this hypothesis.
It should also be kept in mind that their analysis involved too short time series (24 years) for a proper local trend estimation, as local circulation variations (e.g. the North Atlantic Oscillation), the annual cycle, and inter-annual variations, most likely will make the analysis more difficult. Climatic time series from single locations tend to be very noisy, but a clear signal emerges when taking the global mean (by taking the mean, random noise tends to cancel to some degree).
I find it a bit ironic when people use satellite data measurements to argue that GHG is unimportant. They rely on the fact that these measurements are derived using the very same type of physical laws as those predicting an enhanced greenhouse effect due to increased GHG levels (neglecting feedback processes).
I think it’s good that M&M2007 put a focus on the problem with data paucity and quality. There may very well be some non-climatic effects contaminating the measurements, but I am not convinced by their analysis.
So in summary, I think the results of M&M2007 analysis and conclusions are invalid because
- They do not properly account for dependencies.
- They over-fit the regression.
- Their results look unreasonable.
- They “cherry pick” the MSU data that gives the lowest trend