Climate Change & Tropospheric Temperature Trends

Coefficients were then derived for each of these independent variables using least squares methods. For dependent variables, MM use monthly climate records for the period 1979-2000 taken from a network of 218 land based weather stations in 93 countries (shown in Figure 21) chosen from the GISS homogenized surface record (Hansen et al., 1999), and another set of temperature data which they modeled separately, that consists of 5 deg. by 5 deg. gridded data from the IPCC for the cells that correspond to their selected network of GISS stations. This global model was then subjected to a standard multiple regression analysis from which correlations were sought between global warming trends and the various climate, economic, and social influences they modeled.

To no one’s surprise, MM find that the warming observed in the northern hemisphere since 1979 correlates better with economic and/or social variables than with climate variables. To further clarify these results, they sort their data into two groups – one corresponding to global data for the colder half of any given year, and one for the warmer half, after which they further subsample their cold season data to reflect dry (that is, subzero dew point) regions only. They reran their analysis under these conditions, sorted their resulting independent variable correlations by order of importance in each grouping, and then evaluated the actual temperature trend impact of each by removing them one at a time for successive runs of their regression (McKitrick and Michaels, 2004). The results of this exercise are shown in Figure 22. MM note that with economic and social effects removed, the global trend is in surprisingly good agreement with the MSU record (by which, they mean the UAH Version 5.0 TLT record of course). Removal of their Soviet “dummy variable” drops the trend further (to a value remarkably close to the UAH Version 5.0 TMT record, but MM do not comment on this), but they acknowledge that their Soviet variable may not be a true greenhouse surrogate.

Thus, regarding global surface temperature trends for 1979-2000, McKitrick and Michaels are led to two striking conclusions,

1. Outside of cold season dry regions, warming trends are dominated by economic factors, as characterized in their model.
2. During the warm season, warming trends in all regions are dominated by a combination of economic and social factors.

These are bold statements. If MM are right, then nearly all of the warming trends observed worldwide at the earth’s surface, and possibly in the lower troposphere as well, are little more than evidence of our own growing wealth and productivity, mistaken for dangerous global warming. So victory over global warming has finally been achieved and we can all relax, right?

Wrong! Not only did MM fail to make their point, the ensuing drama that followed the publication of this “bombshell” paper will surely go down in history as one of the greatest comedy of errors ever to beset the scientific community. The journal Climate Research, which had already been through one scandal involving yet another seriously flawed paper ¹⁰, published this one in May of 2004. Shortly thereafter the festivities began. MM ran their multiple regression analyses using an econometrics program called SHAZAM ¹¹. SHAZAM makes use of input data files and has its own associated language for characterizing variables and program calls. As noted above, MM used cosine of latitude ( Cos(L) ) as one of their input climate parameters. They derived this within the SHAZAM user interface by trigonometric identity using the program’s built-in Sin(x) variable and give it the label COSABLAT. According to the SHAZAM User’s Guide ¹¹, all trigonometric variables in SHAZAM require their arguments to be expressed in radians. In August of 2004, barely 12 weeks after the paper hit the street, Tim Lambert of the University of New South Wales, Sydney, Australia, obtained McKitrick’s SHAZAM input file from his U. of Guelph web site and discovered that MM had input all of their latitude data to the variable COSABLAT in degrees rather than radians, making virtually all of their derived results useless! When Lambert corrected the errors and reran their analysis MM’s “economic” signal was drastically reduced 12. Figure 23 shows MM’s results after corrections to COSABLAT were made to their inputs. Comparison with their pre-corrected results (Figure 22) reveals that with the corrections made their economic and social signals have been reduced from 0.259 to 0.160 deg. K/decade – in other words, by over one third. At Tech Central Station, DEA proudly tell us that McKitrick and Michaels published this research “after four years of one of the most rigorous peer reviews ever…” (Douglass et al., 2004c). But this “rigorous” peer review process did not even manage to catch a simple conflation of input units that would have been inexcusable in an undergraduate exam!

Well alright then – so they made a simple mistake with units that any of us could have made on a bad day. So what! They still come up with over half of the surface temperature signal tied to economic and social factors rather than climate change. Doesn’t this establish their central thesis anyway?

Only if you neglect the overly simplistic characterization of these effects in their model and it resultant potential for data clustering. Consider MM’s choice of input parameters as given above. Their modeling of effects as complicated as global climate, economic activity, and even social change… boils down to a mere 12 variables. The methodological shortcomings in this approach are almost too numerous to mention. Consider the following,

• Surface pressure in dry regions only is taken as a primary driver of climate response. Moist regions – which comprise the large majority of the earth surface – are summarily lumped into a single “remainder” category.
• Other than a dummy variable for generalized “coastal proximity”, oceanic effects aren’t considered at all, even though oceans cover over 4 fifths of the earths surface and dominate the overall response of the lower atmosphere. Certainly we would expect Michaels to be aware of this given the effort he and his co-authors went to in order to avoid the 1997 ENSO event in their first “tropospheric disparity” paper (Douglass et al., 2004).
• “Coal use” is treated as a proxy for sulfate emissions even though they could have just used direct measurements of anthropogenic sulfate emissions with far less uncertainty (Lelieveld et al., 1997; IPCC, 2001 Chap. 5.2.2.6).
• Economic development activity is characterized by simple land use changes, urban heat island effects, and the impacts of economic development on record keeping with no apparent attempt to correct these effects for the fact that many regional climate responses, including surface temperature, are known to be correlated over larger distances than what can be resolved by simply looking at urban growth and development (Wilks, 1995).
• National GDP growth rate is considered a key economic proxy for regional climate change, yet it is a characteristic of entire national economies. MM make no attempt to describe GDP impacts the regional distribution of economic development within any given country other than factoring it by population. Consider the magnitude of error likely to result from applying a single number nation-specific parameter like this to the United States without considering the separate regional contributions to it of say, the Aleutian Islands, the Sonora desert, and New York City or Los Angeles. Even if we were to factor GDP by say, regional population (as MM do with their annual per capita income inputs), it is still straightforward to identify rural and urban areas, or even separate urban areas with different industry make-ups, that have similar populations but very different heat and/or greenhouse gas or particulate emissions characteristics.
• Other than generalized land use and urban heat island impacts, there is no clear differentiation between the impacts of manufacturing and service based economic growth, nor is any attempt made to discriminate between various greenhouse gas emitting industries other than via simple coal production. Concrete for instance, is known to be a significant producer of greenhouse gas emissions, particularly CO2 (see for instance, Milmoe, 1999). Yet MM do not consider it at all.
• Local literacy rate is identified as a social proxy for climate change. Yet how the two might be related is not addressed, and a myriad of other potential social impacts which conceivably could have a much larger influence (e.g. – cultural trends regarding environmental sensitivities, for which there is a fair amount of survey data, particularly in the U.S. and Europe), are not even considered.
• The former Soviet Union is singled out as a unique contributor to the variability of surface station temperature records - even to the degree that MM consider it to be a step function input to their model. The basis for this is the dramatic socio-economic changes experienced by former Soviet regions during the period of record. But other regions such as sub-Saharan Africa have undergone even more dramatic social upheavals during the same period, suffering similar and/or greater proportional impacts on their existing surface station records, yet these are not considered at all, much less treated as step function changes.
• Coal growth rate is considered to be a key economic proxy for climate change, but automobile production and the resultant emissions increases are not even considered! If coal consumption counts as an economic surrogate for sulfate production, certainly growth in automobile production and use would be a proxy for particulate emissions, smog, and greenhouse gas emissions as well. MM do not even mention it.

And so on, and so on, and so on. This list could be expanded ad-infinitum, but by now the problem should be abundantly clear. The heart and soul of MM’s methodology is the use of multiple regression methods, by which they wish to demonstrate that perceived anthropogenic climate signals are in fact more strongly correlated with social and economic “signals” unrelated to climate change. But it is a well known mathematical fact that multiple regression analysis requires that input variables be independent and identically distributed. If they are not, data clustering (the presence of unaccounted-for correlations between variables that are being treated as independent) is almost certain to happen. As a consequence, their standard errors, and therefore their confidence intervals, are almost certain to be under-estimated, and they will treat statistically insignificant results as valid signals.

Problems like this have plagued econometric large scale analyses of economic and social problems for years because it is generally impossible to account for all the variables needed to make such analyses believable. Such studies almost always degenerate to endless number crunching exercises where input parameters “parameterized adjustments” intended to account for the unavoidable blizzard of unknowns allow for the output of, quite literally, any result desired. MM’s paper is no exception. Consider for instance, their use of cosine of latitude ( Cos(L) ) for a climate variable rather than simply latitude L. There is no apparent justification for this, and MM offer none – they simply do it. We have to wonder what the point is of arbitrarily adding this extra layer of complexity to their inputs. It is true that the area associated with an annular “slice” of latitude of thickness dL would vary as Cos(L), so if MM were seeking correlations to variables that were sensitive to area, they might get a more direct comparison from Cos(L) perhaps. But this would only serve to simplify the mathematical formulation of their desired description of latitude correlated effects – it wouldn’t have any truly meaningful effect on the correlation itself, which is what they’re ultimately after. Though it would be difficult to prove, the best explanation for the use of Cos(L) rather than L appears to be that MM experimented with latitude based input variables and got the results they desired using Cos(L).

Given the issues surrounding multiple regression methods and data clustering in studies such as MM’s, in the very least they should have subjected their model to independent tests of its robustness. The SHAZAM program they used does generate a heteroskedasticity-consistent covariance matrix as part of its output. Ordinary least squares methods can produce wide variation in small trend estimates from separate datasets seeking to measure the same changes even though the two datasets are well correlated (we saw this earlier while comparing separate MSU and radiosonde analyses). If the residuals from these analyses contain significant heteroskedasticity (e.g. – significant variations in standard deviation, or standard error, over subsampled portions of the time series being examined), this would be a big warning sign that the model had been somehow mischaracterized or contaminated by variables that had not been taken into account. Given that MM’s SHAZAM run would have produced heteroskedasticity-consistent results, it is reasonable to expect that their results bias free, and therefore independent of the multiple regression model used to derive them. Furthermore, if their economic and social signals are truly real, rerunning their analysis with part of their dataset should allow for the rest of their independent data to be reproduced.

Tests like these are standard for multiple regression models like MM’s, yet they steered well clear of them. But even so, it wasn’t long before someone else got ahold of their data and did subject it to robustness tests like these. Benestad (2004) obtained MM’s dataset from McKitrick’s web site and reconstructed their analysis using a separate multiple regression model. In particular, Benestad was concerned about the fact that MM did not account for the interstation temperature dependencies caused by correlations of temperature trends over geographic regions larger than MM’s land use and urban heat island variables could resolve. First, he established that his model could reproduce their results using their full data set, thereby demonstrating that he had established a baseline from which further tests could be made. Then, he reran their analysis using 2 separate subsets of their data – one of which was used to calibrate the model, and the other to test its ability to predict the outcome of the variables that had been omitted.

To no one’s surprise, MM’s model failed miserably. Benestad ran five separate analyses of MM’s dataset, each constructed from a subset of MM’s input variables. Dependent variable data from stations within the latitude band of 75.5 deg. S to 32.2 deg. N were used to calibrate the model, and data from the latitude band 35.3 deg. N to 80 deg. N were used to evaluate MM’s results. The following runs were evaluated,

• A model that used all of their input data (McKitrick and Michaels, 2004 Table 4).



Top


Page:   << Previous    1    2    3    4    5    6    7    8    9    10    11    12    13    14    15    16    17    18    19    20    21    22       Next >>