The Big Valley: Altitude Bias in GHCN

Foreword: The focus of this essay is strictly altitude placement/change of GHCN stations. While challenge and debate of the topic is encouraged, please don’t let the discussion drift into other side issues. As noted in the conclusion, there remain two significant issues that have not been fully addressed in GHCN. I believe a focus on those issues (particularly UHI) will best serve to advance the science and understanding of what GHCN in its current form is measuring and presenting, post processing. – Anthony

By Steven Mosher, Zeke Hausfather, and Nick Stokes

Recently on WUWT Dr. McKitrick raised several issues with regard to the quality of the GHCN temperature database. However, McKitrick does note that the methods of computing a global anomaly average are sound. That is essentially what Zeke Hausfather and I showed in our last WUWT post. Several independent researchers are able to  calculate the Global Anomaly Average with very little differences between them.

GISS, NCDC, CRU, JeffId/RomanM, Tamino, ClearClimateCode,  Zeke Hausfather, Chad Herman, Ron Broberg,  Residual Analysis, and MoshTemp all generally agree. Given the GHCN data, the answer one gets about the pace of global warming is not in serious dispute. Whether one extrapolates as GISS does or not, whether one uses a least squares approach or a spatial averaging approach, whether one selects a 2 degree bin or a 5 degree bin, whether one uses an anomaly period of 1961-90 or 1953-1982, the answer is the same for virtually all practical purposes. Debates about methodology are either a distraction from the global warming issues at hand or they are specialist questions that entertain a few of us. Those specialist discussions may refine the answer or express our confidence in the result more explicitly, but the methods all work and agree to a high degree.

As we noted before, the discussion should therefore turn and remain focused on the data issues. How good is GHCN as a database and how serious are its shortcomings? As with any dataset, those of us who analyze data for a living look for several things. We look for errors, we look for bias, we look at the sampling characteristics, and we look at adjustments.  Dr. McKitrick’s recent paper covers several topics relative to the make up and changes in GHCN temperature data. In particular he covers changes over time in the sampling of GHCN stations. He repeats a familiar note: over time the stations representing the temperature data set have changed. There is, as most people know, a fall off in stations reporting shortly after 1990 and then again in 2005. To be sure there are other issues that he raises as well. Those issues, such as UHI, will not be addressed here. Instead, the focus will be on one particular issue: altitude. We confine our discussion to that narrow point in order to remove misunderstandings and refocus the issue where it rightly belongs.

McKitrick writes:

Figure 1-8 shows the mean altitude above sea level in the GHCN record. The steady increase is consistent with a move inland of the network coverage, and also increased sampling in mountainous locations. The sample collapse in 1990 is clearly visible as a drop not only in numbers but also in altitude, implying the remote high-altitude sites tended to be lost in favour of sites in valley and coastal locations. This happened a second time in 2005. Since low-altitude sites tend to be more influenced by agriculture, urbanization and other land surface modification, the failure to maintain consistent altitude of the sample detracts from its statistical continuity.

There are several claims here.

1. The increase in altitude is consistent with a move inland and out of valleys
2. The increase in altitude is consistent with more sampling in mountainous locations.
3. Low level sites tend to be influenced by agriculture, urbanization and other land use modifications

A simple study of the metadata available in the GHCN  database shows that the stations that were dropped do not have the characteristics that McKitrick supposes. As Nick Stokes documents, the process of dropping stations is more related to dropping coverage  in certain countries rather than a direct effort to drop high altitude stations . McKitrick also get the topography specifics wrong.  He supposes that the drop in thermometers shifts the data out of mountainous inland areas into the valleys and low level coastal areas, areas dominated by urbanization and land use changes. That supposition is not entirely accurate as a cursory look at the metadata shows.

There are two significant periods when stations are dropped; Post 1990 and again in 2005. As Stokes show below.

FIGURE 1: Station drop and average altitude of stations.

The decrease in altitude is not caused by a move into valleys, lowland and coastal areas. As the following figures show, the percentage of coastal stations is stable, mountainous stations are still represented and the altitude loss more likely comes from the move out of mountainous valleys .

A simple summary of the total inventory shows this

TABLE 1: Count of Coastal Stations

The greatest drop in stations occurs in the 1990-1995 period and the 2005 period, as shown above McKitrick supposes that the drop in altitude means a heavier weighting for coastal stations. The data do not support this

TABLE 2: Count of Coastal Stations dropped

The great march of the thermometers was not a trip to the beach. Neither was the drop in altitude the result of losing a higher percentage of  “mountainous” stations.

FIGURE 2: Distribution of Altitude for the entire GHCN Inventory

TABLE 3: descriptive statistics for Altitude of the entire dataset

We can assess the claim about the march of thermometers down the mountains in two ways. First, by looking at the actual distribution of dropped stations.

FIGURE 3 Distribution of altitude for stations dropped in 1990-95

TABLE 4:  Descriptive statistics for the Altitude of dropped stations

The character of stations dropped in the 2005 time frame are slightly different. That distribution is depicted below

FIGURE 4 Distribution of altitude for stations dropped in 2005-06

TABLE 5:  Descriptive statistics for the Altitude of dropped stations 2005-06

The mean of those dropped is slightly higher than the average station. That hardly supports the contention of thermometers marching out of the mountains. We can put this issue to rest with the following observation from the metadata. GHCN metadata captures the topography surrounding the stations. There are four classifications FL, HI, MT and MV: flat, hilly, mountain and mountain valley. The table below hints at what was unique about the dropout.

TABLE 6 Station drop out by topography type

There wasn’t shift into valleys as McKitrick supposes, but rather mountain valley sites were dropped.  Thermometers left the flatlands and the mountainous valleys. That resulted in a slight decrease in the overall altitude.

That brings us to McKitrick’s third critical claim. McKitrick claims that the dropping of thermometers over weights places more likely to suffer from urbanization and differential land use.  “Low level sites tend to be influenced by agriculture, urbanization and other land use modifications.” The primary concern that Dr. McKitrick voices is that the statistical integrity of the data may have been compromised. That claim needs to be turned into a testable hypothesis. What exactly has been compromised? We can think of two possible concerns. The first concern is that by dropping higher altitude mountain valley stations one is dropping stations that are colder. Since temperature decreases with altitude this would seem to be a reasonable concern. However, it is not. Some people make this claim, but McKitrick does not. He doesn’t because he is aware that the anomaly method prevents this kind of bias. When we create a global anomaly we prevent this kind of bias from entering the calculation by scaling the measurements of station by the mean of that station. Thus, a station located at 4000m may be at -5C, but if that station is always at -5C its anomaly will be zero. Likewise, a station at sea level in Death Valley that is constantly 110F will also have an anomaly of zero. Anomaly captures the departure from the mean of that station.

What this means is that as long as high altitude stations warm or cool at the same rate as low altitude stations, removing them or adding them will not bias the result.

To answer the question of whether dropping or adding higher altitude stations impacts the trend we have several analytical approaches. First, we could add back in stations. But we can’t add back in GHCN stations that were discontinued. The alternative is to add stations from other databases.  Those studies indicate that adding addition stations does not change the trends:

http://www.yaleclimatemediaforum.org/2010/08/an-alternative-land-temperature-record-may-help-allay-critics-data-concerns/

http://moyhu.blogspot.com/2010/07/using-templs-on-alternative-land.html

http://moyhu.blogspot.com/2010/07/arctic-trends-using-gsod-temperature.html

http://moyhu.blogspot.com/2010/07/revisiting-bolivia.html

http://moyhu.blogspot.com/2010/07/global-landocean-gsod-and-ghcn-data.html

The other approach is to randomly remove more stations from GHCN and measure the effect. If we fear that GHCN has biased the sample by dropping higher altitude stations, we can drop more stations and measure the effect. There are two ways to do this. A Monte Carlo approach and an approach that divides the existing data into subsets:

Nick Stokes has conducted the Monte Carlo experiments. In his approach stations are randomly removed  and global averages are recomputed. Stations were removed based on a randomization approach that preferentially removed high altitude stations. This test gives us an estimate of the Standard Error as well.

Table 7 Monte Carlo test of altitude sensitivity

This particular test consists of selecting all the stations whose series end after 1990. There are 4814 such stations. The sensitivity to altitude reduction was performed by randomly removing higher altitude stations. The results indicate little to no interaction between altitude and temperature trend in the very stations end after the 1990 period.

The other approach, dividing the sample, was approached in two different ways by Zeke Hausfather and Steven Mosher. Hausfather, approached the problem using a paired approach. Grid cells are selected for processing if the have stations both above and below 300m. This eliminates cells that are represented by a single station.  Series are then constructed for the stations that lie above 300m and below 300m.

Table 8. Comparison of trend versus altitude for paired station testing

FIGURE 5: Comparison of temperature Anomaly for above mean and below mean stations

This test indicates that higher elevation stations tend to see higher rates of warming rather than lower rates of warming. Thus, dropping them, does not bias the temperature record upward. The concern lies in the other direction. If anything the evidence points to this: dropping higher altitude stations post 1990 has lead to a small underestimation of the warming trend.

Finally, Mosher, extending the work of Broberg tested the sensitivity of altitude by dividing the existing sample in the following way, by raw altitude and by topography.

1. A series containing all stations.
2. A series of lower altitude stations Altitude < 200m
3. A series of higher altitude stations Altitude >300m
4. All Stations in Mountain Valleys
5. A series of stations at very high altitude. Altitude >400m

The results of that test are shown below

FIGURE 6 Global anomaly.  Smoothing performed for display purpose only with a 21 point binomial  filter

The purple series is the highest altitude stations. The red series lower elevation series. Green is the mountain valley stations. A cursory look at the “trend” indicates that the higher elevation stations warm slightly faster than the lower elevation, confirming Hausfather. Dropping higher elevation stations, if it has any effect whatsoever works to lower the average.  Stations at lower altitudes tend to warm less rapidly than stations at higher elevations. So quite the opposite of what people assume, the dropping of higher altitude stations is more likely to underestimate the warming rather than over estimate the warming.

Conclusion:

The distribution of altitude does change with time in GHCN v2.mean data. That change does not signal a march of thermometers to places with higher rates of warming. The decrease in altitude is not associated with a move toward or away from coasts. The decrease is not clearly associated with a move away mountainous regions and into valleys, but rather a movement out of mountain valley and flatland regions. Yet, mountain valleys do not warm or cool in any differential manner. Changing altitude does not bias the final trends in any appreciable way.

Regardless of the differential characteristics associated with higher elevation, changes in temperature trends is not clearly or demonstrably one of them.  For now, we have no  evidence whatsoever that marching thermometers up and down hills makes any contribution to a overestimation of the warming trend.

Dr. McKitrick presented a series of concerns with GHCN. We have eliminated the concern over changes in the distribution of altitude. That merits a correction to his paper. The concerns he raised about latitude, and airports and UHI will be addressed in forthcoming pieces. Given the preliminary work done on airports. (and here) and latitude to date, we can confidently say that the entire debate will come down to two basic issues: UHI and adjustments, the issues over latitude changes and sampling at airports will fold into those discussions. So, here is where the debate stands. The concerns that people have had about methodology have been addressed. As McKitrick notes, the various independent methods get the same answers. The concern about altitude bias has been addressed. As we’ve argued before, the real issue with temperature series is the metadata, its related microsite and UHI issues and adjustments made prior to entry in the GHCN database.

Special thanks to Ron Broberg for editorial support.

References:

A Critical Review of Global Surface Temperature Data Products. Ross McKitrick, Ph.D. July 26, 2010