The Big Valley: Altitude Bias in GHCN

For a period of nearly 35 years high altitude stations appeared to indicate cooler temperatures. What causes the bias oscillation in Figure 5?

What I would really love to see is a study on diurnal variations from overnight lows to mid-day highs and how they have changed over the years.

E.M.Smith says:
August 19, 2010 at 10:52 am
So, IMHO, we have crappy data and get crappy results from it. Admiring the uniformity of the crappiness does not yield much comfort.
LOL!
I know you were trying to make a serious point. But it’s a good thing there was no milk in my mouth at the time. 😉

Steven,
Thanks for the several posts. If I correctly understand your various points and references, the rating system was adapted from one in use by the French, and the error range stated for each CRN category was the maximum range of error, and not the mean error. And that a station may change rating category through the seasons due to environmental factors, say proximity to trees, which would tend to bias summer temperatures, but less so the winter ones. Reasonable enough, but note that I did not assert a possible error in the +2C range, but more in the range of 3-5 times +.15C, or +.45C to +.75C.
Based upon what you’ve shared with me, it seems clear that without knowing the mean error of the stations in a given CRN category, simply knowing what percent of stations fall in a particular category doesn’t produce a meaningful answer as to the temperature bias in the records. Dr. Leroy’s associate showed a +.1-+.15 error range over CRN2-CRN4 stations, but that was without the inclusion of CRN5 stations. You mentioned a study with a too small sample size which suggested that error range MIGHT be too conservative, and the only way to tell was to gather the data. Which is what was done by Anthony’s Surface Stations project.
One thing seems clear to me from ChristianP’s comments and from the sites I surveyed for the Surface Stations project, is that the status of stations needs to be constantly monitored, and that adjustments need to be regularly updated to reflect current conditions at the site. For example, I was told that the shed at the Telluride site had been built there about 11 months earlier, and that the City had put it there without asking or telling NWS manager for that site. Yet when he found out about it, nothing was done about the station until we did the site survey, and then they immediately pulled the station from that location entirely.
ChristianP also mentioned they were using a modern Stevenson Screen (in two sizes, I take it) and that reminds me that each site I surveyed had been “upgraded” to the MMTS from the Stevenson screens originally in place. In each case the placement of the MMTS was MUCH closer to buildings and other structures which would bias the temperature record upward, if not properly adjusted for. The MMTS has a data cable which must be run inside a building to a data recorder. The cable must be put in a trench. Someone must dig the trench. With a hand shovel! Naturally, most trenches were not dug very far from the structure the data recorder was put in. Not supposed to be that way, but try digging a 50 foot long trench in mountainous soil with a hand shovel sometime. (Much less out to the 100 foot standard!) I’ve looked at other surveys at the Surface Stations project, and this seems to be pretty common for other areas of the country as well. But, this shouldn’t necessarily make a station unfit to use if adjustments for each site are properly and regularly done, but that’s one of the real questions, isn’t it?
Steven, like yourself, I’m willing to be convinced with the data. In the absence of good and convincing data, my default position is that the null hypothesis is preferable: That any warming in the record can be explained by natural climate variations and observation errors. My sincere thanks for your ongoing contributions here at WUWT.
Dan

Rex from NZ says:
August 19, 2010 at 10:38 pm
And what does UHI mean?
Urban Heat Island. Look at the top of the page. There is a “Glossary” tab.

Steven Mosher says:
“Wow. The temperature is FLAT. that is one of the points of doing an anomaly.
To PREVENT the problems of doing a simple average of temperature. NOW, IF
the records had no missing data, well you could use temps. I had section about this in the paper but we pulled it because it was too obvious”
I’m afraid you missed my point. Yes, everything you said about anomaly comparison is true. But it still removes perspective. Hello Winnipeg, just so you know, because of global warming expect highs this summer of 39.7 degrees in August instead of 39.3. Oh, and this winter, your lows will be -39.7 instead of -40.5. I’d survey denizens of Winnipeg to see their reaction but with temps in the 30’s they’ve deserted the city for lake country and in the winter door to door surveys are not appreciated.
Anomalies have plenty of value, but to be meaningful they eventually have to be translated into the actual temperatures that actual people will experience in actual locations or the execrcise is menaingless. When you plot the magnitude of the anomalies against the temperature range that humans have lived at for centuries, you get a perspective as to just how insignificant the change is. Particularly when the coldest parts of the planet warm the most and the warmest parts of the planet the least. Winters warm more than summers and nights more than days. The average across all of this may be 0.6 degrees per century or something in that range on a global scale, but it still comes back to the same thing. An artcic reqion that warms will have slightly warmed summers but milder winters with most of the warming coming in the parts of the year where plant and animal life are both stressed to the limit by cold. OK Mr polar bear, you’ll have to survive -42 this winter instead of -49, but the really bad news is this summer you will have to put up with a swealtering +12 instead of +7.
Mr Polar Bear: “How do I get more of this “problem”?

Hello everyone. Thank you Steve, Nick and Zeke. I am on holiday, and on a slow dial-up, so I will do my best to follow this thread but my online time is limited until September. Here are my main comments.
“McKitrick does note that the methods of computing a global anomaly average are sound.”
– I’m pretty sure I didn’t phrase it like that. I have noted that the various gridding methods currently in use tend to yield similar results once the choice of input data is made. This has been demonstrated by the various groups using GHCN data, including Muir Russell’s team. As to saying whether the methods are “sound”, that goes further than I prefer to.
“the process of dropping stations is more related to dropping coverage in certain countries rather than a direct effort to drop high altitude stations”
– Altitude is likely not the deciding factor in dropping a station. It is more likely that a station is dropped due to cost of collecting the data or something like that, and the change in mean altitude is a knock-on effect.
As I see it, you have looked at 3 topics: did the global mean altitude change, was it due to a relatively large loss of inland and mountain sites relative to coastal sites, and did it affect the trends.
On the first, your Figure 1 shows variability in mean altitude but not the 1990 and 2005 discontinuities in Fig 1-8 of my paper, and hardly any change in the mean, by the looks of it. It would be helpful if you explained how you computed mean altitude. Are you arguing that there was, in fact, no change in mean altitude?
Table 1: What is the date for this distribution of stations? Is it something like, all stations over the entire 20th century? It would be helpful to the reader to know what exactly the sample is. Even more helpful would be the counts and % at some key intervals, like 1970, 1980, 1990, 2000 and 2010. You could make your point much clearer and more succinctly if you did that kind of tabulation.
On the second issue, you say “McKitrick supposes that the drop in altitude means a heavier weighting for coastal stations. The data do not support this” But then you proceed to show the sample of dropped stations, post-2005, is skewed towards inland sites. Only 9% of stations dropped are coastal as opposed to 30% in the larger sample, and 83.7% of dropped stations are inland rather than 64% in the sample. This represents a bias towards dropping inland areas and retaining coasts. If you disagree with this reading, you could make the whole point clearer simply by showing the coastal/inland tabulations at some key intervals.
In Table 4 and surrounding text you appear to be arguing that the distribution is not consistent with falling altitude. Looking at your numbers, compared to the entire sample (Table 3), the sample that disappeared post-1990 has a slightly lower median altitude (183 vs 192) but has a higher mean altitude (441 vs 420), a higher 3rd quartile (589 vs 533) and almost the same max (4613 vs 4670). This indicates to me that the distribution of the 1990 fatalities was skewed to higher altitudes compared to the whole population. Then the stations dropped after 2005 are even more skewed: the mean is 510 (vs 420 for the whole sample) and the 3rd quartile is 681 (vs 533 for the whole sample). You then conclude: “That hardly supports the contention of thermometers marching out of the mountains.” Well, first, that’s a straw man. Second, your own numbers show that the distribution of dropped thermometers was, in fact, skewed to higher altitudes.
Another straw man: In one comment above Mosher says, presumably referring to my paper, “The claim in the paper was that the drop in altitude “comprimised the integrity of the data”.” Where is that quote from? What I said was “the failure to maintain consistent altitude of the sample detracts from its statistical continuity.” That’s a pretty carefully-phrased assertion, especially since I said it in the context of a series of charts showing the magnitude of discontinuities in the nature of the sample around 1990 and 2005. So far you are not convincing me that the altitude distribution remained continuous across those intervals.
Going on to Table 6, the 4-way division of landforms is intrinsically less interesting because it does not directly map onto altitude. Why look at these terms as indicators of altitude, when altitude itself is available? Presumably a “Mountain Valley” is itself a high elevation site, at least compared to a coastal valley. Likewise Flatland can be high-elevation flatland or low-elevation flatland. Your table is hard to read, again because it doesn’t have a date to define the sample with, it doesn’t show a before/after comparison, and the % calculations are not defined. Since columns don’t add to 100% we can eventually figure out what you’re doing, but it isn’t what you need to do to make your argument. You want to argue, presumably, that the different land forms each lost a similar percentage of their stations. But your numbers show that the losses were not similar across categories, they were Flat 34.5%, Hilly 37.8%, Mountain 29.5%, MV 42.2%.
On the 3rd issue, at one point you say “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.”
– No, even if their trends are the same, there can still be a bias in the overall average, if the discontinuity in the sample causes the mean anomaly to change in a way that cannot be corrected.
Regarding the comparisons of trends, I think you have made a reasonable case that, where they can be compared and all else being held equal, high elevation sites in the GHCN sample have similar or slightly higher trends than low-elevation sites. Whether this is a general principal, I do not know. I am aware of Christy’s work on California valley-vs-mountainous trends, and it seems plausible to me that his findings may apply more generally.
But as regards deciding whether “it matters”, this is very difficult and I would caution against making strong statements, since it’s easy to fail to find an effect when you lack the data to measure it directly. In effect you have 2 sets of data: X and Y. X terminates at 1990, Y continues in the form of y. And we know that X and Y have different sample characteristics. If you only use the information contained in y, Y and X to estimate x (the continuation of X after 1990), depending on how you do it you will tend to estimate it based on the portion of X explained by Y, in other words the correlated portion. Ideally what you should have is some other data, W, which is correlated with X but not with Y, and which continues after 1990. Then you can do some proper statistical modeling. It’s a mistake to be overly nihilistic about the surface data, but it’s also a mistake not to properly quantify the possible biases.
As you have alluded to, the big issue needing attention is (in my terminology) the application of ex post rather than ex ante testing to assess the quality of adjustments, gridding, and all the rest. The testable hypothesis is: If the surface data have been properly adjusted, and If the GCM’s are reliable in telling us the dominant influences on surface temperature trends, Then after the adjustment process is complete we should observe certain patterns in the data and not observe others. I expect you all know my thinking on this, based on the research I have done up to now, and why the IPCC’s treatment of the topic in the AR4 was so horrendously inadequate. I have 4 more projects in the works on this: one nearly in print, one in prep for a conference in the fall, and two just at the data collection stage.

I’m currently working on a temp reconstruction with the criteria that it shows the average, max, and min.
The way I’m getting over the station dropout issue is that I’m doing the averaging at the very last step. I’m using the daily GHCNv2 unadjusted.
From that I’m taking max/min temps with the only criteria being that they don’t exceed world records and that they’re balanced. I add all of these temps (they’re in tenths of a degree so they are all integers) together for a given year (I just want to see yearly for the moment) and record the number of observations. These two figures along with the min and max for all the years that station records are then stored in a file.
When I need to calculate a region or country, I open up the temp station files for that area and then add all the temp sums and observations for each year and only then compute an average for a year.
If stations drop out or are added, all that happens is the overall sum of temps and the observation count rise or fall in step with each other, and the computed average is valid. It’s a much less compute intense way of getting a Global Average Temp, with the advantage that none of the information is lost.
I’m still about a month away from releasing it (I still have to earn a living 🙁 ), but what I can say is that it does show that min temps have risen more than max temps, and that that is what seems to be driving global temps up. This is not a signature one would expect from CO2 induced warming.