Monday, August 1, 2022

Claim: A new bombshell report found that 96% of ConUS NOAA temperature is corrupted (false)

Anthony Watts has a new report on US temperature stations. He had 2 pinned posts - the announcement here and a 1hr Heartland video presentation here. The subheading to the announcement reads:

"Official NOAA temperature stations produce corrupted data due to purposeful placement in man-made hot spots
Nationwide study follows up widespread corruption and heat biases found at NOAA stations in 2009, and the heat-bias distortion problem is even worse now"

Dramatic stuff. As Victor Venema pointed out, it is the tenth anniversary of another dramatic announcement, when Anthony said he was suspending everything to rush out a paper on the bad siting of stations. The paper still has not appeared.

In this post, I'll comment briefly on the report, but mainly will discuss the more complete refutation, which is that the resulting ClimDiv US average is very close indeed to that derived from the independent purpose-built network USCRN.

The report.

It is mainly taken up with photos of various sites that WUWT observers found fault with. But the bare statistics are that they looked at 128 stations, out of a total of well over 10,000 that are currently being used, and found 5 that they deemed satisfactory. That gave rise to the headline claim "MEDIA ADVISORY: 96% OF U.S. CLIMATE DATA IS CORRUPTED".

The report is a sequel to an earlier similar study in 2009 which has been fairly central to WUWT's operations. But, aside from the question of whether the stations sighted were really as bad as they claim, there is a further question of whether they are representative. On the face of it, no. The earlier report was in the days of USHCN (pre-ClimDiv), which was a set of 1218 stations from which the national average was derived. This time, 80 of the sample were from that set, leaving only 48 from the remaining 10000+ stations. That might not be so bad, were it not that there is a history, much discussed in the report, of those stations in the earlier report. There is plenty of scope for that to be a biased sample, and there is nothing in the report to show how bias was avoided. However, this is not the main reason for doubting the report.

Comparison with USCRN

USCRN is a network since about 2005 which has about 115 stations in the continental US (conUS) built in an array intended to be representative for the region, including an avoidance of urban activity. WUWT generally thinks well of this (as do we all). The announcement said:
"“By contrast, NOAA operates a state-of-the-art surface temperature network called the U.S. Climate Reference Network,” Watts said. “It is free of localized heat biases by design, but the data it produces is never mentioned in monthly or yearly climate reports published by NOAA for public consumption."

Given that fondness, it was natural that people should ask how ClimDiv, from the supposedly corrupted stations, compared with USCRN. Well, that is interesting. There is a very close match, as the NOAA Temperature Index page shows:

You can see little bits of blue (ClimDiv) peeping out from behind the USCRN, but mostly they totally run together. This may be a bit clearer with the annual data:

Now this comparison is not made anywhere in the report or associated writings. The monthly plot is even featured on the front page (right panel) of WUWT, but the ClimDiv part has been removed.

What is hinted in the report is that the correspondence was achieved by unstated adjustments. Here is Anthony's response to my posting the plot:

"They use the USCRN data to ADJUST nClimDiv data to closely match. And there is only 17 years of it, which means the past century of data is still as useless and corrupted as ever."

The rest of this post is mainly discussing that. No evidence is given, of course, and it's wrong. But it's also absurd. The heading quoted above says that NOAA is deliberately biasing the stations to inflate warming, and yet, the explanation goes, after all that they throw out that effort and adjust to the USCRN.

As for the 17 years, the survey itself was in 2022. It tells about current conditions, and if they are so bad, it is now that they bad results should be appearing. Faults of more than 17 years ago will not be illuminated by scrutinising current stations.

Is the averaging process adjusted?

I do my own calculations here. I described two years ago how I used the monthly average station data provided by both USCRN and GHCN (proxy for ClimDiv) to emulate the NOAA calculations. The one-month calculation, with comparative maps, is here, and the time series calculation is here. The key graph which corresponds to the NOAA monthly graphs, is here. Note that my results are in °C, while NOAA's are in °F.

In the legend, USCRN and CLIMDIV are my rendering of the NOAA data as in the earlier plots. MoyCRN is my calculation of the average from the station USCRN data. GHCN_u is the calculation using the ConUS GHCN V4 unadjusted station data, and GHCN_a uses GHCN V4 adjusted data.

And they all move along together, whether derived from USCRN, ClimDiv/GHCN, calculated by NOAA or me, and whether or not adjusted with GHCN. Again the different behaviour is a bit clearer with a 12 month running mean plot:

So is the station data rigged?

Presumably that would be the claim. It isn't, of course, and as said above it would be completely pointless to do so. But it actually can't be. All the daily data is available within a few days at most of being read. It would be quite impractical to adjust the ClimDiv on even that timescale to match USCRN which would probably be not yet available. And of course the posted daily data comes from data which is mostly posted promptly hourly.

Plus the ClimDiv comes from 10000 or so different operators. They can't all be in the conspiracy, and they can see what is happening to their data. The conspiracy notion is ridiculous.

What if ClimDiv and USCRN were both wrong?

I pushed this line of argument, and this is where it tended to come down to. Don't trust any of 'em. But the causes of error would be quite different - supposedly various bad siting issues for ClimDiv (of many kinds)

Update - difference plot and trend

Prompted by comments, I have plotted the difference USCRN-ClimDiv. Note that the range is much less than the earlier plots. 

Now there is structure to the difference, and a statistically significant upward trend, as Mark B noted in comments. However, it does not indicate "corrupted data due to purposeful placement in man-made hot spots". It goes the other way - ClimDiv is warming more slowly than USCRN!

Appendix - calculation method

I'll give a quick summary of my spatial averaging method here; more details are in my earlier post. It is much less elaborate than NOAA's, but I think quite accurate. I create a fine grid - 20km is a good compromise, 10 km takes longer. Many cells have no station data, especially of course with USCRN. Then I apply the principle that stations without data are assigned a temperature equal to the average of the neighbours. Of course, the neighbours may also be unknown, but I can write down a large sparse matrix expressing these relations, and solve it (conjugate gradients). This is equivalent to solving the Laplace equation fixed at cells with data.

A test of the method is whether it can give spatial maps like those of NOAA. It can - I did that in my earlier post. Here is my comparison for January 2020:

Of course, another test is the good congruence of the time series above.


  1. As we noted in our 2016 paper, if you rerun the NOAA pairwise homogenization algorithm excluding USCRN stations it has no effect on the agreement between the two datasets:

    1. Thanks, Zeke
      I see you got the same conclusion with USHCN, that the overall average agrees well with USCRN with or without homogenization (Fig 1). And you say that Menne et al and Diamond found the same. This is not so true for T_max. Interesting. So to see a (good) effect of homogenization, you have to go to a pairwise level.

  2. Thank you for this analysis. One suggestion, one question.

    The suggestion: add linear trend calculations for each of the time series. (I recognize that for a 17 year series for a tiny piece of the earth, the absolute trend isn't meaningful, but it is a nice way to allow for numerical comparisons between the different trends rather than "eyeball" comparisons). Q

    The question: I can believe that Watts sincerely believes that the siting issues he's identified are a problem (I agree that having USCRN to complement ClimDiv is a good idea, because it isn't clear to me how well pairwise homogenization would work for gradual urban heat island warming/degradation of thermometer housing as vs. station moves or other step changes, though my gut says it isn't a large effect, especially given consistent data regarding ocean heat uptake, stream and lake temperature, etc.)... but does he really believe the "purposeful placement" argument? (shades of the 2010 report, where he accused NOAA of "systematically and purposefully" dropping out high-latitude stations... even though, a) there was no evidence that the station dropout was intentional, and b) there being no evidence that station dropout would lead to a warming bias... but then, he also showed confusion about how temperature anomalies worked multiple times... I would think that being that publicly wrong multiple times would lead to a little humility about these kind of accusations, but I'd obviously be wrong) or that nClimDiv is actually adjusted to match USCRN (and even if it was, what about Berkeley Earth, which shows the same US trends?)?

    1. Thanks,
      Yes, I'll write about the trend. Uncertainty is high, but the trend to June 2022 was 0.46±0.27 C/decade for ClimDiv and 0.61±0.27 for USCRN - ie CRN was higher, but not statistically significant difference.

    2. The difference in trends is not statistically significant, but one can take the trend of the difference and I believe that is statistically significant. The effect is essentially to remove the "common mode" weather signal leaving the much lower variance of the difference in the sensor suites.

      I get a trend difference of 0.16 F/decade with 0.04 F/decade 2 sigma trend significance for USCRN-ClimDiv 2005 to present.

    3. Thanks, Mark
      Yes, that sounds right. I had promised to post a plot of the difference, and I'll do that. There does, by eye, seem to be a trend. Of course, it doesn't fit the WUWT narrative; USCRN shows more warming than ClimDiv.

    4. I've finally posted the difference plot and trend, agreeing with Mark B. The significance of the trend is interesting, and I don't really know what it means. However, it certainly doesn't mean a thumb on the scale exaggerating warming - ClimDiv is warming more slowly.

  3. Eli has always thought of the USCRN as Tom's Trick, Realizing that it was impossible to jump back into the wayback machine and improve the COOP stations in the US Historical Climate Network he set up the US Climate Reference Network. Even at the start, it was obvious that the data from what Watts considered awful stations tracked eerily close to the optimal reference network ones.

  4. "Now there is structure to the difference, and a statistically significant upward trend, as Mark B noted in comments. However, it does not indicate "corrupted data due to purposeful placement in man-made hot spots". It goes the other way - ClimDiv is warming more slowly than USCRN!"

    I really like this analysis but couldn't the significance be due to e.g. spatial sampling rather than adjustments? I'd probably resample the ClimDiv stations to match the USCRN sample size and try to ensure similar spatial characteristics. Do this a bunch of times and see what the trend difference distribution looks like.

    Way too much work, but I think we should interpret the significant ClimDiv-USCRN difference with care.

    MarkR (posting as anonymous)

    1. Thanks, Mark. I did a table of the five sets 2005-2019 (GHCN_u, GCHN_a and MoyCRN are mine). I hope it formats OK. Here are the relative slopes
      USCRN 0.000 0.114 0.114 0.134 0.049
      CLIMDIV -0.114 0.000 0.000 0.020 -0.066
      GHCN_a -0.114 0.000 0.000 0.020 -0.066
      GHCN_u -0.134 -0.020 -0.020 0.000 -0.086
      MoyCRN -0.049 0.066 0.066 0.086 0.000
      So GHCN_a (my adjusted) is very close to ClimDiv, but unadjusted is fairly close too. The trend of USCRN is not only higher than the big station sets, but also quite a lot higher than my calculation of CRN. Here are the standard errors:
      USCRN 0.000 0.013 0.014 0.014 0.009
      CLIMDIV 0.013 0.000 0.005 0.005 0.015
      GHCN_a 0.014 0.005 0.000 0.001 0.016
      GHCN_u 0.014 0.005 0.001 0.000 0.016
      MoyCRN 0.009 0.015 0.016 0.016 0.000

      If the anomalies really do create homogeneity, then different sampling shouldn't create a bias. That's ideal, of course.