Note: Steve McIntyre is also quite baffled by the Marcott et al paper, finding it currently unreproducible given the current information available. I’ve added some comments from him at the bottom of this post – Anthony
Guest Post by Willis Eschenbach
I don’t know what it is about proxies that makes normal scientists lose their senses. The recent paper in Science (paywalled of course) entitled A Reconstruction of Regional and Global Temperature for the Past 11,300 Years” (hereinafter M2012) is a good example. It has been touted as the latest hockeystick paper. It is similar to the previous ones … but as far as I can see it’s only similar in how bizarre the proxies are.
Nowhere in the paper do they show you the raw data, although it’s available in their Supplement. I hate it when people don’t show me their starting point. So let me start by remedying that oversight:
Figure 1. All of the proxies from M2012. The colors are only to distinguish individual records, they have no meaning otherwise.
I do love the fact that from that collection of temperature records they draw the conclusion that:
Current global temperatures of the past decade have not yet exceeded peak interglacial values but are warmer than during ~75% of the Holocene temperature history.
Really? Current global temperature is about 14°C … and from those proxies they can say what the past and present global average temperatures are? Well, let’s let that claim go for a moment and take a look at the individual records.
Here’s the first 25 of them:
Well … I’d start by saying that it seems doubtful that all of those are measuring the same thing. Panel 3/1 (row 3, column 1) shows the temperature decreasing for the last ten thousand years. Panels 4/4 and 4/5 show the opposite, warming for the last ten thousand years. Panel 4/3 shows four thousand years of warming and the remainder cooling.
Let’s move on to the next 25 contestants:
Here we see the same thing. Panels 1/1 and 4/1 show five thousand years of warming followed by five thousand years of cooling. Panel 1/5 shows the exact opposite, five thousand cooling years followed by five thousand of warming. Panel 4/5 show steady warming, panel 5/2 shows steady cooling, and panel 2/2 has something badly wrong near the start. Panel 2/4 also contains visible bad data.
Onwards, we near the finish line …
Panel 2/1 shows steadily rising temperatures for ten thousand years, as does panel 3/4. Panels 4/1 and 5/1, on the other hand, show steadily decreasing temperatures. Panel 4/2 has a hump in the middle. but panel 1/2 shows a valley in the middle.
Finally, here’s all the proxies, with each one shown as anomalies about the average of its last 2,000 years of data:
Figure 5. All Marcott proxies, expressed as anomalies about their most recent 2,000 years of record. Black line shows 401-point Gaussian average. N=9,288.
A fine example of their choice of proxies can be seen in the fact that they’ve included a proxy which claims a cooling about nine degrees in the last 10,000 years … although to be fair, they’ve also included some proxies that show seven degrees of warming over the same period …
I’m sorry, guys, but I’m simply not buying the claim that we can tell anything at all about the global temperatures from these proxies. We’re deep into the GIGO range here. When one proxy shows rising temperatures for ten thousand years and another shows dropping temperatures for ten thousand years, what does any kind of average of those two tell us? That the temperature was rising seven degrees while it was falling nine degrees?
And finally, their claim of turning that dogs breakfast shown in Figure 1 into an absolute global temperature and comparing it to the current 14°C average temperature estimate?
Don’t make me laugh.
I say the reviewers of this paper didn’t use their Mark I eyeball. The first thing to do when dealing with a multi-proxy study is to establish ex-ante criteria for the selection of the proxies (“ex-ante” meaning choose your criteria before looking at the proxies). Here are their claimed criteria …
This study is based on the following data selection criteria:
• Sampling resolution is typically better than ~300 yr.
• At least four age-control points span or closely bracket the full measured interval.
• Chronological control is derived from the site itself and not primarily based on tuning to other sites. Layer counting is permitted if annual resolution is plausibly confirmed (e.g., ice-core chronologies). Core tops are assumed to be 1950 AD unless otherwise indicated in original publication.
• Each time series spans greater than 6500 years in duration and spans the entire 4500 – 5500 yr B.P. reference period.
• Established, quantitative temperature proxies
• Data are publicly available (PANGAEA, NOAA-Paleoclimate) or were provided directly by the original authors in non-proprietary form.
• All datasets included the original sampling depth and proxy measurement for complete error analysis and for consistent calibration of age models (Calib 6.0.1 using INTCAL09 (1)).
Now, that sounds all very reasonable … except that unfortunately, more than ten percent of the proxies don’t meet the very first criterion, they don’t have sampling resolution that is better than one sample per 300 years. Nice try, but eight of the proxies fail their own test.
I must say … when a study puts up its ex-ante proxy criteria and 10% of their own proxies fail the very first test … well, I must say, I don’t know what to say.
In any case, then you need to LOOK AT EACH AND EVERY PROXY. Only then can you begin to see if the choices make any sense at all. And in this case … not so much. Some of them are obviously bogus. Others, well, you’d have to check them one by one.
Bad proxies, bad scientists, no cookies for anyone.
Steve McIntyre writes in a post at CA today:
Marcott et al 2013 has received lots of publicity, mainly because of its supposed vindication of the Stick. A number of commenters have observed that they are unable to figure out how Marcott got the Stick portion of his graph from his data set. Add me to that group.
The uptick occurs in the final plot-point of his graphic (1940) and is a singleton. I wrote to Marcott asking him for further details of how he actually obtained the uptick, noting that the enormous 1920-to-1940 uptick is not characteristic of the underlying data. Marcott’s response was unhelpful: instead of explaining how he got the result, Marcott stated that they had “clearly” stated that the 1890-on portion of their reconstruction was “not robust”. I agree that the 20th century portion of their reconstruction is “not robust”, but do not feel that merely describing the recent portion as “not robust” does full justice to the issues. Nor does it provide an explanation.
Read Steve’s preliminary analysis here:
[UPDATE] In the comments, Steve McIntyre suggested dividing the proxies by latitude bands. Here are those results:
Note that there may be some interesting things buried in there … just not what Marcott says.
Also, regarding the reliability of his recent data, he describes it as “not robust”. It is also scarce. Only 0.6% of the data points are post 1900, for example. This raises the question of how he compared modern temperatures to the proxies, since there is so little overlap.
Finally, about a fifth of the proxies (14 of 73) have the most recent date as exactly 1950 … they said:
Core tops are assumed to be 1950 AD unless otherwise indicated in original publication.
Seems like an assumption that is almost assuredly wrong. I don’t know if that’s a difference that makes a difference, depends on how wrong it is. If we take the error as half the distance to the next data point for each affected proxy, it averages about ninety years … pushing 1950 back to 1860 … yeah, I’ll go with “not robust” for that.
[UPDATE 2] Yes, I am shoveling gravel, one ton down, six to go … and I do get to take breaks. Here’s the result of my break, the Marcott proxies by type:
Best to all,