Kinnard Arctic O18 Series

Kinnard et al 2011 report a highly hockey-stick shaped reconstruction of sea ice. (For a different perspective based on Holocene ocean sediments under the Ellesmere Island ice shelf, see the recent Antoniades et al (PNAS 2011).

Kinnard et al use a regression-based statistical methodology that looks at first blush to be a sort of inverse regression: i.e. of sea ice against a large number of proxies and “proxies”. In any such procedure, there is a serious risk of spurious fits. (After all, you can reconstruct a trend with a sufficient number of white noise series simply by brute force.) Kinnard et al attempt to guard against this by considering the RE statistic. Here they head straight into the home territory of Mannian statistics.

This does not mean that their methodology is wrong, let alone “WRONG”. (I haven’t examined it yet, let alone parsed it.) But it is worrying.

My own preferred technique for examining “multiproxy” data is to first divide the data into subsets of like proxies. The largest subset of Kinnard et al is the 22 O18 series (out of 69 proxies.) This turned out to be a remarkable subset in several ways and I spent much longer examining this data than I had expected or planned. Here is a graphic showing the ten long O18 series in Kinnard et al, arranged from west (Mt Logan, Alaska) to east. The left column shows Mt Logan plus four series in the Canadian archipelago.

The take-home points are that Kinnard et al have provided a huge/unprecedented increase in the number of archived Arctic O18 series – especially long high-resolution series. Whereas Mann et al 2008 had only two Arctic O18 series beginning prior to AD900, Kinnard et al have ten. Most of the ten aren’t new, but, to my knowledge, for the most part, they haven’t been available. And second, that the long Arctic O18 do not give the pronounced HS result of the final Kinnard diagram. Its provenance therefore lies elsewhere: either in other proxies or by sign-flipping and weighting in the regression methodology, resolution of which will require more analysis.

In the meantime, documentation of the new treasure trove of Arctic O18 seems worthwhile in itself.

I’ll look at the proxies from west to east, starting in Alaska.

Mt Logan: In the half-hemisphere from 100E to 80W (going east), Kinnard has only one O18 series – a long series from Mt Logan, Alaska published in 2006 by coauthor David Fisher. This series was discussed a number of years ago at CA: it had an unexpected decrease in O18 values in the modern period, attributed by the authors to changes in regional circulation. To my knowledge, this is the first use of this series in a multiproxy study. (It was not used in Mann et al 2008 or the Ljungqvist articles.

Bona-Churchill: Bona-Churchill is another long O18 series in this hemisphere, but is not used even though Lonnie Thompson is a coauthor. Some years ago, I commented on publication delay of this series (drilled in 2002), speculating that it would have been published in time for AR4 if it had gone the “right” way. Five years later, it remains unpublished and looks like it will remain unpublished through AR5. It is near Mt Logan and I speculated that it would, like Mt Logan, go the “wrong” way in the 20th century. A few years ago, I noticed a graph in an online workshop showing that this was the case. In any event, it wasn’t contributed to Kinnard et al (even though they show a series from Prince of Wales, Ellesmere Island of similar vintage.)

Eastern Canadian Archipelago
There are four long O18 series from the eastern Canadian archipelago, all more or less 80W. From north to south, Agassiz, Ellesmere Island (81N), the new Prince of Wales, Ellesmere Island (78N), Devon Island (75N) and Penny, Baffin Island (69N).

Agassiz: Mann (2008), Ljungqvist (2009) and Vinther (2010) use data ultimately deriving from a core drilled in the 1970s. Mann used the version that was on Fisher’s CD in the early 1990s (now also at NCDC) and “infilled” subsequent data (infill not shown here). Its first value was only 1340. Vinther 2010 re-dated Fisher’s Agassiz cores and combined the short series with less-resolved earlier data. It is negatively correlated to the earlier version, at some points being almost opposite. Kinnard uses a later series with values up to 1987. This is referenced to Fisher’s 1996 contribution to the NATO workshop. To my knowledge, this is the first digital appearance of this data. Christiansen and Ljungqvist 2011 say that they use the Vinther version rather than the very smooth series of L2009 but their data as used isn’t available.

The 1987 Fisher series shows an increase in the 20th century but only after a long secular decline, the increase merely returning to “precedented” values.

Here’s a graphic showing the post-1750 values, which are rather different.

Prince of Wales, Ellesmere Island (78N): This is a previously unpublished series drilled in 2002. Values are shown only until 1995. Again, it shows a 20th century increase but only after a long secular decline, and once again the increase is to precedented values:

Devon (75N): Kinnard have archived a long O18 from Devon Island up to 1994. Again, to my knowledge, this is the first digital appearance of this series. It was not used in Mann et al 2008, which doesn’t appear to have used a Devon O18 series, but used a melt and one other series from this site. Ljungqvist 2009 archived the Fisher series from the 1973 drilling program. Like the nearby Ellesmere Island O18 series, there is a long term secular decline, with an increase in mid-20th century to precedented values (though a decline in the closing values.)

Penny, Baffin Island (69N): Yet another “new” long series. Mann et al 2008 used a short version of this series based on a co.e from the 1970s (1761-1970). The “new” Kinnard version goes from 727 to 1992. Again it shows a sort of increase through the 20th century, but again only to precedented values. The secular decline observable in the more northerly cores is not observable

North of 75N: Kinnard has five O18 series from north Greenland (here north of 75N): two long series (NGRIP and B18) and three shorter series (Camp Century, B26 and B21 starting in 1242-1502). Mann et al 2008 used one short series from this area- Fisher’s composite of Camp Century, North Central and North Site (traceable to DELNORT3.STK) which goes only from 1761 to 1972. Ljungqvist (2009) appears to be a smoothed version of NGRIP.

Central Greenland 70-75N: Central Greenland between 70 and 75N has been a staple in multiproxy reconstructions. Kinnard has four cores from this area – two long cores (Crete ending 1974 and GISP2 ending 1987) plus two short cores (B16, Site A). Mann et al 2008 used the same version of GISP2 as Kinnard; they used a slightly different Crete version: Fisher’s combination of Crete with some nearby sites, extending it from 1974 to 1982. Ljungqvist has a GISP-related series that is smoother. Ljungqvist’s series from Alley combines borehole information with O18 information and needs to be construed as a borehole series rather than a O18 series simpliciter.

Greenland S of 70N: Kinnard has 3 Greenland sites south of 70N – one long series (Dye-3) and two short series (Dye-2 and 20D). Mann et al 2008 used a short series from Dye-2 (identical to the Kinnard version of Dye-2 which is not shown here). Ljungqvist 2009 used a smoothed version of Dye-3 ending in 1960. The Kinnard version of Dye-3 is identical to the Vinther 2010 version.

The effect of 20-year smoothing can be seen in the following graphic showing post-1750 values:

Other Grenland: The Renland series is somewhat east (71N, 27W) of the “Central Greenland” group shown here and starts in the medieval period. Mann et al 2008 also used a Vinther composite of Greenland sites from different regions ( Crete, Milcent and GRIP from the “central” group, Dye-3 from the “southern” group and Renland from the east.

“Eastern” Islands
Kinnard archived 4 O18 series from “eastern islands” – two from Svalbard (17-24E), one from Franz Josef Land (64E) and one from Severnaya Zemlya (95E) -none of which have appeared in previous studies. These series do not shed direct light on the MWP=modern comparison as none have values in the MWP. (I don’t know whether this is due to glacier flow – as a result of which many temperate glaciers e.g. Chile have only quite recent values – or due to reformation of the glacier in the LIA or both.) All are more recent than 1995 (but values are archived only to 1995).

Comparison to Kinnard Hockey Stick

If, as a first rough check, one compares a simple average of the long O18 series to the Kinnard reconstruction (inverted so that “warm” is up), one gets the graphic below – both shown in SD units for simplicity. There is nothing in the long O18 series that yields the pronounced Kinnard hockey-stick.

As noted in the preview, this means that the derivation of the Kinnard hockey stick must come from elsewhere: either from the other ice core proxies (Na+, melt, MSA), from the D’Arrigo tree rings or from the other proxies (a few sediments and documentary series) or from regression procedures. If so, one needs to check out which series have been heavily weighted and whether series have been flipped through the regression methods. Splicing is also something that needs to be looked at: for example, while the O18 series from the Eastern Islands don’t permit a direct medieval-modern comparison, has their stepwise introduction introduced a medieval-modern differential that does not exist in the long series?

Whatever the results of this analysis may be, Kinnard et al deserve credit for ensuring this remarkable expansion of the public archive of Arctic O18 series.


  1. Posted Dec 5, 2011 at 12:47 PM | Permalink

    Steve, I’d like to thank you for pursuing this type of work and urge you to continue it. The Climategate 2 stuff is fun for all–but it’s also something all can do. Nobody else seems to be doing this for public consumption, and I consider it an extremely valuable service.

    • Jim T
      Posted Dec 5, 2011 at 12:52 PM | Permalink


      • Skeptik
        Posted Dec 5, 2011 at 2:01 PM | Permalink

        Probably, carried by acclamation.

        • dearieme
          Posted Dec 6, 2011 at 8:50 AM | Permalink

          Hurray. Hear, hear. Bravo. Approving hubbub.

    • Posted Dec 6, 2011 at 2:03 PM | Permalink

      Completely agree !!!
      When yo guys “talk shop” here about all this stuff, I get lost quickly but Steve’s site really has opened my eyes to the “sausage making” end of CAGW reasearch !!!

  2. Posted Dec 5, 2011 at 1:04 PM | Permalink

    Thanks, Steve.

    You can find pages of references about how the isotope paleothermometer is supposed to work. However There is a substantial problem I think with the use of d18O in precipitation, which has never been mentioned in that pile.

    The main drivers for the d18O are supposed to be temperature at condensation and raining out, changing the isotope ratios in the remaining airmass, also known as the rayleigh effect.

    How about that “temperature at condensation”, is that representative for the ambient surface temperature? For climatologists maybe but not for meteorologists. They call the that the dew point, which is basically only a function of absolute humidity.

    here you can calculate it:

    Obviously comparing deserts and tropical rainforest, humidity is not related a lot to the temperature. So the isotopes in precipitation record basically the humidity at the source.

    Maybe you’d like to take a look at this.


    • BobN
      Posted Dec 5, 2011 at 9:51 PM | Permalink

      Andre – I too have always had trouble with O18 thermometry concept. Some isotope stuff (e.g., foram shells) seem pretty reasonable for paleothermometry and have actually had temperature correlations done in the lab. however, O18 seems more of a theoretical concept without a lot of hard empirical calibration.

      • Paul Dennis
        Posted Dec 6, 2011 at 2:19 AM | Permalink


        on the contrary, there is a lot of empirical evidence for the correlation between mean annual precipitation 18-O (and 2-H) composition and mean annual surface temperature. This was first reported by Dansgaard (Dansgaard, W., 1964. Stable isotopes in precipitation. Tellus.). One can also make an almost trivial model of global precipitation, using the Clausius-Clapeyron relationship for the liquid-vapour phase boundary, the experimentally determined heat of vapourisation (sublimation for ice) and the experimentally determined fractionation factors for isotope partitioning between liquid (ice) and vapour and find that the predicted isotope-temperature relationship is the same as that observed empirically.

        This empirical observation is based in plotting compositions of sites that cover a wide geographic range (e.g. mid-latitudes to poles). It represents an areal correlation. The key question is if you stay at a single site and look back in time (e.g. with an ice core) does the same relationship hold. I think it doesn’t. For example comparing noble gas recharge temperatures (groundwaters) or borehole temperatures (ice core) with isotopically determined temperatures one finds a different relationship.

        • Steve Garcia
          Posted Dec 6, 2011 at 3:11 AM | Permalink

          paul –

          Can you do a bit of explaining on this:

          …on the contrary, there is a lot of empirical evidence for the correlation between mean annual precipitation 18-O (and 2-H) composition and mean annual surface temperature. This was first reported by Dansgaard (Dansgaard, W., 1964. Stable isotopes in precipitation. Tellus.).

          If I understand this correctly, (assuming 18-O is pretty constant in precipitation, which may not be the case) you are saying that mean annual precipitation and mean annual temperature is empirically shown to be correlated.

          [Steve Mc: you’re totally missing the point. 18-O is NOT constant in precipitation and that’s what’s being measured.

          (Numbers from Wikipedia) Now, I happen to know that Guadalajara (36.26 inches/yr) and Chicago (38.35 inches/yr)have almost the same annual precipitation. If I average Milwaukee (34.81 inches/yr) in with Chicago (giving 36.58 inches/yr), the annual precipitation is very close, indeed. Combined, Milwaukee and Chicago have an annual max/min average of 49.3F, while Guadalajara is at 64.9. I am sure I could pick all kinds of locations that mis-correlate like that, but I happened to know the Chicago-Guadalajara near-equal precip, and tried to find an even closer match in the US upper midwest.

          Obviously, I am reading what you are saying wrong. Or am I? I am not doubting, but am not hearing this evidently the way you said it.

        • Paul Dennis
          Posted Dec 6, 2011 at 11:21 AM | Permalink

          Steve G, Steve Mc is right in pointing out that the 18-O content of precipitation is not constant. As water evaporates the lighter isotopes (16-O and 1-H) are preferentially partitioned into the vapour phase. The reverse happens on condensation. Because of the partitioning, which is well understood, it’s possible to model the atmosphere by a simple Rayleigh distillation process. When this is done using experimentally determined values for the isotope fractionation factors one observes an 18-O/temperature dependency of about 0.7 per mille per degree C. This relationship holds out pretty well across high latitude sites. It doesn’t hold out well at low latitudes where rainfall tends to be monsoonal, with a strong convective component. Here the dominant relationship is one of a correlation between the isotope composition and amount of rainfall.

          What I have described is a first order model. The situation in practise is more complex. Evaporation tends to occur under non-equilibrium conditions at less than 100% humidity and condensation under equilibrium conditions at 100% humidity. Andre is concerned about humidity/aridity effects on rainfall isotope composition. I’m not so sure that these effects aren’t wrapped up in the temperature and rainout effect. i.e. High latitude rainfall amounts in polar areas tends to be low simply because the degree of rainout is very large at these latitudes. By the time an air mass has cooled significantly it contains very little vapour.

          Andre has some interesting postings at another website where he begins to describe/discuss his ideas about aridity. I haven’t read them yet so can’t pass an informed comment on his ideas. They sound interesting and worth spending some time on, but this will have to wait for the Christmas break!

          For anyone who is interested I am developing a simple model in Mathematica that describes the isotope composition of precipitation. When finished I can make it available in CDF format (Computable Document Format) which will allow one to explore the differing effects of humidity in source regions, temperature etc. on the isotope relationships.

        • Steve McIntyre
          Posted Dec 6, 2011 at 11:55 AM | Permalink

          One point that would be worth examining. As I recall, the ice cores from the Svalbard, Novaya Zemlya showed evidence of the glacier not existing in the MWP. If any readers can refresh quotations on that, it would be appreciated.

          Because there isn’t much HS-ness in the long records (Yamal, of course), I suspect that somewhere within their complicated methodology is a splice in which these short O18 records with a pronounced HS end up creating a big HS that isn’t in the records that cover both periods of interest.

        • Paul Dennis
          Posted Dec 6, 2011 at 12:20 PM | Permalink

          Steve Mc, I don’t have time to follow this up right now but something looks odd in the Lomonosovfonna deuterium record. See which contains a long 18-O record for Lomonosovfonna. The delta 18-O wiggles around -15 to -16 per mille (vsmow). I would expect the deuterium record to lie around -120 to -110 per mille and not the -160 to -150 per mille shown in the record. Also the shape of the deuterium record with the pronounced dip is not immediately apparent in the 18-O record. The dip in delta D corresponds to about 3 per mille in delta 18-O. It isn’t there!

          I also think there might be subtle differences in the Austfonna record you show and that in the poster.

        • Hu McCulloch
          Posted Dec 6, 2011 at 1:14 PM | Permalink

          As I recall, the ice cores from the Svalbard, Novaya Zemlya showed evidence of the glacier not existing in the MWP.

          I don’t know about Novaya Zemlya, but the Svalbard Lomonosovfonna ice field appears to have begun in 1130 (or 1126 according to the new paper’s data). See
 (also mentioned below).

        • Paul Dennis
          Posted Dec 6, 2011 at 3:20 PM | Permalink

          My mistake Steve! I didn’t see the decimal point in the Lomonosovfonna plot that you showed above. disregard my comment about deuterium.

        • Hu McCulloch
          Posted Dec 6, 2011 at 12:36 PM | Permalink

          Paul — Thanks for dropping by. Your expertise is always appreciated.

          This relationship holds out pretty well across high latitude sites. It doesn’t hold out well at low latitudes where rainfall tends to be monsoonal, with a strong convective component. Here the dominant relationship is one of a correlation between the isotope composition and amount of rainfall.

          But in the Arctic, isn’t the amount of sea ice on the Arctic Ocean going to be a big factor in how far away the atmospheric water vapor has come from, and therefore how much d18O it still contains before the deposition takes place?

          Some of these cores, like Mt. Logan, aren’t on the Arctic Ocean, but many of the others are. How then do we know the d18O in the ice is responding to temperature an not directly to sea ice coverage? Would the two effects reinforce each other or offset each other?

          (I believe that condensation only refers to the vapor-liquid transition, and that the reverse of sublimation is deposition.)

        • Paul Dennis
          Posted Dec 6, 2011 at 1:11 PM | Permalink

          Hu, it’s an interesting comment. My understanding is that some of the Arctic cores, notably Svalbard examples have been interpreted in terms of sea ice extent and how far the moisture has to travel from the source region to the site of precipitation. As I said the kind of model I described is a first order one and is based on the fact that most of the global atmospheric water vapour is sourced from the tropics where there is an excess of evaporation over precipitation. Meridional transport then results in cooling and rainout that leads to the isotope temperature relationship. This is a gross simplification and is based on weighted mean annual precipitation isotope composition and mean annual surface temperature.

          The relationship is something like d18-O = 0.67 x T -13.5.

          Of course individual rain events might not conform to this depending on the synoptic weather patterns.

          There is some data for Svalbard and individual rain events seem to plot close to this line with a relationship of d18-O = 0.57 x T – 12 ( It would be interesting to see what the annual data plots like.

        • Hu McCulloch
          Posted Dec 6, 2011 at 1:19 PM | Permalink

          Paul — Thanks.

          But these presumably hold nearby sea ice constant at their current levels. Is there enough data to add Barents sea ice to the equation?

          Also, does the same formula hold for snow as for rain?

        • Paul Dennis
          Posted Dec 6, 2011 at 1:33 PM | Permalink

          Hu, the formula I gave was first proposed by Dansgaard. The reference is: Dansgaard, W. 1964. “Stable isotopes in precipitation.” Tellus, vol 16, 436-468. It includes precipitation and snow samples. In fact most of the data are snow samples. This is to be expected because there isn’t a very large change in the isotope fractionation factor between liquid and vapour and between ice and vapour. If I knew how I’d post his plot here.

          Interestingly the north Atlantic stations, e.g. Svalbard, Reykjavik etc, plot on a slightly different trend.

          In answer to your question about the Barent’s sea I don’t know. I think it would be an interesting exercise to collate the precipitation data for the Arctic ocean sites and see what effect if any sea ice extent is controlling the isotope compositions. I’d be interested in looking at the deuterium excess to see if this correlates with ice extent. d-excess is defined as deltaD – 8 x delta18-O. I tried this with an MSc student using an ice core from the Antarctic Peninsula to see if the sea ice extent in the Bellinghausen Sea was measurable. Our results suggested that on a century time scale we couldn’t detect any change. This work is not published yet.

        • Pat Frank
          Posted Dec 6, 2011 at 2:33 PM | Permalink

          Paul, the spread between your two equations is about 1 C at constant d18-O. In use of d18-O for reconstruction of the temperature fields of past climates — based on thermal isotopic fractionation and not on mere rescaling to the instrumental record — would you say that (+/-)1 C is a fair lower limit of accuracy?

        • Paul Dennis
          Posted Dec 6, 2011 at 3:08 PM | Permalink

          Pat, personally I think isotopes in ice, or palaeoprecipitation (e.g. groundwater, fluid inclusions in speleothems) are important sources of palaeotemperature estimates, but would be aware that there are limits to the accuracy and precision of those estimates. Your suggestion of +/- 1 degree sounds as though it’s in the right ball park for such proxies, if even then a little optimistic. When looking at precipitation at specific sites one needs to understand what is happening in the source region, the relative importance of different air mass trajectories (synoptic weather patterns) and their seasonal variation, changes in ocean isotope composition due to continental ice volume effects (on glacial-interglacial timescales) and so on.

          Further, the relationships I gave above were for different sets of observations. The Dansgaard relationship is obtained by plotting weighted mean annual isotope composition versus mean annual surface temperature for a large number of stations at different locations ranging from the South Pole, to Greenland, north Atlantic Islands, coastal European sites etc. The relationship for Svalbard was obtained by plotting the isotope composition of ‘rainfall events’ versus local surface temperature at the time of the rainfall. Because they were obtained differently I wouldn’t directly compare them.

          A key advantage of using isotopes is that at least we have some a priori idea of the changes we might expect as a result of thermodynamic considerations of isotope partitioning between phases.

        • Pat Frank
          Posted Dec 6, 2011 at 7:52 PM | Permalink

          Thanks, Paul, for a very informative comment. I’ve saved it to disk for future reference.

          And I totally agree with your final paragraph, worthy of repeating in full, that, “A key advantage of using isotopes is that at least we have some a priori idea of the changes we might expect as a result of thermodynamic considerations of isotope partitioning between phases.

          Thanks so much for making that point.

          dO-18 fractionation, like that of other isotopes, is understood in terms of a powerful and falsifiable physical theory. That means the zeroth-order temperature model can be, and has been, tested against experiment. This places proxy isotope fractionation models squarely within the realm of science, at least in principle, and makes them categorically distinct from any so-called proxy thermometric method that relies upon mere numerical normalization to the instrumental record.

        • Steve McIntyre
          Posted Dec 6, 2011 at 2:51 PM | Permalink

          The Fisher 2006 article on Mt Logan adopted a similar approach – delO18 depletion depended on distance from moisture source.

        • Geoff Sherrington
          Posted Dec 9, 2011 at 5:29 AM | Permalink

          In terms of Dansgaard’s mechanisms, I would not expect isotope composition to change in snow deposited from one side of the Antarctic to the opposite, because at a given time there is nothing very different happening in the atmosphere that is of relevance – except, to add further weight to the assertion, a huge amount of mixing in the atmosphere by those winds that do not rise above freezing point. I’m thinking that there is no fractionation from evaporation (as happens nearer the tropics) because it’s too cold for evaporation; and there is no condensation, because that process exhausted itself before the atmosphere moved to positions above the ice sheet, or at its edge. If anything significant is happening, it is sublimation, which I see as adding material of unknown isotope composition to the highly mixed atmosphere, and a complication with the theory of calculationg the ratios in what is (or should be) left behind. In short, most processes don’t happen because we are below temperatures of interest.
          This does not conflict with your Mt Logan, southern Yukon, comment.

        • BobN
          Posted Dec 6, 2011 at 3:17 PM | Permalink

          Paul – Thanks for the link to the Svalberg report and data. I agree that the delO18 is modertaely correlated (r=~0.6), but for any given delO18 value, the range of possible temperatures is pretty large (~15degC), so, at least for individual rainfall events, any temperature estimate is going to be poorly constrained. Is the concept that, on an annual basis, the errors will average out to allow a more tightly constrained estimate of the average annual temperature? And has anyone actually evaluated this.

          I look into the Antartic dissertation when I have more time and also try to track down a open source copy of the Dansgaard paper. I remember reading a number of papers on delO18 thermometry back in grad school in the 80s for a project and was somewhat skeptical at the time as well.

        • Paul Dennis
          Posted Dec 6, 2011 at 3:30 PM | Permalink

          BobN, Your right there is a big spread in isotope composition of individual rain events at any given temperature. This doesn’t surprise me and is what is normally observed. I suspect that averaging over monthly and annual cycles will considerably tighten up the relationship.

        • Steve Garcia
          Posted Dec 6, 2011 at 8:27 PM | Permalink

          Try this link to the Dansgaard paper, “Dansgaard, W. 1964. “Stable isotopes in precipitation.” Tellus, vol 16, 436-468”:

        • BobN
          Posted Dec 7, 2011 at 4:42 PM | Permalink

          Thanks for finding that Steve. I give it a read.

        • Andre
          Posted Dec 6, 2011 at 9:36 AM | Permalink

          The point is not that temperature and isotopes don’t correlate, the problem is that isotopes also correlate with aridity/humidity shifts. As said, most essentially isotopes correlate with the condensation temperature, obviously the absolute humidity in the atmosphere also follows the seasonal pattern, which is much more pronounced in the arctic, which causes an indirect relationship with temperatures. This can be seen in the GNIP precipitation database, where correlations crumble in the moderate lattitudes and go completely haywire in the tropics.

          See also for instance the work of Michiel Helsen in Antarctica

          Click to access thesis.pdf

          Things however get a lot more obscure if we really zoom in on all the details about the periods where it matters, ie the last glacial transition and Younger Dryas etc.

        • Steve McIntyre
          Posted Dec 6, 2011 at 10:08 AM | Permalink

          Paul, I’d recommend caution in relying on borehole temperature reconstructions. These rely on inversion of near-singular matrices. In one style of inversion, the results are very sensitive to the number of retained principal components (a familiar CA problems). While I haven’t fully parsed all the methods, my impression is that the observed temperature profiles are compatible (in some statistical sense) with very different temperature histories.

        • Paul Dennis
          Posted Dec 6, 2011 at 11:09 AM | Permalink

          Steve, thanks for the interesting comment about borehole temperatures. It’s not something I’ve looked at closely but it sounds as though the final result might not be very robust with regard to retained PC’s. Not with standing this published work on UK aquifers shows that reconstructed glacial-interglacial temperature differences determined using stable isotopes invariably are considerably smaller than those obtained using noble gas mixing ratios as per my comment above, but this is getting off topic. The Kinnard et al archive looks as though it’s going to be very useful and thanks for bringing it to our attention.

        • Phil B
          Posted Dec 6, 2011 at 11:26 AM | Permalink

          Steve, I believe your comment here refers to non-ice borehole reconstructions.

          Steve: To other readers, PhilB was involved in our quick look at these methods a few years ago – where we did look only at non-ice “borehole” reconstructions. (Mostly they are from mineral exploration drill-holes.) Phil, it is my understanding that reconstructions from holes in ice present problems over and above the ones in rock and that the problems carry over. There is considerable movement within the ice that may or may not affect the reconstruction. (Hugo Beltrami thinks that it is a serious problem.) Parsing the methods in ice borehole reconstruction isn’t easy. One such study relied upon by the NAS panel in 2006 (by Clow on Law Dome, Antarctica) remains unpublished and data unavailable 15 years later.

        • Phil B
          Posted Dec 6, 2011 at 12:43 PM | Permalink

          Steve, I have not looked at the ice borehole temperature reconstruction only the non-ice. Will you be presenting the artic “instumental” ice record that Kinnard used? My memory of the artic ice record (~1850 – present) is that 30-40% of the data was infilled with an average. The data was measured with three different methods, but was conveniently spliced together for a contiguous look to the graph. I noticed that Kinnard et al was behind a $32 paywall. Is there a free access somewhere?

        • Curt
          Posted Dec 6, 2011 at 9:38 PM | Permalink

          In Dahl-Jensen’s work on the Greenland borehole reconstructions, she did not try to invert the matrix because of its near singularity. Instead, she brute-forced it with a Monte Carlo method of plugging in over three million possible solution vectors and seeing which agreed best.

          It strikes me that the ice boreholes largely avoid the possibility of water flow screwing up the conduction relationships, but add the complexity of glacier flow. D-J’s matrix included a model of that flow. I have no idea of how well.

        • Phil B
          Posted Dec 7, 2011 at 2:17 PM | Permalink

          Thanks Curt, obviously Dahl-Jensen hasn’t taken a linear algbra class. If the matrix was singular, then you can have 3 million different solution vectors with no change in your residuals. With a near singularity matrix with singular value ratios of a million or so, you can have 3 million different solution vectors with small changes to the residuals. The real issue is that model is ill posed and the results shouldn’t be used the way they have been. In linear system theory parlance, the model has an observability problem.

        • BobN
          Posted Dec 7, 2011 at 4:49 PM | Permalink

          Curt- I think you’re right regarding ice being better than bedrock borehole for possible temperature reconstructions since you don’t have the groundwater flow issue, fractures acting as discontinuities, or changes in lithology affecting the heat conductance. Having used temperature logs quite a bit in hydrogeologic investigations (to identify zones of groundwater influx/efflux), I was really suprised finding out a few years ago that anyone would try to use them for surface temperature reconstructions.

          As to ice, I think you are better off since all heat flow should be via conductance, but permafrost is probably the only situation where other complicating factors such as ice movement don’t come into play.

        • Hu McCulloch
          Posted Dec 7, 2011 at 5:55 PM | Permalink

          Steve — this problem sounds mathematically similar to one that arises in the estimation of multivariate stable distributions — see my paper on “Estimation of the Bivariate Stable Spectral Measure by the Projection Method” at .

          In my paper, the problem is to find a solution that should be nonnegative to a system of ill-conditioned linear equations whose exact solution is often of very mixed signs. But if you pose it instead as a quadratic program that minimmizes the sum of squared equation discrepancies subject to nonnegativity of the solution, it works like a charm.

          Similarly, the borehole problem is likely to give a very unsmooth and ill-conditioned exact solution. But it could easily be expressed as a quadratic program that simultaneously minimizes the roughness (measured by the sum of squared temperature changes), plus some tuning factor times the squared equation discrepancies.

        • Geoff Sherrington
          Posted Dec 9, 2011 at 5:32 AM | Permalink

          Anyone who has sat by a drill rig and thought about the process would not even start to think about temperature reconstructions. Apart from the ill-posed mathematics, the physical conditions are too diverse to be quantified.

  3. Kenneth Fritsch
    Posted Dec 5, 2011 at 1:13 PM | Permalink

    If I can read between the lines of “Canadian nice” here, I think SteveM has put out a few lines of what should be some very interesting future analyses.

    • JEM
      Posted Dec 5, 2011 at 4:01 PM | Permalink

      There is certainly a lot of “it’s not just where you pick your cherries but when you pick them” in these series.

  4. Don McIlvin
    Posted Dec 5, 2011 at 1:39 PM | Permalink

    If one compares the O18 long series and its seemingly consistent 2000yr downward slope, and some dendrochronology based long series that appear more wavy (Example; Some Finnish very long series also well to the North), there seems to be a fundamental inconsistency that needs explanation as to the adequacy in using such series as temperature proxy.

    • Frank
      Posted Dec 6, 2011 at 12:08 PM | Permalink

      A consistent 2000-year downward trend certainly makes sense when we take into account the drop in solar irradiation at high Northern latitudes that has been occurring throughout the Holocene. The standard Milankovitch graph shows the variation in July irradiation at 65 degN, but total annual radiation follows a similar cycle. The maximum change in annual radiation is about 13% or about 30 W/m2. Eyeballing the curves, annual irradiation in high Northern latitudes has fallen about 16 W/m2 (7%) over the last 8000 years or 4 W/m2 (2%) over the last 2000 years. Meridional mixing will diffuse some of the extra warmth to lower latitudes (where the Milankovitch cycle has much weaker effects), but a 2% change in radiation is comparable to the forcing from 2X CO2. .

      • Don McIlvin
        Posted Dec 8, 2011 at 1:07 AM | Permalink

        Finnish 7644 yr Scots Pine Chronology

        The Finnish long chronology is made up of tree ring data from northern Lapland about 69N to 70N latitude. So the solar influence would I think be similar.

        The point I was trying to make is this dendro study is waving showing MWP and LIA, and other similar periods. The O18 seems does not seem to do so.

        Maybe I am just looking at it incorrectly.

  5. Paul Linsay
    Posted Dec 5, 2011 at 1:39 PM | Permalink

    Whie noise, all of it. Hurst exponent = 1/2.

    • Posted Dec 6, 2011 at 4:46 PM | Permalink

      I have estimated the Hurst Coefficient for the Series 18 data given in the SI for the paper. Using sample lengths from about 30 years to 992 years ( I used 1984 of the data points ), I get the Hurst Coefficient to be about 0.57. That’s close to the value for random data, which is 0.50.

      The Figure below should be linked to a larger Figure that shows the results of the calculation.

  6. Bruce
    Posted Dec 5, 2011 at 2:02 PM | Permalink

    “with an increase in mid-20th century to precedented values”

    Precedented Values! Phrase of the year.

  7. Andy
    Posted Dec 5, 2011 at 2:27 PM | Permalink

    Anyone want to take a bet on it being the tree rings? Assuming there are no upside down sediments that is.

    • Ken
      Posted Dec 5, 2011 at 3:33 PM | Permalink

      I don’t want to be to sarcastic, because I generally agree with almost everything I read here, but tree rings? Really?

      Are these the magical trees that grow in the arctic and migrate south in winter? On the other hand I am sure if the team could convince people of global warming, convincing them of trees in the arctic would be easy.

      Personally I think these trees are killing the polar bears as well.

  8. Bernie
    Posted Dec 5, 2011 at 2:38 PM | Permalink

    Do you have a sense of how many other long O18 ice cores series are out there that Kinnard did not to include?

    Steve: it looks very comprehensive other than Thompson’s Bona Churchill. The coauthors are ice specialists. It’s amazing that so many series have been unavailable.

  9. Pat Frank
    Posted Dec 5, 2011 at 3:12 PM | Permalink

    [Mt. Logan] had an unexpected decrease in O18 values in the modern period, attributed by the authors to changes in regional circulation.

    Something I don’t understand. If, e.g., Mt. Logan produced an unexpected change attributed to regional circulation, why should that not imply other changes due to shifts in “regional circulation,” at Mt. Logan and elsewhere, even if the changes they’ve produced are ‘expected‘?

    [Steve – I agree and commented similarly in my posts on Mt Logan at the time.

    So, the question is: is there an implicit assumption in O-18 proxy studies that ‘circulation changes‘ here are compensated by ‘anti-circulation changes‘ over there, so that the combination of multiple proxies averages out all such non-thermal effects in the final composite?

    That assumption, or something like it, certainly seems to be implicit in dendro-proxy studies.

    This whole business reminds me of those road-runner cartoons, where Wile E. Coyote would run over the cliff and straight out into the air, but not fall to the canyon floor until he came to a stop, felt around, and consciously realized the nothing beneath his feet.

    All these proxy studies seem to blindly proceed over the cliff of unknown variables and out into the open air, assuming the ground beneath their feet.

    My thanks added to the pile, by the way.

    • Craig Loehle
      Posted Dec 5, 2011 at 5:01 PM | Permalink

      In “climate physics” (vs Wily Coyote physics), you stay in mid-air as long as you refuse to admit there is nothing under your feet (aka never admit you are wrong, never).

  10. 007
    Posted Dec 5, 2011 at 3:42 PM | Permalink

    Are you saying it may just be a canard???

    (I’m so sorry.)

  11. Hu McCulloch
    Posted Dec 5, 2011 at 5:06 PM | Permalink

    A PDF is online for free at .

    Lonnie Thompson is a co-author even though Bona Churchill is not included.

    • Hu McCulloch
      Posted Dec 7, 2011 at 3:24 PM | Permalink

      Correction — it’s only free if you logon from a subscribing university computer. The SI and data are there for all to see, however.

      • RuhRoh
        Posted Dec 7, 2011 at 6:05 PM | Permalink

        Calling FergAl or similarly skilled folks;

        Please consider dissecting Figure S11.
        There are quite a few ‘interesting’ graphs where the big black line (preferred answer) is covering up the colored line behind it which might be in conflict with the preferred answer.

        It can be found for free at the supplementary Info;

        Click to access nature10581-s1.pdf


        • RuhRoh
          Posted Dec 7, 2011 at 10:34 PM | Permalink

          Well, I spent some time poking at that figure S11 from the SI.

          It is plots with various proxies withheld.

          From eyeball analysis, the only time that the plots are perfectly overlaid is in that sharp negative deflection at the end.
          This seems to true for all 7 plots. Otherwise, you can see the colored line peeking out from behind the black line.
          This is done at 600% zoom.

          The graphs seem to be a .png file which is not a vector format I think. (?)
          How is it possible that no withheld proxy class causes divergance from the aggregate? Seems rather fishy to me…

  12. Johan
    Posted Dec 5, 2011 at 5:27 PM | Permalink

    Hm, eyeball statistics tells me Devon definitely looks like a hockeystick. Wrong angle of the stick relative to the icy surface, that’s all.

  13. Steve Garcia
    Posted Dec 5, 2011 at 5:42 PM | Permalink

    Steve, I agree with thomswfuller, Tim T and Skeptic above. Thanks for pursuing your lines of inquiry about their lines of inquiry. You’d have made a great Pinkerton man (not Mann…lol).

    In your role as educator (intentional or not) to us all, you bring up basic questions in my head:

    “Eastern” Islands…(I don’t know whether this is due to glacier flow – as a result of which many temperate glaciers e.g. Chile have only quite recent values – or due to reformation of the glacier in the LIA or both.)

    I have worked with laminar and turbulent flows a bit. I am troubled in hearing that they are taking ice cores from glaciers, because glaciers have a huge amount of internal shear going on, and I cannot imagine that this does not taint the hell out of those ice cores. One layer in such a core may have started miles and miles from its neighbor layers (up or down). Bot only that, but nearer the base of the glacier flow the friction with the underlying rock/soil will retard the layer flows there in something like a parabolic curve of unknown parameters), so the “older” layers will be much more chaotic. With all of the shear going on, the gases trapped will be spread out thinly over the length of the layer-to-layer shearing – and probably collect in pockets large and small. This all seems to argue for huge uncertainties, due to the collective chaotic conditions.

    Steve, is this what is really done? Why would they take readings in the middle of a flowing river of ice instead of in stable ice nearby?

    • Paul Dennis
      Posted Dec 5, 2011 at 6:09 PM | Permalink

      Steve Garcia,

      I can’t comment on the details of these studies archived by Kinnard et al. until I’ve looked at the metadata of core locations etc. Many ice core sites, however, are chosen to be at the summit and close to the accumulation areas in order to avoid as far as possible effects associated with deformation due to ice flow.

      • Steve McIntyre
        Posted Dec 5, 2011 at 7:04 PM | Permalink

        Steve G, Paul is right that specialists try to take cores at the summit. Sometimes they miss e.g. one of the Dunde cores was dated back only to the early 20th century.

        My point is a little different. Even at a summit, the flow in temperate glaciers is very rapid and, as a result, temperate glaciers typically yield dateable core for only a few hundred years. The layers in tropical cores thin negative exponentially (at least). Thompson estimates dates for Dunde, Dasuopu etc by estimating the thinning parameter. Any error in this parameter can yield huge errors as one goes back.

        I am unconvinced that the Kilimanjaro glacier is nearly as old as Thompson’s estimate. I made this comment in the AR4 review – it was one of the few comments that they paid attention to. They removed the reliance on Thompson’s date though no questionoing of Thompson’s dating has ever appeared in the litchurchur/

        • Steve Garcia
          Posted Dec 5, 2011 at 7:29 PM | Permalink

          Thanks, Paul and Steve M.

          Steve, I would agree that the flows are rapid, even at the summits. Summits are as often as not just about the steepest parts of the glaciers, as it is bare rock more or less and with little to no till laid down. It is bare at least partly because the glacier scours it.

          Steve, at, referenced by Speed below, you say, “If the cartoon in Figure 1 above is an accurate representation of mountain glacier flow, I find it hard to picture how the annual layering merely thins, but since there is observable annual layering, it must deform somehow so that the layering is kept intact.”

          (FYI, Figure 1 doesn’t display anymore.) Until the Reynold’s number is reached for a fluid, the flow is laminar, and up till then a fluid can keep its layering incredibly well. We used to inject 5 layers of plastic at about 5,000 psi and starting out at about .250″ at the sprue, then filling out for up to about 6″ long a wall that the middle layer thinned reliably to about .003″ thick, and going around some radiused corners. That layer began as about .018″ thick. This all happened in less than a second, yet since the flow rate was below the Reynold’s number, no mixing/turbulence occurred. Some of this kind of phenomena doesn’t seem to be intuitively probable, but they DO happen, if parameters are controlled well enough. It comes down to the Reynold’s number, which has to do with several factors, including the viscosity at temperature, the consistency, the pressure, and the flow rate. But it is also based on a consistent shape, which in upper mountain valleys is all over the place. As the valley shape changes in hard-to-define ways, I would not want to be the one trying to determine if the flow is laminar or not. If they make assumptions, I’d like to see what those are – I’d be at least a tough sell.

          Steve, as to Kilimanjaro, I would tend to agree with you, from things I’ve read in the past about the ages of glaciers. The ages as based on receding rate histories seem to imply much shorter lives than we are led to believe. Basically, I think it appears to be a much more complicated picture than the work done so far would seem to assume.

        • DocMartyn
          Posted Dec 5, 2011 at 7:51 PM | Permalink

          I find this rather surprising.
          X-Ray/Gamma Ray spectroscopy of the ice would, non-invasively, identify metal peaks.
          There should be a pattern of every large impact (iridium/Platinum & Fe/Cu/Ni and every large volcano (Se for sure and Aluminum), recorded as a metal bands.
          Strontium 90 would give you the 1945 date.

          Steve; Isoptopes from atomic tests are used to date the top parts of glaciers, as are tephra from volcanos when available. Dating the first few hundred years isnt the issue. The Himalayan glaciers tend to be about 140 meters thick, with much of their time history placed in the basal few meters which are dated by fitting a thinning parameter. While this may seem surprising to you, it is nonetheless the case. Small percentage errors in estimating the thinning parameter can lead to centuries of dating error at the base (but not at surface.)

        • dearieme
          Posted Dec 6, 2011 at 8:58 AM | Permalink

          It is said that you can identify the fall of the Western Roman Empire by the decline of lead in Arctic Ice. That would mean that you could date the ice from the decline in lead. If that date (albeit rough: who knows when the last lead worker in Londinium threw in the towel?) is inconsistent with other dating methods, those others are wrong.

        • Craig Loehle
          Posted Dec 6, 2011 at 10:03 AM | Permalink

          For one of my papers I corresponded with some authors of the most recent Greenland dating of the long cores there. I asked about dating error and the author was quite offended that I would imply they did not have it nailed. But at least he answered my email and sent me the data.

          Steve: my comments were directed at temperate and tropical glaciers where the situation is much more extreme. These glaciers are relatively thin and receive relatively a “lot” of precipitation relative to their thickness and the turnover is rapid. While there may be somewhat corresponding issues with Greenland glaciers, they will be much diminished.

    • Speed
      Posted Dec 5, 2011 at 6:38 PM | Permalink

      Here’s an interesting graphic from lecture notes of Jed Schneider and Kathryn Clapp here showing the flow of ice through a mountain glacier. If this is representative of flow in mountain glaciers, one wonders at what exactly is being recorded in a vertical drill hole away from the summit (such as Guliya Core 2).

      The comments are interesting.

  14. Steve Garcia
    Posted Dec 5, 2011 at 5:43 PM | Permalink

    I wish Woredpress had a function – terrible typoing in that comment. Sorry! Especially “older” should have been “oldest”.

  15. tty
    Posted Dec 5, 2011 at 7:01 PM | Permalink

    Does anyone understand how the WC01 and POL03 datasets were merged in Fig 1 of the supplementary information?

    The POL03 dataset shows essentially unchanged ice extent in the Siberian sector of the Arctic throughout the 1900-2000 period, while WC01 shows a precipitous decline for the whole Arctic post c. 1970, but the merged dataset is essentially identical with WC01 after c. 1975. They mention “re-scaling” POL03, but one woud still think it should have at least some effect on the final curve.

    By the way the link to the POL03 dataset in the supplementary information is wrong, there is a blank inserted before “index” that shouldn’t be there. Clearly none of the reviewers bothered to check the link.

  16. Posted Dec 5, 2011 at 10:00 PM | Permalink

    Re: There is nothing in the long O18 series that yields the pronounced Kinnard hockey-stick.”

    If two possible proxies yield totally different results, at least one of them must be invalid. So why would one combine the two?

    And why would anyone use tree ring growth as a proxy for the effect of rising atmospheric carbon dioxide concentration on temperature or sea ice melting when we know from numerous controlled studies that carbon dioxide stimulates tree growth independently of any effect it has on temperature?

  17. BobN
    Posted Dec 5, 2011 at 10:02 PM | Permalink

    Steve – Are these ice core proxies all from glaciers and/or ice sheets surrounding the Arctic Ocean? And if so, how are they consider proxies for sea ice extent? Maybe I am failing to make the obvious connnection.

  18. Kevin O'Neill
    Posted Dec 5, 2011 at 10:12 PM | Permalink

    Steve — what justification is there for only including the long series in your ‘Compare O18 to Reconstruction’ graph? Ignoring almost half the O18 proxies seems rather dubious.

    What does the ‘O18 vs Reconstruction’ look like if all the O18 proxies are included?

    if you’re interested in comparing the MWP to the modern period, then the only relevant proxies are the ones that cover both periods. It wasn’t that I was ignoring the shorter proxies. (I did show 4 shorter proxies from the Eastern Islands.) But combining the short proxies with the longer proxies raises methodological issues that do not arise by using proxies that cover both periods. Plus if you get a different answer when you include shorter proxies, then the different answer would seem to me to be an artifact rather than the “right’ answer.

    The splicing methodology used by Kinnard looks rather suspect to me.

  19. William Larson
    Posted Dec 5, 2011 at 11:11 PM | Permalink

    The concern I have, trivial or not, is why, in some of these series, the archived data do not extend up to the date of the coring. I am troubled by this apparent “truncation” of these series. I have taught myself here at CA to be wary of any (apparently) missing data.

  20. Posted Dec 6, 2011 at 11:59 AM | Permalink

    I listened to presentations and participated in discussions with Gerry Holdsworth, the person involved very early with the Mt Logan cores, and Fritz Koerner the person involved in the early drilling on Ellesmere Island. We were part of a group involved in the National Museum of Canada’s project “Critical Periods in the Quaternary Climate History of Northern North America”, which was part of the National Museum of Natural Sciences Project on Climatic Change in Canada During the Past 20,000 years. This ground breaking project was canceled when Environment Canada withdrew funding to pursue global warming.

    Holdsworth talked about the extreme difficulties in getting access to the glaciers, the shallowness of the glacier, the severe wind conditions and scouring of layers of snow and the large temperature range plus the problems of plastic flow and turbidity. There is also the problem of summer melt and access of the water at the top of the glacier through the Bergschrund. As I recall he considered there were few ‘good sites’ for drilling.

    Fritz Koerner, who was also a key participant in the Canadian Geological Survey Polar Shelf Project, gave a few talks on the problems of ice coring and analysis. I recall him telling us that scouring and melt were also a problem especially because of warm air created by polynas off shore and constantly shifting sea ice. I also recall him warning about CO2/temperature because his preliminary findings were that temperature changed before CO2 as later emerged from the Antarctic data, although I am not aware he ever published this.

    The bottom line for both people, especially Koerner, was one of extreme caution and to produce the work but make people very aware of the severe limitations. The critical issues mostly overlooked because few warn of the limitations include; that precipitation, both snow and rain, are as important in glacier construct and dynamics as temperature (it’s the tree ring issue again); the amount of water in and around a glacier – even at high altitude or latitude, this meltwater includes supra glacial, englacial and subglacial; the amount of melt of the surface layers varies with the changing albedo caused by the amount of entrained debris; the importance of wind speed and direction which are much different at altitude and over time and include dramatic katabatic flow – the annual mean wind speed at Cape Dennison, Antarctica is 19.3 m/s (43 mph, 70 kph). Of course, these wind speeds and directions change as the glacier dynamics and global conditions change. Consider the global wind patterns and speeds at the glacial maximum just 20,000 years ago. How different were the dynamics even during the Little Ice Age, problem Holdsworth confronted on the short record on Mt Logan.

  21. Hu McCulloch
    Posted Dec 6, 2011 at 1:07 PM | Permalink

    This study uses two series from Svalbard’s Lomonosovfonna ice field, the subject of my 2009 post, “Svalbard’s Lost Decades,” at . Readers may want to review that post plus the exensive and very helpful discussion by author Aslak Grinsted and others.

    The new Data Supplement 1 contains raw data for “MSA” (#17) back to 1126 AD, which must be one of the washout ratios that indicated very warm summers up to 1200, along with d18O back to 1440 (#53). I haven’t had a chance to examine them closely yet, and I can’t find a definition of MSA in either the paper or the SI. There were in fact two washout series, Na-Mg which showed extreme warming, and Cl-K which showed warming only comparable to the current century.

    One problem back in 2009 was that Elisabeth Isaksson, who controlled the data and is a co-author on the Kinnard paper, had never archived the full data. At last, at least “MSA” has been archived for the full core back to 1126. However, d18O is still only reported back to 1406. Back in 2009 Grinsted indicated that Isaksson was withholding the earlier portion of the d18O series for a later publication, and referred the issue to her.

    • Hu McCulloch
      Posted Dec 6, 2011 at 2:16 PM | Permalink

      On further reading MSA is methanesulfonic acid (see ref # 61 in the SI). So while this is undoubtedly interesting, it is not the washout ratios discussed in the Grinsted article that indicate very warm summers in the 12c.

    • Hu McCulloch
      Posted Dec 6, 2011 at 7:44 PM | Permalink

      Kinnard archived 4 O18 series from “eastern islands” – two from Svalbard (17-24E), one from Franz Josef Land (64E) and one from Severnaya Zemlya (95E) -none of which have appeared in previous studies.

      “lom” in your “Eastern Islands” d18O chart looks the same as Lomonosovfonna in Grinsted (2006) and later articles, so it has appeared already. See . However, this may be the first time it has been archived.

      Nevertheless it is apparently missing its early portion going back to 1130 or so, which coauthor Elisabeth Isaksson has evidently been withholding.

  22. alex verlinden
    Posted Dec 6, 2011 at 1:30 PM | Permalink

    I’m in complete and utter awe …

    there’s lots of data, and there’s 1 result from those data …

    all the data are in black, the result is in red …

    nowhere in black, I can spot an “unprecedented” rise of the data in the 21st century, and still, the 21st century result of that all goes off the screen to the upside …

    simply amazing … absolutely great job …

    • Hu McCulloch
      Posted Dec 6, 2011 at 2:19 PM | Permalink

      Alex, Steve’s graph above merely compares the d18O records (and then only the longer ones) to the final reconstruction. There is lots of other data, so the HS must be coming from this other data that he has not examined yet. I suspect that if you stay tuned you will learn more.

      • Tom Gray
        Posted Dec 6, 2011 at 5:01 PM | Permalink

        If there is a common temperature signal through the proxy measurements, shouldn’t there be some measure of similarity between them. This is a real question from a layman. If a group of proxies do not exhibit a correlation with temperature in the 20th century and other groups do, is that not inidicative of something?

    • Steve McIntyre
      Posted Dec 6, 2011 at 2:53 PM | Permalink

      there are other data series that are not canvassed in this first post. However, I don’t see anything in the Na+, CL- or MSA series that would result in such a big HS.

      • RomanM
        Posted Dec 6, 2011 at 5:34 PM | Permalink

        Steve, the hockey stick will not show up in the data unless you know where to look for it. 😉

        The authors take the 0-1 scaled (infilled and presumably smoothed) data where it is the most complete (from 1843 to 1995) and extract the PCS. The first PC does not look like a hockey stick (Fig.S3 in the Supplementary pdf)and is identified as the “temperature signal”. However, the second PC (which I may point out is uncorrelated with the first PC) can be oriented to look like the “August ice extent” so is deemed the correct item to use in making the reconstruction.

        The ice extent seems to be a construct cobbled together from three sources: the satellite record, a Russian dataset on ice thickness (1936 – present) and ice extent (1900 – present) at Arctic locations near Russia and finally, a third set which purports to go back to 1870 but for which the data website states:

        Please note that large portions of the pre-1953, and almost all of the pre-1900 data is either climatology or interpolated data and the user is cautioned to use this data with care (see “Expert user guidance”, below).

        There are three periods for which the sources of the data change fundamentally:

        1972-1998: Satellite period – hemispheric coverage, state-of-the-art data accuracy
        1953-1971: Hemispheric observations – complete coverage from a variety of sources. The observational reliability varies with each source, but is generally accurate.
        1870-1952: Climatology with increasing amounts of observed data throughout the period.

        Seems like a bit of a stretch…

        • Hu McCulloch
          Posted Dec 6, 2011 at 9:28 PM | Permalink

          “Because of its strong persistence and relatively short length (n = 126 years), the observed sea ice record has a low degree of freedom (about 9), which impedes discarding long segments for validation, as done elsewhere.” (p. 513)

          If there are only 9 effective DOF (in the Bertlett-Quenouille sense after adjusting for serial correlation, presumably), it must be hard to meaningfully estimate more than 9 parameters. 😉

          And if the early sea ice data is not very reliable…

  23. Hu McCulloch
    Posted Dec 7, 2011 at 4:03 PM | Permalink

    Since proxies and sea ice are both noisy reflections of exogenous temperature, neither way of running the regression is pristinely correct.

    Kinnard et al regress sea ice on proxy Principal Components, which will tend to understate the slope of the transfer function from proxies to sea ice. But multivariate CCE (or just constructing a proxy index by regressing sea ice on multiple proxies and then regressing the proxy index on sea ice and inverting) would have the opposite bias.

    If the relative size of the two types of noise could be quantified, this could be an appropriate application for Total LS (pace Roman 🙂 ). This data by itself cannot tell us the relative size of the two noises. However, regressing both sea ice and proxies on temperature during the calibration period might give one an adequate handle on this ratio.

    • Steve McIntyre
      Posted Dec 7, 2011 at 4:10 PM | Permalink

      Hu, one thing that seems vital to me in any regression-based methodology is for the authors to clearly show the weights that arise in the reconstruction – the weights are the “dual” of the reconstruction in a linear algebra sense.

      At the end of the day, the reconstruction is a linear combination of the underlying proxies. So we should be shown both the sign and the weight of the coefficients.

      ONe of the advantages of regional averaging is that it keeps all the mechanics in plain view.

  24. Hu McCulloch
    Posted Dec 9, 2011 at 10:47 AM | Permalink

    As noted in the preview, this means that the derivation of the Kinnard hockey stick must come from elsewhere: either from the other ice core proxies (Na+, melt, MSA), from the D’Arrigo tree rings or from the other proxies (a few sediments and documentary series) or from regression procedures. If so, one needs to check out which series have been heavily weighted and whether series have been flipped through the regression methods.

    In Table S1 of Kinnard’s SI, the 22 d18O proxies are evenly split (11/11) between positive and negative correlations with their August sea ice index. In the SI text they rationalize either sign, so that they are willing to live with “flipped” signs.

    In fact, almost all their proxies show nearly equal splits, which they also rationalize in the SI text:
    Ice core melt (#5-#12) is 4 positive vs 4 negative.
    Ice Core MSA (#13-17) is 3 plus, 2 minus.
    Ice Core Na+ (#18-32) is 10 plus, 5 minus.
    Tree Ring Width (#59-69) is 6 plus, 5 minus.

    The most notable exception is Lake Varve Thickness (#55-58), which is 0 plus, 4 minus. However, in theory lake varves should go either direction with temperature (and therefore sea ice?), depending on whether the lake is glacier fed or show-fed. They give no indication of lake feed type, so it’s not clear whether these are as all as expected, all flipped, or some of each.

    (A glacier-fed lake will have heavy inflow in warm years, and therefore thick varves. But a snow-fed lake will have the most turbulent inflows and therefore thickest varves in cold years in which there is lots of snow relative to rain and therefore a big spring thaw. But then an open Arctic Ocean would cause more precipitation in general and therefore thicker varves.)

    Ice Core Cl- (#3 and 4) are both positive (2 plus, 0 minus). This is backwards, since less sea ice should mean more salt spray and more Cl-, but they are able to rationalize even this sign.

    Barents Sea ice limit latitude (#1, back to 1581) is naturally negatively and strongly correlated with sea ice extent.

    The Icelandic Koch Index of sea ice severity (#2) goes back to 1100 (!) and comes in weakly negative at -.09. I’m not sure how it’s measured or what its sign should be, but it doesn’t look significant either way.

    The last column of Table S1 also gives “maximum” correlations with local gridded ice conditions (just for the satellite period?). These also come in mixed about 50-50 for most proxies, and even lake varves are now mixed 1-3. These correlations are oddly bimodal, with not a single one less than 0.30 in absolute value. However, they apparently represent the maximum absolute correlation over several local gridcells, which could explain the bimodality.

    So with perhaps only 2 exceptions (#1 and 2), the authors are apparently willing to live with either sign on each proxy. This doesn’t exactly instill confidence in the value of the proxies.

    • Craig Loehle
      Posted Dec 9, 2011 at 11:16 AM | Permalink

      “This doesn’t exactly instill confidence in the value of the proxies.”
      No wonder they got a hockey stick. If you can specify a priori how the proxies should go under different circumstances, fine (maybe) but after peeking at the cards it is the sharpshooter’s fallacy.

    • RuhRoh
      Posted Dec 9, 2011 at 2:55 PM | Permalink

      After close (800x zoom) inspection of Figure S11 in the SI
      (Reconstructions with Proxy Classes Withheld),

      it seems that only graph Sll(f) (Lake Varves Withheld) has a full colored pixel at the tip of the hockey stick,
      but perhaps this is image compression artifact.

      The data graphs seem to be a .png image onto which the text is applied by XML, which limits my ability to zoom and interpret. I don’t know how to recover from the (presumably lossy) compression. Maybe FergalR has a trick for this also, but not me.

      I had noticed this little teaser prior to your ruthless dissection of bipolar weightings in each class but the varves.

      I guess it is not inconsistent with the varves having a ‘unipolar’ role in the HS.


      • RuhRoh
        Posted Dec 9, 2011 at 3:10 PM | Permalink

        I still can’t understand why all 7 the S11 graphs have such perfect matching on the blade portion of the hockey stick? How is that possible?

        Was there some kind of ‘calibration’ to that time interval?

        Figure S11 seems to imply that classes of proxies were ‘fitted’ on a class by class basis, as if to ensure that exclusion-by-class ‘verification’ would look good. (?)
        This speculation is purely from ‘answer-analysis’…

    • Steve McIntyre
      Posted Dec 9, 2011 at 3:09 PM | Permalink

      If the proxies are flipped in sign all over the place, it looks like a sort of multiple inverse regression against a very large number of proxies. If so, then the methodology runs the risk of being akin to the example from Phillips 1998 discussed here where a function generated from a sine curve was regressed against multiple white noise series, yielding a “fit” through brute force. Kinnard has all the signs of being something like that.

  25. Kevin O'Neill
    Posted Dec 9, 2011 at 9:03 PM | Permalink

    Devon, Prince-of-Wales, Dye 3, NGRIP, GISP2, and B18 all have characteristics of hockey sticks – despite the large variations all end with visually apparent increases.

    I don’t know the areal or latitudinal weighting applied, but from the long ice-cores alone it’s obvious there was a significant increase in temperature the last couple hundred years.

    Why that would surprise anyone is beyond me. How many different lines of evidence do we need before we realize something rather unusual has been going on?

    • Steve McIntyre
      Posted Dec 9, 2011 at 10:59 PM | Permalink

      Re: Kevin O’Neill (Dec 9 21:03),

      a couple of points. 1) the Kinnard HS reconstruction (not shown in this post – see the article) has a HUGE blade reaching to unprecedented levels. From a statistical point of view, since the reconstruction is a linear combination of proxies, it has to come from individual proxies or from overfitting to the target through their regression methodology. Though the 20th century O18 values of the long O18 series are higher than 19th century values, they are not unusual relative to MWP values and thus not classic HS series. Nor can these series yield the Kinnard-shaped HS through regional averaging.

      Understanding the properties of the proxies and methodology is a legitimate question to ask in connection with this type of study.

      • Kevin O'Neill
        Posted Dec 10, 2011 at 6:57 PM | Permalink

        Steve says – “the reconstruction is a linear combination of proxies”

        While the article is paywalled, the supplementary materials are not. Nature

        The supplementary materials indicate EOF, PCS, and PCR analysis was performed. Wouldn’t the reconstruction then be a linear combination of the resulting Empirical Orthogonal Functions – not a linear combination of the proxy data itself?

        The last 140 years of data isn’t dependent on proxies – we have direct observations. Doesn’t this reconstruction (say from 500AD to 1870AD) mesh with all the other data we have? I.e., foraminfera from arctic sea bed cores, bowhead whale fossils, plankton studies, etc.

        What would be surprising is if it contradicted previous studies, rather than confirming them.

        • RomanM
          Posted Dec 10, 2011 at 7:40 PM | Permalink

          For the record, EOFs are a linear combination of the proxies, so what Steve said is quite correct.

          With regard to “all the other data we have”, reconstructions are not compared to “data” but to other reconstructions which can often utilize much of the same data and similar methodology so the comparisons are not quite independent. The quantification of the “meshing” is not always specified with the evaluation being a simple offhand remark and differences between the reconstructions rarely addressed in a scientific manner. Reconstructions which do not share the same features are often discounted as being “incorrect” without sufficient justification.

          It’s not a simple case of just looking at the “observations”.

  26. Alexander Harvey
    Posted Dec 9, 2011 at 9:53 PM | Permalink



    I don’t know if you are interested but some general questions regarding the inversion can be addressed, e.g. temporal resolution. This is important at least as far as knowing how safe the end sections of the reconstruction are. AFAICR temporal resolution, is in practice unlikley to be better than a band [0.6t, 1.6t] where t is time into the past.

    The upshot being that if shallowest depth and deepest depth correspond to the interval [t1,t2] only [1.6t1,0.6t2] is safe.

    The practice of predetermining the undisturbed gradient, as used in the 500 yr recon, is inherently unsafe. Strictly speaking that gradient should be an ouptut of the analysis not an input to it. Or put another way, a reconstruction where it is predetermined should match the result when it isn’t. I found this not to be the case.

    Lastly it was apparent to me that some procedure is necessary to screen each whole for quality. The most likely issue being a discontinuity in the coefficient of thermal conductivity. This wouldn’t matter so much were it not for the constant thermal flux up through the crust. A small change in conductivity puts a nice V in the temperature profile for a single discontinuity.

    I think people do realise how problematic these reconstructions are but I suspect not quite how problematic they are.

    I gave up on it some time back for I could not find any reliable criteria for screening boreholes or retaining components that didn’t leave me with none of either.

    If you are interested please comment, the email address is not one I regularly monitor and is terminally over spammed.


    • Steve McIntyre
      Posted Dec 9, 2011 at 11:06 PM | Permalink

      Re: Alexander Harvey (Dec 9 21:53),

      thanks for the comment. I enjoy your comments at Keith Kloor’s by the way.

      If you have some technical commentary, I’d be happy to post it separately as this topic is a bit of a loose end for me. I experimented with differing numbers of retained principal components in inversion (a battleground issue in connection with a number of other paleo topics) and thought that the comparison was interesting – though I haven’t written these notes up.

      Quite separately from the inversion issues, as I’ve mentioned from time to time, some of the classic “boreholes” are in mining areas where I have some experience. Gold mining areas are in “shear zones” and there are fractures (with water) everywhere.

      I did some analyses looking at drill holes near important Canadian mines – again that I didn’t write up, but were interesting.

    • oneuniverse
      Posted Dec 13, 2011 at 10:28 AM | Permalink

      re: T reconstruction by inversion of borehole T profile

      Given that there are many possible temperature histories consistent with a given borehole T profile, might it be possible to address the problem to some extent by re-measuring the T profile of the borehole (or perhaps preferably, of another freshly drilled borehole nearby) some time later (say 50 years later) ?

      Applying the inversion procedure to the two profiles yields two sets of possible temperature histories. Ignoring the last century (say), one can then exclude solutions not in the (approximate*) intersection of the two sets to obtain a smaller set of possible solutions.

      *: approximate intersection : x member of A and y member of B are in the approximate intersection of A and B if the difference between x and y, for some appropriate metric, is appropriately small 🙂

  27. Posted Dec 10, 2011 at 7:42 PM | Permalink

    I have read the post and every comment on the this interesting thread. I don’t have much to add other than – don’t ever use non-ice boreholes as thermometers -ever. I don’t have an opinion about ice boreholes other than the fact that the weight differential of isotopes combined with variations in weather patterns leaves me appropriately skeptical.

    Steve, way off topic so clip if you want but Chapter 4 zero order draft is available here:

  28. Posted Dec 12, 2011 at 1:20 PM | Permalink

    Perhaps Steve will have a new thread soon on Kinnard’s Partial Least Squares (PLS) method, in which case this comment may be premature.

    I’ve been reading the 2001 article in Chemometrics and Intelligent Laboratory Systems (58:109-130) that Kinnard cites, and Kinndard’s summary of the method on p. 513 of his article seems to follow Wold et al.

    PLS does cherry-pick to some extent, by effectively flipping all signs as required to get a positive correlation, then forming an average of the predicted predictand across all predictors in the first stage, and then repeating this a few times on predictor and predictand residuals from each step.

    PLS not quite as cherry-pickish as it sounds, however, since the “weights” it uses just ammount to univariate regression coefficients, and then it takes an unweighted average of these preidctions to form a proxy summary at each stage. A further weighting of each proxy’s prediction by its simple regression R2 would give results similar to forward Recursive Least Squares, but this is not done.

    In order to evaluate the fit or “skill” of the model, it is necessary to do some sort of Monte Carlo tabulation of critical values for R2 (or Wold et al’s Cross
    Validation Q2 statistic). I personally don’t like Fritts’s RE or CE statistics beloved by dendroclimatologists, and have more confidence in R2 (or Q2), but maybe that’s just my lack of understanding.

    On p. 15 of the SI, Kinnard et al report that they do something along these lines, at least for the RE statistic:

    Monte Carlo simulations were used to test the statistical significance of the reconstruction. The original proxies selected as predictors in each PLS model were replaced by nonsense predictors and the calibration and validation procedure repeated 300 times. The nonsense predictors were derived by randomizing the Fourier phases of the proxies, which resulted in surrogate proxies having equal mean, variance and power spectrum, but randomized phases, effectively destroying any phase relationships between the proxies and th eice extend predictand. The 95% and 99% significance thresholds on the RE statistic were derived from these Monte Carlo tests (Fig. S9).

    Aside from the choice of RE over R2 or Q2, it would seem a lot more natural to me to hold the predictors (the proxies) fixed, and then to generate random predictand series (sea ice) under the null that the predictors have zero explanatory value. For this purpose it is merely necessary to model the predictand under the null, in terms of its autoregressive structure. What is relevant here is therefore not the behavior of the prediction residuals, but the behavior of the predictand itself without reference to the prodictors. Then with each simulated predictand series, repeat their PLS procedure and compute the selected test statistic.

    Steve has already discussed PLS some at
    Unfortunately, some symbols (notably u-umlaut in Burger’s name plus Burger’s coauthor’s name) got garbled in the conversion to WordPress, but it’s a start.

    • RuhRoh
      Posted Dec 12, 2011 at 2:51 PM | Permalink

      Do you have the impression that the Kinnard keeps the proxy classes grouped together, at least initially?

      I am stuck with the idea that the Kinnard ‘method’ was (at least partially) chosen to ensure the compelling ‘classes-withheld’ graphs in Figure S11 of the SI. Each withheld-class reconstruction seems to reproduce the hockey stick blade with very high fidelity, and misses elsewhere.

      That seems highly improbable to me, that withholding each proxy class would have a negligible impact on the blade, unless they were initially weighted within class to ensure ‘separability’…

      • Hu McCulloch
        Posted Dec 12, 2011 at 6:29 PM | Permalink

        It’s not my impression that Kinnard et al grouped the classes in any way, except for the “delete one class” fig S11.

        It could be that there is something in the PLS methodology that generates Hockey Sticks. Or perhaps sea ice was really flat before 1900, and this is reflected in most of the proxies.

    • Hu McCulloch
      Posted Dec 13, 2011 at 12:44 PM | Permalink

      PLS does cherry-pick to some extent, by effectively flipping all signs as required to get a positive correlation, then forming an average of the predicted predictand across all predictors in the first stage, and then repeating this a few times on predictor and predictand residuals from each step.

      PLS not quite as cherry-pickish as it sounds, however, since the “weights” it uses just ammount to univariate regression coefficients, and then it takes an unweighted average of these preidctions to form a proxy summary at each stage. A further weighting of each proxy’s prediction by its simple regression R2 would give results similar to forward Recursive Least Squares, but this is not done.

      I’ll take that back — after flipping signs so that all correlations are in effect positive, PLS will flatten the predictor series that don’t happen to correlate strongly with the predictand, and beef up those that do. Higher order terms will then perfect the shape.

      This procedure could be ferreting out a subtle but valid relationship, or could just be generating spurious results with a flat handle. It would take an appropriate Monte Carlo simulation to determine whether the results are spurious or not.

  29. Hu McCulloch
    Posted Dec 12, 2011 at 2:11 PM | Permalink

    Steve —
    What would your “Comparison to Kinnard HS” look like if each d18O series were flipped before averaging so as to correlate positively with Sea Ice? This sign flipping is part of what PLS does.

One Trackback

  1. By New Light on Svalbard « Climate Audit on Apr 13, 2013 at 10:59 PM

    […] in 2011, I reviewed the comprehensive Arctic O18 data from Kinnard et al 2009 here. I observed that the Kinnard Hockey Stick was not observable in the long O18 data and therefore had […]

%d bloggers like this: