Interesting work Kenneth. Thanks for posting.

]]>As a follow up to showing the original Mann 2008 proxy series without in filling and without the proxy groups Luterbacher, Schweingruber and Tiljander (excluded for reasons already discussed in detail above) and a composite plus scale (CPS) of the individual series, I want here to show these 1029 proxies divided into 40 groups with 39 groups formed by proxy type and locality/region with at least 5 proxies per group and a 40th group made up of the remainder of the proxies not fitting into the group of 39. The remainder group was from different types of proxies (187) and dispersed around the globe. As a result the 39 groups should give a picture of local historical temperature if the proxies had a discernible temperature signal while the 40th group should do the same for the historical global temperatures.

In the first 4 links/graphs below I show the series for the 40 groups with each labeled such that the group could be identified from the Mann 2008 SI and the 40th group is identified as Remain. The picture presented by these group series is that the groups as a whole show no discernible temperature signal and particularly in comparison to the earlier shown simulated ARMA and ARFIMA models. It should also be remember that the y axis is in standard deviation units and converting those units to degrees C would reduce the scale.

I calculated a CPS of the 40 proxy groups for comparison with the earlier presented CPS of the individual series and those graphs are presented in the last link/graph below. This comparison should show any discrepancies that might arise from the lack of area weighting in the CPS for individual proxies. I show the upper and lower 95% confidence intervals outlined in red and shaded grey. If I try to show the actual series with the CIs, the graph becomes difficult to read. I have also included a spline smooth the CPS series in both graphs. The two graphs are very similar and are in line with what one would expect from combining proxies that are responding to variables other then temperature in a manner that tends very much toward a noisy series without a signal and particularly a discernible temperature signal. The variability of the CPS series should be evaluated based on the simulations of ARMA and ARFIMA models shown previously.

It should be noted that the CIs for the CPS from individual proxies shown here are much reduced from that shown previously. In the previous graph I inadvertently did not divide by the square root of the number of data points used in determine the mean. Making this error doubly egregious was exclaiming my surprise that CI ranges were so wide.

]]>Eli Rabbet, a bit late in the game here, but I thought I’d link to some results that were driven by your comments about variance.

If I scale the spectra so that the low-frequency aligns with Loehle (I think I have a better method now, due to a suggestion by “A Fan”), here’s what I get:

As I said on Brandon’s blog:

While the reconstructions have been scaled to match the low-frequency portion of Loehle & McCollough, the two temperature series are shown with no scale adjustment. Given uncertainties in the relationship of the pseudo-temperature scale to the global temperature scale, the level of agreement was a bit surprising to me.

Also — note that Moberg does not agree well with the other high-frequency reconstructions. Loehle predictable rolls off steeply below 50-Hz.

Ljungqvist and Mann 2008 EIV appear to agree well with each other (given uncertainty) and with the global temperature series, both in slope and in magnitude.

There seems to be an issue with Moberg. You were right about that, but it seems to be more substantive than just sample size.

]]>Would it not be better if he keeps digging? IMHO

]]>In medical studies, knowing how survivor bias can influence a study affects whether it is taken seriously or not (think what’s different between those who died versus those who didn’t and if that affects the inferences or conclusions made in the study). It would be good to know, if possible, if there are systematic differences between the proxies that survive to the end and those that don’t. Is there enough data to do a t-statistic (difference between the means) or other methods to test for survivor bias?

]]>“What, if any, survivor bias is introduced by that large drop-off in the number of proxies?”

Not sure I understand the question. The drop-off starts at about 1975 and this is on the original and not in-filled data. Remember the authors of Mann 2008 felt compelled to in-fill for every proxy with missing data up to 1995. That action makes a farce of the reconstruction and particularly since the time period it covers is the modern warming period where we want to most how the proxies respond. As the data becomes more sparse at the end the reconstruction the more influence a few proxies can have.

]]>Kenneth,

What, if any, survivor bias is introduced by that large drop-off in the number of proxies?

]]>I did a composite plus scale (CPS) by the methods described in the post above on the Mann 2008 temperature reconstruction (before in-filling) after excluding the 71 Luterbacher proxies (instrumental), the 4 Tiljander proxies (not a temperature proxy for the modern era) and the 105 Scweingruber MXD proxies (truncated at 1960 in Mann 2008 due to divergence). I show the composite for 1400-1995 with a spline smooth and 9 simulations using a well fitting ARIFMA model in the first link below.

In the second link below I show the 2 sigma confidence intervals (CIs)for the composite and the number of proxy data used each year (1400-1995) in calculating the composite. I was shocked by the very large range of the CIs compared to the mean composite series and the spline smooth. I have not done CIs for CPS, but I could not find any basic error in the method used. The method was merely calculating the row standard deviations from the standardized proxy matrix where the columns were the individual proxy data standardized in units of standard deviations.

The resulting CIs would be expected from individual proxies which gave very random responses each year around a mean close to 0. I want to do some more simulations in the near future, but I cannot see where there is any discernible temperature signal in this group of proxies. I judge that it takes numerous manipulations of the proxy data (tricks) used in Mann 2008 to make the data appear to have a signal.

]]>Layman, I was contemplating excluding all 105 Schweingruber proxies, but I wanted to make the point of the truncation at 1960. I can easily do a CPS without those proxies. Actually your analysis, as I recall, found the divergence could have started in the 1920s. That is a critical point because that would tend to counter the arguments made by those who go to great lengths to connect the divergence to something in the modern warming period and thus claim that the cause of divergence was less likely to occur in the past and is somehow unique to the warming period.

]]>