I wasn’t suggesting anything that radical. Willis was proposing that he would replace the proxies in a cluster with a single average for that cluster. That seems sensible otherwise, 10 tress of the same species in the same area have ten times the weight.

My suggestion was, that you could achieve, probably much the same effect, by a different method. Namely first average proxies in a grid cell. It isn’t exactly the same. If Bristlecone or Tiljander proxies happen to not all be in the same grid cell they’ll have a bit more impact. There would be a difference in the other direction, if there were two different clusters in the same geographical location. If you average by clusters that would give two signals. If you average by grid-cell, you’d only have one.

I wasn’t proposing that Willis starts a whole new research project or starts hunting down potential new proxies. I was simply offering an alternative way of attempting to treat the issue that sometimes you have a cluster of 10 similar trees in one place, and other times you only have one tree or 1 lake sediment record. If you are simply averaging the data together then it matters that you have weighted the data equally.

I think, perhaps, I wasn’t explaining myself very well. The issue here is, should 100 similar proxies (say all bristlecones or all juniper trees) really be thought of as 100 times more significant than 1 speleothem. By doing simple averaging you are giving each individual proxy the same weight. If you identify proxies as members of a cluster and then average within each cluster, and then average the clusters together, then you are giving each cluster the same weight. Which strikes me as more reasonable than the first simple average.

But it still calls for a certain degree of judgement. I wasn’t accusing you of cherry picking, I was trying to find a way of of avoindg that being an issue.

My alternative suggestion, is that proxies are identified by lat+long to a given 5° by 5° grid-cell. You are then clustering the proxies together by grid-cell. Take the average for each grid-cell, then average all the grid-cells together. Course, you’ll be doing this with anomalies not actual temperatures. But this is roughly how the instrumental records seem to work. if one cell has 100 thermometers it doesn’t get more weight than the 1 thermometer in the next cell.

That is different from your approach, but probably not by much. If a given proxy family covered a large area then it would get a little bit more weight. Although not by as much as if you use all the individual proxies separately.

The approach I suggest would also reduce the influence of geographically close but unrelated proxies. There might be a location somewhere on the globe, with plenty of long lived trees, interesting caves, deeply sedimented lakes and plentiful branches of Starbucks. But should our estimate of global temperature be more weighted by this geographical region, simply because more different proxies have been produced from that location? The actual number of proxies seems pretty small – when compared to the number of thermometers used to create the instrumental record. So I’m not sure that this can be a major issue. Nevertheless, in principle, it seems wrong that one part of the globe should be over-represented in the global average simply because it has produced more data.

In addition, as I showed in the head plot, if there is a signal, it is revealed by averaging. If we knew more we could likely extract the signal better, but we don’t. For my money, if it can’t be seen in an average, we’re getting into sketchy territory. And in this group of proxies, averaging does reveal a common signal.

You ask why not do it like the instrumental record … we can’t, because we have no common scale. With temperature records, I know that 30°C in Vladivostok is the same as 30°C in Tahiti, and a ten degree drop in either one the same. With proxies, all we have are records which may or may not contain a signal, and which do not have a common scale.

You also talk about cherry picking. I have not done the job yet, but I’ve been thinking about ex ante criteria for averaging.

There are some easy cases. If there are two (or more) samples from the same location, of the same type, by the same investigator, I’d average them. Start from there and work up to the hard cases. I’m doing an investigation, not looking for a final answer yet.

The key to me is the similarity of the cluster, and its location among the other clusters. For instance, the Argentinian Cypress cluster (above) are all from nearby sites, and have a very distinctive shape. They form their own cluster, and it’s the second to last cluster assimilated (Graybill is last, at the top).

Now, I see nothing wrong ex ante with the cluster … but clearly we don’t want to have seven of them balanced against single proxies from other geographical areas.

From what I have seen, I don’t think the details of exactly who is averaged will make a whole lot of difference. Once you pull out the heterogeneous proxies (or alternately the Graybill and Tiljander clusters), the overall shape seems fairly robust. But, I haven’t done the shovel work yet. I will do it a bit at a time, so I can compare the effects of various actions.

Re: BarryW (#150), you ask about trying various cutoffs. I’ve played with it quite a bit. As Figure 1 shows, the majority of the signal is in the ends of the histogram. Where you put a cutoff point matters very little. You see in Fig. 3 above the results of setting the cutoff at 0.005 and 0.995, and they differ very little.

More later, gotta catch a plane …

w.

Reed, Also try here</a

Instead of using &gt 0.005 as the cut off have you tried a higher value (say &gt .5). Could the values below .5 be canceling the effect of those above .5?