Yesterday I left my exposition of Mannian CPS just before trying to do a NH composite. There was another stupid pet trick left, that made replication of this step a little trickier than one would think.
UC had sent me a file entitled NHmean.txt (which I’ve placed online), which was produced during gridboxcps.m, and proved to be 9 unspliced of the NH reconstruction using iHAD as a “target”.
In this case, I was able to locate information at Mann’s website that could be reconciled exactly to this output (after performing a little trick.) The third column in the dataset at http://www.meteo.psu.edu/~mann/supplements/MultiproxyMeans07/data/reconstructions/cps/NH_had.csv (the column entited “cps.iHAD.”) started in AD1500 and was a splice of three columns of our NHmean.text – the first 100 years was from the AD1500 step, the next 100 years from the AD1600 step and the remainder from the AD1700 step.
The AD1800 step was left out. How? It doesn’t seem possible from the source code. But who knows? It is what it is. Another stupid pet trick.
The second reconstruction in this spreadsheet (headed “cps.HAD”) has a correlation of exactly 1 to the cps.iHAD reconstruction. It starts from the same composite (which I’ll discuss below) and is rescaled a little differently – to match the HAD instrumental series (instead of the iHAD series). It’s 1850-995 standard deviation and mean are a little different, but otherwise the series is identical. One version supposedly reconstructs land temperatures, while the other reconstructs land+ocean temperatures: the size of the glove changes depending on the person wearing it.
Related (spliced) reconstructions using CRU (instead of hadCRU) as a target are at http://www.meteo.psu.edu/~mann/supplements/MultiproxyMeans07/data/reconstructions/cps/NH_land.csv . The column headed “cps.iCRU” has a correlation of exactly 1 to the two HAD series, but is scaled a little differently. It starts in AD600 instead of AD1500, but like the two versions already discussed, omits the AD1800 step.
The column headed “cps.CRU” starts in AD200 and, unlike the other three series, had a high ( more than 0.9 correlation to the other series) but differed for inexplicable reasons – something that shouldn’t happen given the source code.
Another weird and stupid pet trick lay here. The post-AD1700 portion had an exact correlation of 1 to the other versions – so the first leg in the splice was the AD1700 portion. For an inexplicable reason, Mann skipped the next two steps in this version – the AD1400-AD1700 portion had an exact correlation to the reconstruction step beginnning in AD1400 – so the AD1400 network was used for 300 years.
At this time, I only have Matlab output going back to AD1000 and the next steps did not have an exact correlation to any of the steps starting in AD1000 or later. The highest correlation was to the AD1000 step – it looks like this recon uses a pre-AD1000 step (perhaps AD800?) for the period up to AD1400. But only in the reconstruction with “CRU” target.
Weird. It’s hard to figure out how this would happen given the source code. I wonder if the climate models work like this.
Making a Composite
After this diversion through new pet tricks, back to making the NH composite. Because UC sent me a Matlab version (that proved to have an iHAD target), I used this instrumental series (at http://www.meteo.psu.edu/~mann/supplements/MultiproxyMeans07/data/instrument/NH_iHAD_reform) as a target in gridboxcps.m.
I made a weighted average of the 15 gridcell series for the AD1000 network discussed yeaterday, weighting them by the cos latitude of the gridcell. This weighted average had a correlation of 0.9999968 to the column in the Matlab dataset NHmean that started in AD1000. So the composite matched perfectly up to scaling. Re-scaling to match the mean and sd of Mannian smoothed iHAD, the two matched up to a slight rounding discrepancy of about 0.006, as shown below. So we’ve pretty much manage to replicate all the little pet tricks so far.
Next we’ll try to figure out why some reconstructions start in AD1500 and some in AD600.