Esper in Morocco

The other series in Trouet that contributes to their NAO reconstruction is the Morocco chronology from Esper the non-Archiver. In the Scotland speleo series, we observed an awkward detrending step prior to the NAO reconstruction. In the Morocco series, there proves to be some awkward splices: a splice of two different series versions at a hinge point of AD1300; and the use of different normalization curves before and after AD1600.

The net effect of these splices is impossible to assess on the present record as none of Esper’s measurement data is archived and, in detail, his methods remain obscure. Nonetheless, evidence is presented here to indicate that these splices result in a substantial detrending relative to an RCS chronology based on the 40% of the data available from prior archiving by Stockton.

Unfortunately, the pending “data” contribution from Trouet et al is limited to the NAO reconstruction and the spliced Esper PDSI reconstruction, neither of which suffice to permit proper assessment of the article. Esper the non-Archiver has not archived his underlying Morocco measurement data.

Let’s start with the Esper drought (PDSI) reconstruction. The same series is illustrated in multiple versions (maybe the referees thought that this was different information, but it’s all the same.) In each case, there is a closing a downtick representing seemingly unprecedented drought.

or here

or here (being compared to another NAO reconstruction with which it has zero correlation):

Alert CA readers may recall that a Morocco chronology (morc014) was a component of MBH99 (back to AD1000), one that was re-used in Juckes et al 2007 (non-inverted) and I’ve compared this data as well (see the graphic below). If you squint back and forth, you can see considerable correspondence between the decadal peaks and downticks, but the MBH/Juckes version looks like nothing more than noise in terms of MWP-LIA-Modern variability (and indeed, white noise performs just as well as a bristlecone condiment).

Any dendros reading this post will observe that the chronologies used in the Mann-Juckes recon couldn’t have been “conservatively standardized”. While one would have thought that left and right had no natural orientation in dendochronology (which has a hard enough time deciding whether a series should be oriented up or down), “conservatively” in this context means something like RCS standardization, where one age model is determined for the data set and used to standardize all the series (as core by core standardization with short splines will yield things that look like the MBH version.)

As noted above, Esper hasn’t archived his measurements, so let’s look at the very large (40%) sample of measurement data that is archived and carry out our own RCS analysis.

Esper 2007: Tree-ring data used in this study include ~64,000 annual ring width measurements from Cedrus atlantica trees sampled in 1985 [Glueck and Stockton, 2001], and a newer collection of ~100,000 measurements sampled in 2002.

First, retrieve the data – the following script picks the data up from WDCP and organizes it into a data frame for easier analysis – columns being core ID, year, age and ring-width. My retrieval function make.rwl is a very old function for me (~2003); I’d write it more cleanly now, but it works – here downloading 63,734 measurements from 7 Stockton sites. The total number of measurements match the number reported above in Esper et al 2007 and the sites are listed in the Esper SI so this is apples and apples. The newer Esper data supposedly focuses more on older trees, but without any data, it’s impossible to comment on the reliability of this statement.

id=c(paste(“morc00″,1:3,sep=””), paste(“morc0″,11:14,sep=””))
source(“”) #use make.rwl
for(i in 1:length(id)) {
#[1] 63734 4

In this example, plot of average ring width against age shows a very nice negative exponential form, as shown below:

The above plot was given by:

tree=Data #I’m used to using tree as a name for this sort of data
fm=nls(rw ~ A+B*exp(-C*age),data = tree,
start = list( A=mean(tree$rw,na.rm=T)/4,B = mean(tree$rw,na.rm=T), C= .01 ),
alg = “default”, trace = TRUE,control=nls.control(maxiter=200, tol=1e-05, minFactor=1e-10));
B=fm$coef;round(B,5) #546.76899 1501.98114 0.01079< – B[1]+B[2]*exp(-B[3]*(1:max(tree$age)) )
plot(as.numeric(names(aging)),aging,type="l",ylab="Ring Width",main="Morocco RW by Age",col="grey50",xlab="Age")

Next we’ll calculate an RCS- chronology which, as I understand it, consists of standardizing all ring widths by fitting one growth curve. I’ve used my own algorithm here, which is based on careful study.

test= RCS.chronology(tree,method=”nls”)

I’ve occasionally posted RCS emulations in the past, sometimes provoking screeches from dendros that I’ve done something or other “WRONG”, but I was unsuccessful in eliciting any data sets where there are both archived measurement data and RCS chronologies and, to my knowledge, none of the public dendro software uses these methods contains an RCS module. The underlying math is pretty simple – here I’ve fit one age curve to the data and standardized on that basis. This yields the following RCS chronology – one which looks like neither the MBH version nor the Esper version, though, once again, if you squint back and forth, you can match the decadal variations.

What accounts for this difference? Well, Esper has more recent data that yields a downtick at the end of his series. Add a downtick to the above chronology and we’re still left with a conundrum.

Esper has considerable supplementary material online here that may have some clues.

First, Esper describes the elimination of about 34,000 measurements as follows:

Common variance between the nearby site chronologies from Tiz, Col, Tou, and Jaf supported combination of the old growth data into a merged dataset TCTJ integrating 326 tree-ring series consisting of ~134,000 annual width measurements.

On another occasion, Esper stated:

However as we mentioned earlier on the subject of biological growth populations, this does not mean that one could not improve a chronology by reducing the number of series used if the purpose of removing samples is to enhance a desired signal. The ability to pick and choose which samples to use is an advantage unique to dendroclimatology.

One hopes that this was not the justification for deleting 34,000 measurements.

Second, Esper used a different age-curve for data prior to AD1300 and after AD1300. This is described as follows:

the combined Norm-RCS record is represented by Norm back to 1300 and by RCS Old prior to this year.

Esper purports to justify this splice by observing a high correlation between these versions. Needless to say, a high correlation can co-exist with a drifting apart of the two versions and this would need to be examined in detail.

Thirdly, although one of the purposes of RCS methods is to have uniform age curves, Esper uses a different age curve for “Old” data (152 tree-ring series before AD 1600 and 174 series after AD 1600, respectively), as shown below:

It seems fairly clear that this use of different normalization curves before and after AD1600 results in a considerable de-trending of the RCS chronology. How is this justified?

Analysis of the RCS method applied to all TCTJ data indicated that resulting chronologies contain increasing long-term trends that are possibly exaggerated by the inclusion of measurement series from younger trees. This potential bias has been assessed by splitting the TCTJ data into Young and Old sub-samples, integrating 152 tree-ring series less than AD 1600 and 174 series greater than AD 1600, respectively. Alignment of these sub-samples by cambial age revealed substantial differences in initial growth rates and age trends (Figure S4). The Young data contain much wider rings (initially about 1.6 mm) and a steeper age trend than the Old data (first rings about 1.1 mm). …

These differences might to some extent arise from the tendency of greater pith offset (PO) – which is the difference in years between the innermost ring on a core sample and the true center or pith of a tree at sampling height – in larger trees [Esper et al., 2006]. In other words, there is a tendency that more innermost tree-rings were missed on core samples from old trees than on samples from young trees. The key reason for this tendency is the sheer size of Moroccan cedar trees, with individuals frequently exceeding diameters of 4-5 m at breast height. As a consequence, differently old tree-rings are related to each other within the RCS procedure [Esper et al., 2003a]. This is particularly important as the young trees, with greater growth values cover only the recent end of the chronology. And as such, they may impart a bias towards a positive trend in recent times.

This last sentence should be quite worrying to dendros. In this particular case, Esper, expecting the data to go down in the post-1980 period, identified a potential bias in young trees relative to old trees. But what of the far more common case where increasing ring widths are expected? How often do we see dendros (Esper himself for that matter) assessing the data against a “bias towards a positive trend in recent times”. Doesn’t seem to happen.

Update: I’ve run separate RCS by each site. The average RW at TZK is about 6 times the width at Col du Zad. The upward trend in the overall RCS plot is therefore mainly due to inhomogeneity in site location, with an increasing mix of high growth sites. This seems like a far more plausible stratification than Esper’s peculiar attempt to stratify Young and Old trees.

Jan Esper, David Frank, Ulf Buentgen, Anne Verstege, Juerg Luterbacher, and Elena Xoplaki, 2007. Long-term drought severity variations in MoroccoGRL 2007 url


  1. AnonyMoose
    Posted Apr 7, 2009 at 9:01 PM | Permalink

    snip –

    Steve: PLEASE – no philosophizing about tree ring widths in general.

  2. bender
    Posted Apr 7, 2009 at 9:03 PM | Permalink

    Esper the non-Archiver

    the“? Isn’t he just one of many? Perhaps a “Hall of Shame” should be started for non-archivers? Some are worse than others. They could be graded and issued demerits.

  3. bender
    Posted Apr 7, 2009 at 9:42 PM | Permalink

    In the Morrocco RW example above annual ring width declines as a function of age, but the variance increases (!?). What is the mechanistic explanation for the increasing variance over time? And if it is not climatic signal, shouldn’t it also be detrended? In these multi-author papers it sure would be nice to know who decided what.

    • Craig Loehle
      Posted Apr 8, 2009 at 6:41 AM | Permalink

      Re: bender (#3), It has been found in Forestry that after the sapling stage trees tend to show constant basal area increment each year, but since this area is in a ring around a growing tree, the ring width decreases as the tree grows. I would tend to model this with a 1/d^2 but dendros like an exponential. In very old trees the variance goes up because they start suffering from disease, storm damage, etc.

  4. bender
    Posted Apr 7, 2009 at 9:45 PM | Permalink

    The ability to pick and choose which samples to use is an advantage unique to dendroclimatology.

    Esper was the one who said that? What age is this cheeky boy? I would like to talk to him.

    • Posted Apr 7, 2009 at 9:57 PM | Permalink

      Re: bender (#4),

      I’m more curious to know if all of his peers feel the same way.

    • theduke
      Posted Apr 7, 2009 at 11:25 PM | Permalink

      Re: bender (#4),

      According to the link below, he started publishing in 1994.


      Or, as someone close to me once replied when he was arrested for a foolish crime and asked how old he was, “Not old enough.”

      • Geoff Sherrington
        Posted Apr 8, 2009 at 3:28 AM | Permalink

        Re: theduke (#6),

        A drunken youth, friend of son, asked by police “What are you doing here?” replied “I’ve often asked my Mum but she won’t tell me”.

        Is my bow drawn too long?

    • Steve McIntyre
      Posted Apr 7, 2009 at 11:32 PM | Permalink

      Re: bender (#4), bender, c’mon, :) surely even you can’t object

      if the purpose of removing samples is to enhance a desired signal.

      • bender
        Posted Apr 9, 2009 at 8:27 AM | Permalink

        Re: Steve McIntyre (#7),

        enhancing signal
        enhancing desired signal

        There’s a difference, and Esper’s choice of words – “enhancing desired signal” – is most questionable. So, yes I do object. What the hell is he playing at?

    • OldUnixHead
      Posted Apr 8, 2009 at 3:13 PM | Permalink

      Re: bender (#4),
      If you can believe Wikipedia, he was born in 1968.

  5. Geoff Sherrington
    Posted Apr 8, 2009 at 3:25 AM | Permalink

    These differences might to some extent arise from the tendency of greater pith offset (PO) – which is the difference in years between the innermost ring on a core sample and the true center or pith of a tree at sampling height – in larger trees [Esper et al., 2006].

    So the correction for pith to a tree a decade old is the same as one hundreds of years old? One curve covers all?

  6. John A
    Posted Apr 8, 2009 at 4:29 AM | Permalink

    snip –

    Steve: PLEASE – no philosophizing about tree ring widths in general.

  7. David L. Hagen
    Posted Apr 8, 2009 at 11:59 AM | Permalink

    Seconding bender (#2), may I recommend a section on your site with a running list of climate related publications, promised supplementary data, posting of data, requests for data, performance on requests, FOIA actions etc. This seems to be an ongoing and important contribution to the climate modeling field.

  8. Alan Bates
    Posted Apr 8, 2009 at 12:57 PM | Permalink

    Following on from David #12:

    “This seems to be an ongoing and important contribution to the climate modeling field…”

    …and will allow those papers to be discounted when formal assessments are made.

    (In my dreams, I know, but surely if a series of calculations is not repeatable it is not science?)

  9. Steve McIntyre
    Posted Apr 8, 2009 at 1:49 PM | Permalink

    I’ve added separate plots for the 7 available sites to the above post, yielding a variety of patterns. As noted in the update, it seems to me that a stratification by site is a more sensible sort of stratification than the ad hoc split between Yong and Old.

  10. Posted Apr 9, 2009 at 5:23 AM | Permalink

    Just a quick observation on the exponential fit plot. The errors are evidently (i.e. by eye) serially correlated and the sequence should be differenced before regression is performed. I don’t think it’d make a huge difference to the estimated parameters (the fit looks pretty good), but we’re trying to be methodologically correct, right?

  11. Skepticus
    Posted Apr 9, 2009 at 8:10 AM | Permalink

    OT care to comment on:

  12. Geoff Sherrington
    Posted Apr 10, 2009 at 12:49 AM | Permalink

    I might be taking this out of context, but I think not. If not, it is a nonsense.

    The ability to pick and choose which samples to use is an advantage unique to dendroclimatology.

    How do you think early gold miners went about their panning, since antiquity and probably before statistics were invented?

  13. Mark T
    Posted Apr 10, 2009 at 8:39 AM | Permalink

    Yeah, but they weren’t doing a weighted average, biased toward their productive locations, to tell you the hills all had gold, either. ;)


  14. Chas
    Posted Apr 10, 2009 at 2:00 PM | Permalink

    For the RCS he used a 100-year spline rather than an exponential -which seems to me to be overdoing it a little bit.

One Trackback

  1. By Is dendroclimatology a science? | Hoystory on Apr 13, 2009 at 2:49 AM

    [...] you shouldn’t be surprised. However as we mentioned earlier on the subject of biological growth populations, this does not [...]


Get every new post delivered to your Inbox.

Join 3,193 other followers

%d bloggers like this: