date: Tue, 19 Dec 2006 13:26:07 UT from: grlonline@agu.org subject: Review Received by Geophysical Research Letters to: t.osborn@uea.ac.uk Content-Transfer-Encoding: binary Content-Type: multipart/alternative; boundary="_----------=_116653476782846" MIME-Version: 1.0 X-Mailer: MIME::Lite 3.01 (F2.6; B2.11; Q2.03) Date: Tue, 19 Dec 2006 13:26:07 UT Message-Id: <116653476798@gems2> Dear Dr. Osborn: Thank you for your review of "Correction for proxy number and coherence in a large-scale temperature reconstruction" by David Frank, Jan Esper, and Edward Cook [Paper #2006GL028692], which we have safely received. A copy of this review is attached below for your reference. Thank you for your time and effort! Sincerely, James Famiglietti Editor Geophysical Research Letters ------------------------------------------------------------------------ Science Category: Science Category 2 Presentation Category: Presentation Category A Annotated Manuscript: No Anonymous: Yes Referrals: No Confidential Referrals: Highlight: No Highlight: Formal Review: Review of Frank, Esper and Cook "Correction for proxy number and coherence in a large-scale temperature reconstruction" This manuscript addresses the important issue of artificial changes through time in the variance of a frequently-used reconstruction of Northern Hemisphere temperature changes during the last 1200 years or so. Artificial change in variance is just one of a number of issues that remain to be dealt with in relation to this (and indeed some other) temperature reconstructions, but nevertheless it is entirely appropriate for a manuscript to focus on this one issue. The method used to quantify, and to compensate for, the artifical changes in variance is mostly appropriate, though further consideration to the issue of non-stationarity should also be given. The magnitude of the artificial changes in variance is sufficiently large to warrant this publication of a revised reconstruction that removes them, though it should be made clearer throughout that this is not the only issue to consider when evaluating the information provided by this and other reconstructions Similar adjustments to remove artificial changes in variance, arising from changing sample size through time, have been made in some other climate reconstructions (which should be cited to accurately represent the research record), though perhaps not in such a systematic way as presented here. Overall, therefore, I consider that this manuscript is appropriate for publication, subject to minor revisions. Major comments (1) As suggested in the introductory paragraph above, it should be made clearer throughout the manuscript that there are other important sources of uncertainty/error in the reconstruction that are not addressed here. Related to this, I would suggest that the word "correction" be replaced with the word "adjustment" in the title and at other appropriate places within the text because use of the word "correction" suggests that the new record is now "correct" whereas it will still be subject to considerable error (the estimated reconstruction error has not been shown in any of the figures either). (2) On pages 4 and 11 of the manuscript, it is noted that the method used to modify the records is applicable in cases where the individual time series are stationary. On page 11, this is followed by a mention of Jones et al. (2001), who applied the adjustments only to the high-frequency residuals from a smoothing filter, under the assumption that the high-frequency component will be stationary, but allowing non-stationarity on longer time scales. This is followed by the statement that tree-ring data contain noise on short and long time scales, which I presume is their justification for applying the adjustments to all time scales. It should be made clear that just because some data may contain noise on all time scales, this does not mean that they are then non-stationary on all time scales. Certainly it means that some component of the variability may contain artificial changes in variance on all time scales, which should be removed where possible but it does not mean that the entire time series variation is a random stationary deviation from a mean. I appreciate that it is not possible to know the non-stationary signal, and hence identify the stationary deviations from the signal, but nevertheless the manuscript would be improved by greater consideration of this issue. On page 11 it is stated that "we approached a rough stationarity requirement by setting the long-term mean of series to zero". This is a strange statement, since adjusting the mean level of a non-stationary series yields a series that is still non-stationary!! Perhaps a more defensible statement would be "we continued our analysis under the assumption that the individual tree-ring series were stationary; if this assumption is invalid, then some real climate changes during periods with small samples of data (e.g., the MWP) may be adjusted towards the mean level of our reconstruction by a greater amount than they should be Climate deviations during periods of small sample size might, therefore, be underestimated. We did, however, adjust the long-term means of each series to equal zero prior to making the adjustments." Overall, the stationarity assumption and its implications must be given greater consideration in the manuscript. (3) The manuscript appropriately cites previous work that has discussed the dependence of variance on sample size, put forward the adjustment method, and applies it to instrumental temperature data (Wigley et al., 1984; Osborn et al., 1997; Jones et al., 2001; Brohan et al., 2006). It would be appropriate to cite the adjustment method paper more prominently at the start of the discussion section (page 11). More importantly, however, it is necessary to point out that the method has previously been applied to temperature reconstructions. Careful reading of Briffa et al. (1998; Nature, 393, 450-455) indicates that they made variance adjustments when making a large-scale average of chronologies or of regional series, though it is not stated that they adjusted the individual tree-ring chronologies for this effect. Similarly, Jones et al. (1998; Holocene, 8, 455-471) also state that they made such adjustments when forming a large-scale average of individual series as did Briffa et al. (2001; J. Geophys. Res., 106, 2929-2941). It must be made clear that the current manuscript is not the first example of this method being used with proxy-based temperature reconstructions. (4) Overall, the text of the manuscript needs to be improved. Although it appears to be mostly understandable, there are cases throughout where the text must be much more precise and tight, to avoid ambiguities arising from slightly sloppy wording. Make sure, for example, that you take into account words with specific statistical meanings. The "mean-value function" is mentioned a number of times, for example, but it should be clarified that you don't mean the time-average of the reconstructed values, but instead you mean the average of the sample of data in one particular year. A careful reading of the manuscript by all the authors, with a view to considering whether a reader of the manuscript could understand precisely and unambiguously what has been done, or what is meant, should easily find the various problematic sentences. Minor comments (1) Abstract, final sentence: perhaps change "appear to" to "may", since you haven't demonstrated that these other reconstructions do contain similar biases. (2) Page 2, line 45: would "measurement records" be better than the more technical "predictors"? (3) Page 2, line 50: replace "namely" by "typically", since not all records deteriorate monotonically into the past. (4) Page 3, line 66: you cannot "eliminate changes in the proxy network", but you can "eliminate the influence of changes in the proxy network". (5) Page 3, lines 68-69: I don't understand "while remaining fixed to the ECS dataset". (6) Page 3, line 78: replace "variance of a mean collection" with "variance of the mean of a collection". (7) Page 5, line 130: suggest you add "tree-ring" before "measurement" to remind the reader what data you are referring to. (8) Page 6: where RUNNINGr is discussed, you should also note and discuss the possibility that there could be real, climate-driven variations in inter-series correlation that you would want to retain, rather than considering them all as sources of artificial variance change. Also, why did you not consider a RUNNINGr adjustment when making the large-scale averaging set, where only the MEANr adjustment was shown. Did you try it and consider it to be inappropriate? For what reason? (9) Page 8, line 196: should "that make their mean highly sensitive" really be "that make their variance highly sensitive"? (10) Page 9, lines 226-229: the sentence about "self-consistent best estimates" needs to be explained more clearly. (11) Page 9, lines 229-231: expand this sentence about the overall temperature range (note "range" is better than "amplitude") to explain that the reason why the range is unaffected by the adjustments is that the "warmest" and "coldest" reconstructed temperatures occur (give dates) during periods with large samples of data. (12) Page 10, line 243: Jones et al. (1998) is not an EOF-based method, since they did composite plus scaling (they did show some EOFs of the proxy data, but didn't actually use them in making their NH reconstruction). (13) Page 12, line 289: I don't think the problem is just that the "site locations are generally less known", but rather that they may be more scattered with greater separation between them. (14) Figure 1C: Why do the standard deviations computed in running 100-year windows extend to each end of the record - I would expect them to stop 50 years from each end. (15) Figure 2D: The caption should say which line is which (red or black). (16) Figure 2E: The y-axis label says "Rel. variance" but should it be "Rel. St. Dev."? (17) Figure 3A: The caption states that this is a log scale, but the figure does not appear to have a log scale. (18) Figure 3B: The caption states that these are the "corrections", but I think that they are the "corrected series". (19) Supplementary material, lines 30-34: Make it clearer that the reason for the reduced variance earlier on is that they used many different regression models, each re-calibrated using the different subsets of proxy data that are available in each period. (20) Suppl. Mat., line 39: "speculative" not "speculate". (21) Suppl. Mat., line 48: what does "unadjusted MEANr adjusted" mean? (22) Figure S1: Are the units "degrees C" rather than "Z"?