This is an important and provocative paper, presenting an interesting and innovative analysis that offers a fresh take on a long-standing problem. Specifically, whether the paleoclimatic reconstructions that employ transfer functions, modern calibration datasets, multivariate fossil assemblages, and space-for-time substitution are heavily biased by assumptions of stability in the correlation structure among climatic variables in space and time. Prior papers by these authors and others (e.g. Juggins 2013) have argued that shifting correlation structures among variables can lead to major and often unrecognized biases.
This paper makes a new contribution by working entirely with a set of modeled climates and vegetation (PFTs), for which the actual climate evolution for the Holocene is known. Hence, the authors can apply standard transfer function methods to the modeled vegetation PFTs and see how well the reconstructed climates compare to the simulated climates. The results clearly show that in this modeled context, the PFT-based climate reconstructions are really only able to reconstruct one variable (MTWA) well, in the focal region of Siberia.
This paper is timely given the recent reconstructions of Northern Hemisphere mean annual temperature by Marcott et al (2013) and the challenge by Liu et al. (2014) that the pollen-based paleotemperature reconstructions primarily represent summer instead of winter temperatures.
I’ve read both the paper and the commentary and found it all fascinating. My overall assessment is strongly for publication –this is the kind of paper that will spark debate but I think in a net positive direction; it should move the conversation forward. However, the paper can be expected to meet resistance from the proxy-based paleoclimatic community, who can easily dismiss italong the grounds outlined by the original two reviewers: the model is very simple, it contains only 9 PFTs, etc. I personally am not entirely convinced by the authors’ response. I personally find the findings more illustrative than definitive, given the simplicity of the vegetation model and its PFTs, and think that the paper should present its findings as such.
Hence, to really have the impact it deserves, this paper needs to further strengthen its introduction discussion. A careful weighing of the paper’s own limitations and caveats will go a long way to strengthening the paper’s overall argument and moving the conversation forward.
1. Introduction: This should be expanded, to better set relevance and context. Specifically:
a. It should mention the space-for-time assumption and clearly establish the underlying assumption that variables correlated in space also correlated in time. Explicitly state that we have a good a priori reason to expect this assumption to be violated for temperature, given a) the strong correlation of summer and winter temperature in space today and b) the known anticorrelation of summer and winter insolation in the NH over the past 10,000 years due to precessional forcing. (this is mentioned very late in ms., on P15L3-5)
b. Could more fully describe Juggins 2013 – this paper is cited in passing, but its key point about confounding variables isn’t really explained.
c. Cite the Marcott et al. and Liu et al. papers in the intro, as a way of signaling the importance of this topic in current data-model comparisons and conversations about Holocene warmth vs. 21st-century temperature changes. This topic is introduced in Discussion, but should be introduced earlier.
2. Uniformitarianism is appropriately a central theme of this paper. However, there are several distinctly different kinds of uniformitarianism. Lyell founded the discipline of geology by assuming that processes observable today also operated in the past. Assuming uniformitarianism of processes is fine. However, Lyell also assumed that rates were roughly uniformitarianism through time, which is false. Gould called these two forms of uniformitarianism ‘methodological’ (uniformitarianism applied to processes, true) and ‘substantive’ (uniformitarianism applied to rates, false) (http://philpapers.org/rec/GOUIUN). This paper is dealing with a third kind of uniformatiariansm, assuming that covariance structures are uniform through time. This, like the ‘substantive’ uniformitarianism, is clearly false.
a. So, I strongly suggest a more nuanced treatment of uniformitarianism that distinguishes these concepts.
b. Also, remove quotes around this phrase.
c. Adding and defining a new phrase (‘correlative uniformitarianism’?) can give others an easy way of citing the arguments in this paper.
3. In Discussion, strengthen your case by pointing to other papers that have also explored the stability of correlation structures and the effect of confounding variables.
a. Salonen et al. 2013 Holocene for demonstratining effect of alternate calib datasets and continentality on . TJul reconstructions.
b. Blois et al. 2013: explicitly tested the space-for-time assumption by running generalized dissimilarity models on spatial vs. temporal datasets of species turnover. They argued that GDMs fitted across space could predict emergent patterns of diversity across space, but also found that the modeled relationships between diversity and turnover varied quite a bit among climate variables.
c. More clearly set up the issue of assuming stable correlations and the issue of secondary variables. Insolation. See P15L3-5.
d. Should acknowledge that Siberia may be an end-member/worst-case region for paleoclimatic transfer functions, in which summer light and warmth is really critical. Hence, this may be a worst-case system for multivariate transfer functions where one variable really dominates (MTWA) and the others are very secondary. In other regions, multiple climatic controls on vegetation may be important and disentanglable by transfer functions.
4. I am not convinced by the N2 analyses presented by the authors. I have two major concerns:
a. The paper never establishes what is a meaningful difference in N2. It simply states that the N2 in pollen data and the N2 in the modeled PFTs are about the same. However, I suspect that for paleoclimatic transfer functions, even a difference of 0.5 in N2 could be important, given that this represents in some sense the degrees of freedom available to the transfer functions. So if e.g. N2 is ~1.5 for PFTs and ~2.5 for the pollen data, that would imply an extra degree of freedom or so in the pollen data and more multivariate power.
b. For any given time or place in the pollen data, N2 might be about 2, but over the entire global region, I’d suspect that N2 is >>2. But in the model, Global N2 can never be higher than 8, and in the extratropics, can only be 5.
5. In Abstract and Conclusions: Add caveats. Note that number of PFTs of study are fewer than in modern pollen datasets. But nevertheless, the issues raised here about confounding variables are consistent with those from empirical studies of calibration datasets.
a. In Conclusions, add a ‘more work is needed’ sentence with a pointer to a LPJ-GUESS study. Well posed ‘future work’ statements can be very effective at moving the field forward and spurring future work, either by the authors or by other teams.
MISCELLANEOUS COMMENTS
Summer Temperature. One takeaway message of this paper by a naïve reader could be that summer temperature (MTWA) is the most critical variable, and so pollen-based paleoclimatic reconstructions should restrict themselves to MTWA. The abstract itself implies that summer temperature is the critical variable. However, the rest of the paper shows more complexity and caveats to this inference:
• These are modeled results, and the model may be more sensitive to summer temperature than real-world vegetation.
• Figure 4 shows interesting deviations from this, particularly in the tropics and subtropics, where MTCO seems to be better reconstructed than MTWA.
• Fig 6 shows that the variables explaining variance in vegetation vary regionally, e.g. MAP in the tropics.
• Siberian focus really emphasizes MTWA, as noted above.
I suggest adding nuance to the abstract and adding a section of the Discussion specifically focused on the question of whether MTWA is always the best variable for pollen-based vegetation reconstructions. That would directly speak to the discussion by Liu et al. and Marcott et al.
Throughout paper, be very careful to not mix up PFTs and species – they are very different. For example, when referring to Hill’s N2 for the model simulations, use ‘effective number of PFTs’ . ‘Taxa’ also would be a good option.
P10L5: A RMSEP of 3C would be mostly unacceptable in real-world Holocene paleotemperature reconstructions, given that the Holocene signal of temperature change is on the order of 1-2 degrees in many places. Fig. 8 shows a similar trend, on the order of 2C. Adjust wording and note that this may be the case where the PFT-based paleoclimatic transfer functions are doing much worse than real-world pollen-based transfer functions.
P14L28-30: ‘real world’ is vague. Clarify that what you mean is that these results may not be applicable to pollen-based paleoclimatic transfer functions, because of their higher richness.
Figure 1: Clarify that these maps are from the model simulations. Fix axis title that says “# species” – it should say “# PFTs” or “# effective PFTs”
Figure 2:
This figure packs in a lot of information. It needs more information in axis titles and legend.
Fig 2B: Not clear that MTWA explains the most variance in modern vegetation, as claimed in Fig. legend.
Fig 2C: Clarify axes – is this ‘temperature’ MTWA, MTCO, or MAT?
Figure 2D: Vertical axis? What is this a % of?
Figure 2E: Define the PFT acronyms in legend.
Figure 2G: What is 2G? It’s not mentioned in legend. Vertical axis title? What are the dashed vertical lines?
Figure 7, exploring R2 – seemed less critical – could delete this figure.
See also the minor editing for grammar on the attached PDF
Regards: Jack Williams |