Still waiting for trilobites

First they find a living  coelacanth, thought to have become extinct in the Cretaceous.

Then they find dawn redwood, growing in a grove in China.

Now Kenneth Mertens has found living cysts of a dinoflagellate thought to have become extinct in the early Pleistocene. This species, a remnant of the warmer Pliocene oceans, found refuge in the Indo-Pacific warm pool.

Modern and fossil cysts of Dapsilidinium pastielsii

Modern and fossil cysts of Dapsilidinium pastielsii (from Mertens et al 2014)

I’m still holding out for a trilobite, but I’ll settle for an ammonite. Even a small one.


Posted in Peer reviewed literature, Silliness, Tropical Ecology | Tagged , | Leave a comment

Climate skeptics confuse time and space

Climatologies — gridded datasets of,  for example, mean monthly or annual temperature or precipitation — are very useful in large scale ecological work. I use the World Ocean Atlas extensively in my work on marine climate proxies and the New et al (2002) data from the Climate Research Unit for work on terrestrial proxies. However, there are always variables that one might like to include in an analysis that are not available.

Beckmann et al (2014) fill one of these gaps with a UVB climatology derived from satellite data. UVB, which causes sunburn and skin cancer in humans, is ecologically important, with effects at various scales from physiological to population dynamics. This new climatology will be useful, not least to the EECRG’s PARASOL project.

Beckmann et al assess the amount of new information in their climatology by correlating it with existing climatologies for temperature and precipitation, both globally and locally. At the global scale they find that UVB correlates with mean annual temperature, but is otherwise largely independent of existing climatologies, and at a local scale the correlation between UVB and temperature can either be positive or negative because of topographic effects.

It was that finding that UVB correlates with mean annual temperature that excited the climate skeptics at the Hockeyschtick and WUWT. HS writes, and Watts copies:

A paper published today in Methods in Ecology and Evolution describes a new satellite dataset of solar UV-B radiation for use in ecological studies. According to the authors, “UV-B surfaces were correlated with global mean temperature and annual mean radiation data, but exhibited variable spatial associations across the globe.” The finding is notable, since climate scientists dismiss the role of the Sun in climate change by only looking at the tiny 0.1% variations in total solar irradiance [TSI] over solar cycles, ignoring the large variations in solar UV of up to 100% over solar cycles, and which according to this paper, correlates to global mean temperature. Thus, the role of the Sun and solar amplification mechanisms on climate is only at the earliest stages of understanding.

And both mislead their audience. Again. I guess that neither would have had access to the journal (nor do I – there are other means of getting papers), but a sentence in the abstract makes it very clear that this is a spatial rather than temporal analysis.

“We correlated our data sets with selected variables of existing bioclimatic surfaces for land and … ocean regions to test for relations to known gradients and patterns.”

Nick Stokes, undertaking his Sisyphean labour, pointed out that Watts was wrong. Needless to say, Watts has not corrected his article.

Beckmann et al (2014) glUV: a global UV-B radiation data set for macroecological studies. Methods in Ecology and Evolution 5, 372–383

Posted in climate, Fake climate sceptics, Peer reviewed literature, Silliness, solar variability, WUWT | Tagged , | 5 Comments

Perhaps the most depressing palaeoecology paper ever

One of the major rationales for palaeoecological analyses is to provide data that can be used to validate climate models — if the models can predict past climate from periods with different climate forcings our confidence in their projections of future climate change under increased greenhouse gas concentrations should be enhanced. The Last Glacial Maximum, 21 kBP (LGM), is often used as a target because the climate forcing and response are large. Another often used target is the mid-Holocene, 6±0.5 kBP, towards the end of the Holocene thermal maximum when orbital forcing gave warmer summers and cooler winters at high latitudes. The orbital forcing was stronger earlier in the Holocene but this is a less suitable target as the climate was complicated by the remnants of the Laurentide Ice Sheets and massive pulses of meltwater, such as at the 8.2 kBP event.

At the LGM, most proxies, especially at mid and high latitudes, give a coherent response (colder) that is larger than the uncertainty, and is in agreement with the sign and approximate magnitude of the modelled climate.

The mid-Holocene is a more difficult target as the climate change is much smaller than that of the LGM. A new paper by Hessler et al, under review at Climate of the Past Discussions, examines a global compilation of mid-Holocene sea-surface temperature (SST) reconstructions inferred from different proxies and finds significant discrepancies between proxies and that the reconstructed anomalies are often smaller than the uncertainty. They conclude that the SST reconstructions cannot serve as a target for testing climate models.

Hessler et al’s compilation includes reconstructions inferred from alkenones, calcium-magnesium ratios in foram tests, and dinocyst and foram assemblages. Their conclusion that the proxies (which they insist on calling sensors) cannot be used to validate the models is in contrast to a recent paper in CPD which initially claimed that the mid-Holocene climate models’ simulations of sea ice had no skill because they could not match dinocyst-inferred sea-ice cover (the replies to reviewers suggests that the final paper will be more balanced).

Hessler et al’s conclusions are depressing. If palaeoceanographic proxies cannot give consistent reconstructions for the mid-Holocene, are they of any value in the Holocene? If not, compilations such as the global Holocene temperature reconstruction from Marcott et al (2013) are questionable as they include many marine proxies.

I’ve written several papers questioning how good proxies are. Now I want to consider if Hessler et al might be throwing the baby out with the bathwater: what might be going wrong with their analysis that makes the proxies appear to be of doubtful utility?

Chronological error. Lack of a match between proxies from different cores can always be blamed on chronological error. Perhaps they would match if the chronology was better. The inclusion criteria from Hessler et al required at least two radiocarbon dates (or other chronological control points) within the last 10 kBP. This is rather minimal, but presumably many fewer records would be included with stricter criteria. No attempt was made (probably wisely given the grief that Marcott et al received) to remake the age-depth models. It is perfectly possible that some of the chronologies are in error by more that a thousand years. I don’t think that is a big problem – there is no particular reason to expect the climate at 5 kBP or 7 kBP to be very different from 6 kBP. This would be a problem is they were trying to find records of a short-duration event like the 8.2 kBP cooling.

Under-estimation of proxy error. The uncertainty estimates for both the dinocyst and foram assemblage-based reconstructions are underestimated. The most important cause of this is the spatial autocorrelation in the calibration set which violates the assumption of independent observations. The true uncertainty is probably 50-100% larger than that reported. This is both good and bad news. The good news is that the proxies are less likely to contradict each other as the uncertainties on the reconstructions are wider. The bad news is the larger uncertainties will mean that even fewer of the reconstructions would be deemed statistically significant.

The proxies’ sensitivities differ. Much as we would like to reconstruct SST, some of the proxies may not directly care about it. In at least part of the ocean, including the Norwegian Sea, forams are more sensitive to sub-surface conditions than those at the surface. Worse, surface and sub-surface temperature trends are not necessarily the same. The seasonal sensitivity of the proxies may also differ. Hessler et al reconstruct summer and winter SST with some proxies; these are largely fanciful as there is no evidence that the assemblages contain sufficient information to make independent seasonal reconstructions. In the Nordic Seas, most foram production occurs in summer, but the transfer functions imply that winter SST is more important than summer SST. What is happening is that the temperature at the depth at which the forams are growing is set during winter overturn, so it is the summer temperature at depth which is important. These issues mean that even in the (exceedingly) implausible event that all the reconstructions were perfect, they might still appear to contradict each other.

Bad apples. It takes one bad apple to spoil the whole barrel. Is one bad proxy enough to make everything appear inconsistent? The authors have taken the diplomatic strategy of treating all proxies equally, but they are probably not all equally reliable. Nor are all reconstructions equally good – some may have non-analogue communities or other issues. The dinocyst reconstructions from the Nordic Sea at the LGM are seriously strange (warmer than modern), I wouldn’t want to bet much on the Holocene reconstructions being much better. The dinocyst reconstructions in Hessler et al are more varied than the other reconstructions, but they are not entirely responsible for the inconsistencies.

It may be too much to hope for coherent responses on a core-by-core basis from noisy proxies. Perhaps the regional signals are valid for at least some proxies, seasons and depths, but determining which proxies, seasons and depths are valid will take some effort. I think there is utility in Holocene reconstructions, but they need care in interpretation. Any hope that the climate modellers had that they could take mid-Holocene reconstructions off the shelf and use them to validate their simulations is unlikely to be fulfilled.

Hessler, I., Harrison, S. P., Kucera, M., Waelbroeck, C., Chen, M.-T., Andersson, C., de Vernal, A., Fréchette, B., Cloke-Hayes, A., Leduc, G., and Londeix, L. 2014. Implication of methodological uncertainties for Mid-Holocene sea surface temperature reconstructions, Clim. Past Discuss., 10, 1747-1782, doi:10.5194/cpd-10-1747-2014


Posted in climate, Peer reviewed literature, transfer function | Tagged , , | 4 Comments

Variance inflation factors and ordination model selection

Variance inflation factors (VIF) give a measure of the extent of multicollinearity in the predictors of a regression. If the VIF of a predictor is high, it indicates that that predictor is highly correlated with other predictors, it contains little or no unique information, and there is redundancy in the set of predictors.

We don’t really want to have redundant predictors in a constrained ordination: they make the analysis more difficult to interpret, and the more predictors we have, the less constrained the ordination is, the more it resembles an unconstrained ordination.

Model selection in ordinations is tricky. Techniques such as forward selection can guide the choice of predictors to include, but also bias the results (see Juggins 2013).

Recently I have seen a paper and reviewed a manuscript using VIF to determine which predictor to drop from a constrained ordination in a backwards selection.

The procedure they were using was

  1. Generate a constrained ordination with all available predictors.
  2. Calculate the VIF of each variable.
  3. If any variable has a VIF over a threshold (typically 20), drop the variable with the highest VIF
  4. Repeat until all remaining variables have a VIF below the threshold.

I want to show that this is a procedure that can go badly wrong using an example based on the SWAP diatom-pH calibration set.

First I load the data from the rioja package and make two fake pH variables that are highly correlated with the observed pH by adding Gaussian noise to the observed pH.


env<-data.frame(pH=SWAP$pH, fakepH1=rnorm(length(SWAP$pH), SWAP$pH, .2), fakepH2=rnorm(length(SWAP$pH), SWAP$pH, .2))
pairs(env)#highly correlated.

Now I fit a constrained correspondence analysis using cca, and calculate the VIF.

mod<-cca(sqrt(SWAP$spec)~., env)
#      pH  fakepH1  fakepH2
#26.82960 12.63575 14.49851
CCA of the SWAP calibration set with the predictor pH between two fake predictors correlated with pH

CCA of the SWAP calibration set with the predictor pH between two fake predictors correlated with pH

The VIF for pH is greater than the threshold 20, and so under the above procedure is dropped. If we fit a new model without pH we find the VIF of the remaining predictors has fallen and all are now below the threshold. An anova shows that both predictors are statistically significant

mod2<-cca(sqrt(SWAP$spec)~.-pH, env)
# fakepH1  fakepH2
#6.822985 6.822985

anova(mod2, by="terms")
#Permutation test for cca under reduced model
#Terms added sequentially (first to last)
#Model: cca(formula = sqrt(SWAP$spec) ~ (pH + fakepH1 + fakepH2) - pH, data = env)
#          Df   Chisq       F N.Perm Pr(>F)
#fakepH1    1  0.3406 11.0050     99   0.01 **
#fakepH2    1  0.0530  1.7122     99   0.01 **
#Residual 164  5.0754
#Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

What’s happening is that pH is highly correlated with both fake predictors, whereas the fake predictors are less strongly correlated with each other so have a lower VIF.

Any procedure with a high risk of throwing out the true predictor and instead accepting two fake predictors is not a useful method.

Is my toy example realistic? I would argue it is. One of the studies I saw used the temperature of all the months of a year as predictors for a calibration set. Imagine that June is the true predictor, but June temperature will be highly correlated with May and July temperatures. The procedure risks returning three or four months spaced evenly around the calendar rather than just June.

VIF is a very useful indicator that there is multicollinearity in a data set. It is a poor indicator of which predictor should be dropped from a model.

Posted in EDA, R | Tagged | 3 Comments

On the niche of Thalassiosira faurii, perils in palaeoecology

Palaeoenvironmental conditions can be reconstructed from microfossils preserved in sediments using the relationship between species and the environment. Usually the species’ relationships with the environment — their niches — are insufficiently constrained by experimental data, so we are forced to rely upon observations of species abundances and environmental conditions in a modern calibration set and assume that we can make useful inferences of the niche from these data.

Sometimes this assumption fails. This post tells of one such case.

I did my PhD, and some subsequent work, trying to infer climate from diatom assemblages in sediment cores from closed-basin lakes in Ethiopia, and later Tanzania. Closed-basin lakes are lakes that don’t have a river flowing out of them (or groundwater leakage), and should be sensitive to climate change:  deep and relatively fresh when the climate is humid; and shallow and saline during arid periods. Diatoms, algae with beautiful siliceous cell walls or frustules that preserve well in most lake sediments, are sensitive to salinity and water depth and so should be possible to use to constrain past changes in hydrology. Françoise Gasse and coworkers (Gasse et al 1995) developed a transfer function to reconstruct conductivity (a variable related to salinity) from diatom assemblages. I used this transfer function in my thesis and subsequent work.

The first paper I wrote was on a 6500-year long diatom stratigraphy from Lake Awassa, a large but fairly shallow caldera lake in the Ethiopian rift valley. Even though the lake has no out-flowing rivers, it is fairly fresh, suggesting that there is groundwater leakage, probably to the north. I was expecting to find evidence that Lake Awassa was deeper and fresher during the early Holocene, which was more humid in most of North Africa, with lakes scattered across the Sahara. I was hoping to be able to make a continuous climate reconstruction across the end of the termination of the early Holocene humid period – was it an abrupt or gradual change? Instead, the diatom-conductivity model suggested that the lake was more saline during two phases of the mid-Holocene, before the expected termination of the humid period. Either the timing and nature of the termination inferred from several sites across North Africa were incorrect, or there was something funny going on in Lake Awassa. I decided that the latter was more plausible.

Both of the apparently saline phases in Lake Awassa occurred immediately after very fine-grained tephra layers, so I investigate the potential for a volcanic origin for the saline phases. I found sufficient evidence from other sites to make what I thought was a plausible case for pulses of saline groundwater somehow related to the volcanism making the lake more saline.

One of the main salinity indicators in my Lake Awassa core was Thalassiosira faurii. Most Thalassiosira are marine species, but some live in, often brackish, lakes.

Thalassiosira faurii (Roubeix et al 2014)

Thalassiosira faurii (Roubeix et al 2014)

Gasse et al (1995) estimated the conductivity optimum of T. faurii to be ~9000 μScm-1, calculated by weighted-averaging. With the European Diatom Database (EDDI) we can look at the data behind this estimate, and plot them (you might need to enable java for unsigned applets – sorry I wrote/borrowed the java code for these plots long before security holes in java were known).

T. faurii relative abundance against salinity in the African diatom calibration set

T. faurii relative abundance against salinity in the African diatom calibration set

The marked sample and the other two samples with a high relative abundance of T. faurii are all from a salt swamp in Niger. Without these samples, the optimum would be somewhat lower. This dependence on one site, where T. faurii does not reach the maximum relative abundance (60%) found in Late Awassa, is not ideal.

Roubeix et al (2014) investigated the salinity niche of T. faurii collected from Lake Langano, another rift-valley lake 60 km to the north of Awassa, by measuring its growth rate when grown at different salinities under controlled conditions in a laboratory.

Growth rates of Thalassiosira faurii (black diamonds) and Anomoeoneis sphaerophora (open diamonds) versus conductivity of the culture medium. The dashed line shows the variations of pH between cultures. Vertical bars represent the standard deviations of the triplicates. A star means that the cells were still alive after 15 days, whereas a cross means that they were all dead. The two triangles indicate the conductivity and pH of lake water from which the diatoms were isolated.

Growth rates of Thalassiosira faurii (black diamonds) and Anomoeoneis sphaerophora (open diamonds) versus conductivity of the culture medium. The dashed line shows the variations of pH between cultures. Vertical bars represent the standard deviations of the replicates. A star means that the cells were still alive after 15 days, whereas a cross means that they were all dead. The two triangles indicate the conductivity and pH of lake water from which the diatoms were isolated.

At least the clone grown by Roubeix et al (2014) did not have an conductivity optimum of ~9000 μScm-1; it died at conductivities this high. The optimal conductivity was estimated at just  400 μScm-1. This is the optimum in the absence of any competition, the realised optimum in lakes many be somewhat higher, but is unlikely to be much above ~1000 μScm-1, as by ~2000 μScm-1 there is no growth. Roubeix et al use this new information to reinterpret the conditions in Lake Abiyata, a saline lake adjacent to Langano, during the Younger Dryas. Previous work by Chalié and Gasse (2002) had inferred saline conditions during this interval.

Roubeix et al also question my reconstruction from Lake Awassa. If T. faurii has a much lower conductivity optimum, then the conductivity spikes I reconstruct in the mid-Holocene may be an artefact. While I recognise that my proposed mechanism of pulses of saline hydrothermal water is speculative, I’m not entirely convinced that I was wrong as the salinity spike is not based on the presence of T. faurii alone. Instead, I found T. faurii in association with the saline indicators Navicula elkab (optimum ~14000 μScm-1) and T. rudolfii (optimum ~11400 μScm-1). Both these conductivity optima estimates seem to be well supported by data from a variety of sites.

The discrepancy between the old observational and new experimental optimum of T. faurii may reflect taxonomic problems. Perhaps T. faurii is not a single species but a suite of morphologically similar cryptic species with different salinity niches, with the diatoms from the salt swamp in Niger representing one end of the spectrum and the clone from Langano representing the other. Unless morphological characteristics that distinguish salt-tolerant from non-tolerant variants of T. faurii, it will not be possible to guarantee which variant is present in the palaeoecological records and any reconstruction will be rather uncertain.

Cryptic species are known in several of the taxonomic groups used for reconstructing climate, so there is a risk of cryptic species with different ecological preferences may be a widespread problem in palaeoecology.

Part of the problem relates to the lack of external data. If I examine a pollen calibration set and find the optimum temperature for lime (Tilia) to be 5°C, there is enough literature on the ecology of Tilia to recognise that there is a problem. With diatoms, dinoflagellate cysts and many other groups, there is little or no literature on the ecology of the different species, so we cannot easily recognise when the optima we calculate are spurious. Yet another reason why care is required when using transfer functions to reconstruct palaeoenvironmental conditions.

Gasse, F., Juggins, S. & Ben Khelifa, K. 1995. Diatom-based transfer functions for inferring past hydrochemical characteristics of African lakes. Palaeogeogr. Palaeoclim. Palaeoecol. 117: 31–54.

Roubeix, V., Chalié, F. & Gasse, F. (2014) The diatom Thalassiosira faurii (Gasse) Hasle in the Ziway–Shala lakes (Ethiopia) and implications for paleoclimatic reconstructions: Case study of the Glacial–Holocene transition in East Africa. Palaeogeogr. Palaeoclim. Palaeoecol. 402: 104–112.

Posted in climate, Palaeohydrology, Peer reviewed literature, transfer function | Tagged , | Leave a comment

Smilodon and the climate skeptic

I had been meaning to write about La Brea Tar Pits, having spent a morning with the mammoths last month, paying homage to the bones of the Pleistocene megafauna, the relics of extinction.


Mammoth columbi  (my photo)

Researchers from the museum have published two papers on the evolution of two of the predators found at La Brea, Smilodon fatalis the sabre-toothed-cat, and Canis dirus the dire wolf, in response to climate change in the late Pleistocene. The papers are based on the cranial morphology of the two species, which, especially the dire wolf, are present in large numbers in the tar pits. For Smilodon, they find a large morphotype during warmer periods of the late Pleistocene and a small morphotype in the colder periods, which they suggest may relate to the choice of prey species.


Smilodon fatalis (my photo)

Anthony Watts has written an incisive and intelligent commentary about the press release announcing these two papers in a post “The La Brea Tars Pits gets themselves in a sticky wicket over climate change and adaptation“.

One of the most shrill arguments from alarmists is the idea that climate change will wipe out species because they can’t adapt. The claims run from polar bears to tortoises, to plants and coral. Yes, if we listen to these arguments, Nature so poorly equipped it’s creatures that they can’t adapt to a slightly warmer future.

Except when the last ice age ended, and it got warmer, and the saber-toothed cats and wolves got bigger because the prey got bigger…instead of disappearing due to “climate change”.

OK, so that wasn’t particularly incisive or intelligent. Watts  is comparing the evolutionary response of large, free-roaming populations of carnivores over millennia, to what might be expected for the remnant populations of wildlife, constrained by agriculture and urbanisation to small and fragmented areas, over a century.

Posted in Peer reviewed literature, WUWT, Fake climate sceptics, Silliness | Tagged , , | 3 Comments

INTIMATE training school 2014

Don’t try searching for “INTIMATE dating” on Google; you won’t find what you are looking for.

This is what you want – a week long training school in palaeoecological and palaeoclimate techniques, organised by the INTIMATE  network (INTegrating Ice core Marine and TErrestrial records) in Romania. The training school will be held by the volcanic crater Lake St. Anne between May 30 and June 5, 2014. I’ll be giving a session on developing transfer functions and interpreting palaeoecological data using R.

The course is open to end-MSc students, PhD fellows, and early-stage postdocs. The application deadline is 27th April. 

Here is a video of last year’s INTIMATE training school.

Posted in climate, transfer function | Tagged , | Leave a comment