Robust climate reconstructions of the most recent centuries and millennia are invaluable for placing modern warming in the context of natural variability. Here we present an extended and revised database (version 1.1) of proxy temperature records recently used to reconstruct Arctic temperatures for the past 2,000 years. The datasets are presented in a machine-readable format, and have been extended with the geochronologic data and consistently generated time-uncertain ensembles, which will be useful in future analyses of the influence of geochronologic uncertainty. A standardized description of the seasonality of the temperature response for each record, as reported by the original authors, is also included to motivate a more nuanced approach to integrating records with variable seasonal sensitivities. Despite the predominance of seasonal, rather than annual, temperature responders in the database, comparisons with the instrumental record of temperature suggest that, as a whole, the datasets best record annual temperature variability across the Arctic, especially in northeast Canada and Greenland, where the density of records is highest.
Background & Summary
An accurate understanding of the past one to two thousand years of Earth's climate history is critical for placing recent warming in the context natural climate variability. Consequently, extensive efforts have been made to reconstruct regional1, hemispheric2,
Here we present an Arctic proxy temperature database for the past 2,000 years. The database is a revised version of the one used to reconstruct temperature in the Arctic for the past 2,000 years, which was recently included as part of the global summary by the Past Global Changes (PAGES) 2k Consortium1. In addition, we expanded the database by including consistently determined chronological uncertainty estimates for every record, except tree-ring records. These data are needed to quantify the influence of age uncertainty in climate reconstructions, but are rarely accessible to researchers aiming to develop large-scale climate reconstructions. This database also complements the recent Arctic Holocene Transitions (AHT) database11, a well-formatted collection of Arctic paleoclimate records for the Holocene. The overlap between the two datasets is minimal (9% of the sites in the AHT database are also included in this collection) because the AHT database includes records that extend further back at lower resolution; all records go back to at least 6000 years ago, and most extend 9000 years. Additionally, the AHT database only includes the geochronology data for radiometrically dated records, and does not include age ensembles for addressing age uncertainties. To our knowledge, the collection presented in this data descriptor is the first compilation of proxy climate data to include age ensembles, or age uncertainty estimates of any kind for layer-counted records.
Data aggregation and formatting
The database presented here is a revised version of the one used for the Arctic region of the PAGES 2k Network1 (Figure 1). Each revision is described below and in Table 1. The records selected were required to meet several criteria. Specifically all records:
are from north of 60°N;
extend back in time to at least 1500 AD;
have an average sample resolution less than 50 years;
have at least one age control point every 500 years;
have been published in a peer-reviewed journal, where evidence is presented documenting that the record is sensitive to temperature. This evidence may be statistical (e.g., correlation with nearby instrumental temperature data), or mechanistic (e.g., description by the authors of mechanisms by which the archive senses temperature change).
In several cases, the fifth criterion above is not met throughout the entire record (e.g., following AD 1720, agriculture nearby Lake Korttajärvi disrupts the temperature sensitivity of the record12). In these cases, we excluded the section of the record that violates this criterion.
In this study, we substantially expand the PAGES Arctic 2k database by including formatted geochronology data (e.g., radiocarbon ages and associated data) for the radiometrically-dated records, and systematically determined age-ensembles for all of the radiometrically-dated and layer-counted records in the database.
For each radiometrically-dated record, we developed a new age-depth model using the original geochronology data from each site and the Bayesian ACcumulatiON (BACON) algorithm13. BACON is a Bayesian age-modeling routine written for the software package R that takes advantage of prior knowledge about the distribution and autocorrelation structure of sedimentation rates in a sequence. The algorithm employs an adaptive Markov Chain Monte Carlo algorithm that allows for Bayesian learning to update the sedimentation-rate distribution.
The new age models do not replace those of the original study. Indeed, it is likely that the original investigators incorporated expert knowledge into the development of the original age models that we cannot replicate. Although the revised best-estimate age models may, in some cases, be inferior, there are two advantages to our approach. First, by systematically determining ages using a consistent methodology, we eliminate the aspect of age uncertainty and bias when comparing two records due to choices made during age modelling and the nuances of the many approaches originally used. Second, for each site, we extract a subset of age-ensemble members, which will facilitate future efforts to quantify the influence of age uncertainty in Arctic mean and temperature field reconstructions. This is important because nearly all of the original age models did not consider age-uncertain ensembles, and the data are not available for the few that did.
The PAGES Arctic 2k database includes 26 records from annually banded (varved) lake sediment and glacier ice for which the chronologies are developed by layer counting. The timeline for tree-ring records are also based on layer counting for which cross-dating among many samples makes tree-ring chronologies robust with negligible error14,15. Age uncertainty for annually banded sediments and ice cores typically increases with age. Although such records can often reach subannual precision, replication is more difficult and costly than with tree ring records, and consequently, cross-dating is rare, but possible with sufficient replication.
To develop time-uncertain ensembles for the layer-counted records, we used BAM (Banded Age Model), a probabilistic model of age errors in layer-counted chronologies16. The model allows a flexible parametric representation of such errors (either as Poisson or Bernoulli processes), and separately considers the possibility of double counting or missing a band. For each layer-counted chronology, we used BAM with published over- and under-counting estimates from the original study of each record (Table 1). When such estimates were not available, we applied conservative estimates of 1% for both over- and under-counting.
Arctic-wide temperature reconstruction
Changes from PAGES 2k Consortium (2013)
Here we present an Arctic regional temperature reconstruction that revises the one published recently by the PAGES 2k Consortium1. The revisions include updating records using more recent published studies from three sites17,
The interpreted temperature relation of the series from Hvítárvatn28 was corrected from positive to negative.
A 50-year offset in the ages of the record from Lone Spruce Pond29 was corrected.
The coordinates of the Copper River tree-ring reconstruction24 were corrected.
For this study, we did not add any new records to the database, or those that satisfy other criteria. We refer to this revised database as version 1.1. Additional records, including those sensitive to other aspects of the climate system (e.g., precipitation), will be included during the ongoing phase 2 of the PAGES 2k project. We suggest the next version of the database that includes additional records be designated as ‘version 2’.
The PAGES 2k Consortium1 used the Pairwise Comparison method (PaiCo9) to reconstruct the average Arctic mean-annual temperature for the past 2,000 years. PaiCo is a type of composite-plus-scale method8 that is unique because it does not require annually sampled data, nor the assumption that the proxy-temperature relation is linear (only monotonic). These features made it ideal for the Arctic 2k reconstruction. Here we use PaiCo to replicate the Arctic temperature reconstruction1, including the changes to the proxy database described above, to evaluate how the revisions influence the reconstruction.
Overall, the database revisions have a fairly minor impact on the relative variability in the reconstruction, but they do affect the long-term trend (Figure 2). The primary change is a relative increase in reconstructed temperatures for most of the record, especially between AD 1–1300. This results in an amplified long-term cooling trend that preceded 20th century warming; 0.47 °C/kyr in the revised reconstruction compared to 0.29 °C/kyr in the original. Decadal—scale variability in the revised reconstruction is quite similar to that determined by Kaufman et al. 7; however, the variability is about twice as great in the revised PAGES Arctic 2k reconstruction (Figure 2d). This is likely due the averaging and scaling procedures used in the earlier study7.
The PAGES Arctic 2k database presented here (v 1.1) is archived at the National Oceanic and Atmospheric Administration's World Data Center for Paleoclimatology (WDC-Paleo) http://ncdc.noaa.gov/paleo/study/16973, and the data are formatted according to WDC-Paleo's most recent standards http://www.ncdc.noaa.gov/data-access/paleoclimatology-data/contributing. The database is also archived on figshare [Data Citation 1: Figshare http://dx.doi.org/10.6084/m9.figshare.1054736]. For each record, there are self-describing and machine-readable ascii-files that include extensive metadata (e.g., source, title, investigators, publications, site and chronology metadata, variable descriptions) as well as the time-series and chronology data (when appropriate). Additionally, each site (except tree-ring records) has a corresponding netCDF file that archives the age-model ensembles. These files include up to four large matrices, depending on archive type and resolution:
AgeYoungEns: An ensemble of age estimates corresponding to the upper extent of each sampled interval. Each column is a different ensemble member.
AgeOldEns: Same as AgeYoungEns, but for the lower extent of each sample.
BaconAgeEnsemble: Ensemble of age models determined by BACON13. Each column is a different ensemble member (radiometrically dated only).
BaconAgeEnsDepths: Depths corresponding to ages in BaconAgeEnsemble (radiometrically dated only).
AgeEns: An ensemble of age estimates for the annually-resolved, layer-counted records as determined by BAM16. Each column is a different ensemble member (layer counted only).
DataEns: An ensemble of time—series perturbed by the simulated age uncertainty in AgeEns. Each column is a different ensemble member (layer counted only).
The PAGES Arctic 2k temperature database includes records that infer past temperature variability
from five types of natural archives. Each of these archives respond to temperature changes in
different ways, and that signal is recorded in each archive's chemical, physical, or biological properties. An overview of the records comprising the database is presented in Table 1. A novel aspect of this collection is the specification of the seasonal correlation of each record as described in the original publication. As shown in Table 1, the seasonal response of the proxies is quite variable, yet most synthesis and reconstruction efforts, including both the original and revised reconstructions described above, disregard the potential for seasonal differences among records that bias inferred climate changes in the past. The first step towards a more realistic treatment of seasonality is a uniform handling of these metadata, and we hope that future compilations will make this a priority. Although the records are well-summarized in Table 1 and in each records file in the database, the full details behind the collection, analysis and interpretation of each of the 56 records in the database is beyond the scope of this compilation, and we refer readers to the original publications for that information12,17,
Evidence that the records in the database reflect past temperature variability can be found in the original publications associated with each record. Here, we examine the extent to which the database as a whole captures observed temperature variability in the region. To do this, we calculated field correlations and their significance between each record in the database and the Natiaonal Aeronautics and Space Administration's (NASA) Goddard Institute for Space Studies Surface Temperature Analysis (GISTEMP) product with 1,200-km smoothing67 during the period of overlap (AD 1880–2000). In this analysis, the time series for each site, as well as the Arctic-wide reconstruction, were correlated against the temperature record for every grid cell north of 60° N. Significance at each grid cell was determined using a Student's T-test following correction for autocorrelation68. All calculations were performed at the temporal resolution of the proxy values; annual-mean temperatures were used for the annually-sampled records, and averages of multiple years corresponding to the sampling of the low-resolution records were calculated to correlate with the lower-resolution records.
This analysis shows that the revised PAGES Arctic 2k temperature reconstruction does an excellent job of capturing observed temperature variability in the Arctic, with significant (P<0.05) correlations over most the Arctic (Figure 3a,b). This is consistent with patterns observed from the summary of individual record field correlations (Figure 3c,d), although several of the sites demonstrate insignificant correlations over much or even all of the Arctic (Supplementary Figure S1). These records are typically those with low resolution and time uncertainty, which confounds this analyis. The interpretation of temperature sensitivity at these sites is derived from expert understanding of the system, rather than statistical comparison with instrumental data. In both the PAGES Arctic 2k temperature reconstruction, and as a whole from the individual sites, the highest correlations were calculated over northeast Canada and Greenland, where data density is highest. Interestingly, despite strong data coverage, and several sites with strong local correlations (Supplementary Figure S1), the temperature variability in Fennoscandia is not particularly well represented in the database. This may be due to out-of-phase decadal-scale temperature variability between Fennoscandia and the western part of the North Atlantic. Indeed, instrumental temperatures from near Greenland and northeastern Canada are poorly correlated with temperatures from Fennoscandia (Supplementary Figure S2). Some of this feature is due to the choice to compare the reconstruction to annual temperatures, thereby integrating some of the strong out-of-phase relationship that characterizes the region during the winter. However, a weaker, but similar pattern is present when analyzing summer (JJA) data only (Supplementary Figure S3). We also examine how the reconstruction correlates with instrumental summer (JJA) temperatures (Supplementary Figures S4). As expected, the reconstruction correlates better with summer than annual temperatures over Fennoscandia, however, the results are mixed elsewhere. Correlations with winter half-year (ONDJFM) temperatures strongly resemble annual correlations, but with fewer significant correlations across the Arctic. This resemblance is likely due to the dominance of winter temperature variability in the Arctic69. Overall, the reconstruction and records as a whole appear more representative of annual than either winter or summer temperatures. This is not because the records are sensing annual temperatures, rather, it is likely an artifact of including both summer and winter sensitive records in the compilation. Indeed, the spatial heterogeneity of the response highlights the biases introduced due to the variable seasonal response of proxy types and individual sites, and the shortcomings of index reconstructions, and highlights the need for a more nuanced consideration of spatial and seasonal variability in paleoclimate syntheses.
Finally, it should be noted that whereas these analyses are useful for quantifying some aspects of temperature sensitivity, they are poorly suited to determine the extent to which the records reflect long-term (centuries to millennia) changes in past temperature, or the stability of the modern relation back through time.
How to cite this article: McKay, N. P. and Kaufman, D. S. An extended Arctic proxy temperature database for the past 2,000 years. Sci. Data 1:140026 doi: 10.1038/sdata.2014.26 (2014).
McKay, N. & Kaufman, D. Figshare http://dx.doi.org/10.6084/m9.figshare.1054736 (2014).
We thank those who discovered and shared errors and updates to the original PAGES Arctic 2k database. Code and support for BACON and BAM was kindly provided by Maarten Blaauw and Maud Comboul, respectively. Kevin Anchukaitis and participants of the PAGES 2k Climate Reconstruction Methods Workshop contributed to this study. WDC-NOAA Paleoclimatology helped format and guided the design the data files. Support for PAGES activities is provided by the US and Swiss National Science Foundations, US National Oceanographic and Atmospheric Administration and by the International Geosphere-Biosphere Programme. N.P.M. was supported by NSF award ARC-1107869. We thank Sami Hanhijärvi and Atte Korhola for compiling the original version of the PAGES Arctic 2k database, and the many colleagues who kindly made digital versions of their data available for this product. The comments and suggestions of two anonymous reviewers improved this data descriptor and data collection.
This work is licensed under a Creative Commons Attribution 4.0 International License. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in the credit line; if the material is not included under the Creative Commons license, users will need to obtain permission from the license holder to reproduce the material. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0 Metadata associated with this Data Descriptor is available at http://www.nature.com/sdata/ and is released under the CC0 waiver to maximize reuse.