final_program_abstracts[1]

11 IMSC Session Program
Statistical analyses of the abrupt temperature increases of the
last interglacial
Tuesday - Poster Session 3
Daniel Peavoy
The North Atlantic, during the last glacial period, was subject to large and sudden
climate changes known as Dansgaard-Oeschger (DO) events. These large increases in
temperature, occuring within one hundred years or less, are clearly visible in the
North Greenland Ice core Project (NGRIP) δ 18 O record, which is a proxy for
temperature dating back to the previous interglacial (110 Kyr before present). It is
important to analyse these events in order to understand their pattern of recurrence
and to determine whether similar features are present in the Holocene climate.
Previous conclusions regarding the periodic or random ocurrence of DO events have
been based upon their identification by eye or by using thresholding methods. Since
the climate background temperature changes over the period due to Milankovitch
forcing and the DO events vary in magnitude, it is difficult to distinguish them
unambiguously.
Here we propose to identify DO events by their distinct dynamics rather than their
magnitude. For this we introduce an unobserved climate state, corresponding to either
usual fluctuations or DO event dynamics. Conditional on this hidden state the data is
assumed to evolve according to one of two independent AR(1) processes. We make
no assumptions about the parameters in this model, instead they are estimated from
the NGRIP data using Bayesian methods. Consequently, we are able to present the
posterior probability of an event as a function of time, displaying our uncertainty in
their occurence. We find that there is considerable agreement between times of high
probability in our model and the original identification of DO events.
We consider two different models for the evolution of the hidden state. The first,
assumes the events are a random Poisson Process with unknown rate, the second, that
they occur periodically with unknown phase, period and probability. The two models
are compared in a Bayesian framework using Monte Carlo methods. Finally, we
assess the likelihood of the purported 1450 year period with regards to the second
model.
Abstracts 100