Pushing the Limits of AMS Radiocarbon Dating with Improved Bayesian Data Analysis

Abstract

We present an improved version of the continuous autoregressive (CAR) model, a Bayesian data analysis model for accelerator mass spectrometry (AMS). Measurement error is taken to be Poisson-distributed, improving the analysis for samples with only a few counts. This, in turn, enables pushing the limit of radiocarbon measurements to lower concentrations. On the computational side, machine drift is described with a vector of parameters, and hence the user can examine the probable shape of the trend. The model is compared to the conventional mean-based (MB) method, with simulated measurements representing a typical run of a modern AMS machine and a run with very old samples. In both comparisons, CAR has better precision, gives much more stable uncertainties, and is slightly more accurate. Finally, some results are given from Helsinki AMS measurements of background sample materials, with natural diamonds among them.